2312.11805v5

<details> <summary>Image 1 Details</summary>  ### Visual Description Icon/Small Image (184x28) </details> ## Gemini: A Family of Highly Capable Multimodal Models Gemini Team, Google 1 This report introduces a new family of multimodal models, Gemini, that exhibit remarkable capabilities across image, audio, video, and text understanding. The Gemini family consists of Ultra, Pro, and Nano sizes, suitable for applications ranging from complex reasoning tasks to on-device memory-constrained use-cases. Evaluation on a broad range of benchmarks shows that our most-capable Gemini Ultra model advances the state of the art in 30 of 32 of these benchmarks - notably being the first model to achieve human-expert performance on the well-studied exam benchmark MMLU, and improving the state of the art in every one of the 20 multimodal benchmarks we examined. We believe that the new capabilities of the Gemini family in cross-modal reasoning and language understanding will enable a wide variety of use cases. We discuss our approach toward post-training and deploying Gemini models responsibly to users through services including Gemini, Gemini Advanced, Google AI Studio, and Cloud Vertex AI. ## 1. Introduction We present Gemini, a family of highly capable multimodal models developed at Google. We trained Gemini models jointly across image, audio, video, and text data for the purpose of building a model with both strong generalist capabilities across modalities alongside cutting-edge understanding and reasoning performance in each respective domain. Gemini 1.0, our first version, comes in three sizes: Ultra for highly-complex tasks, Pro for enhanced performance and deployability at scale, and Nano for on-device applications. Each size is specifically tailored to address different computational limitations and application requirements. After large-scale pre-training, we post-train our models to improve overall quality, enhance target capabilities, and ensure alignment and safety criteria are met. Due to the varied requirements of our downstream applications, we have produced two post-trained Gemini model family variants. Chat-focused variants, referred to as Gemini Apps models, are optimized for Gemini and Gemini Advanced, our conversational AI service formerly known as Bard. Developer-focused variants, referred to as Gemini API models, are optimized for a range of products and are accessible through Google AI Studio and Cloud Vertex AI. We evaluate the performance of pre- and post-trained Gemini models on a comprehensive suite of internal and external benchmarks covering a wide range of language, coding, reasoning, and multimodal tasks. The Gemini family advances state-of-the-art in large-scale language modeling (Anil et al., 2023; Brown et al., 2020; Chowdhery et al., 2023; Hoffmann et al., 2022; OpenAI, 2023a; Radford et al., 2019; Rae et al., 2021), image understanding (Alayrac et al., 2022; Chen et al., 2022; Dosovitskiy et al., 2020; OpenAI, 2023b; Reed et al., 2022; Yu et al., 2022a), audio processing (Radford et al., 2023; Zhang et al., 2023), and video understanding (Alayrac et al., 2022; Chen et al., 2023). It also builds on the work on sequence models (Sutskever et al., 2014), a long history of work in deep learning based on neural networks (LeCun et al., 2015), and machine learning distributed systems 1 See Contributions and Acknowledgments section for full author list. Please send correspondence to gemini-1report@google.com (Barham et al., 2022; Bradbury et al., 2018; Dean et al., 2012) that enable large-scale training. Our most capable model, Gemini Ultra, achieves new state-of-the-art results in 30 of 32 benchmarks we report on, including 10 of 12 popular text and reasoning benchmarks, 9 of 9 image understanding benchmarks, 6 of 6 video understanding benchmarks, and 5 of 5 speech recognition and speech translation benchmarks. Gemini Ultra is the first model to achieve human-expert performance on MMLU (Hendrycks et al., 2021a) - a prominent benchmark testing knowledge and reasoning via a suite of exams - with a score above 90%. Beyond text, Gemini Ultra makes notable advances on challenging multimodal reasoning tasks. For example, on the recent MMMU benchmark (Yue et al., 2023), that comprises questions about images on multi-discipline tasks requiring college-level subject knowledge and deliberate reasoning, Gemini Ultra achieves a new state-of-the-art score of 62.4%, outperforming the previous best model by more than 5 percentage points. It provides a uniform performance lift for video question answering and audio understanding benchmarks. Qualitative evaluation showcases impressive crossmodal reasoning capabilities, enabling the model to understand and reason across an input sequence of audio, images, and text natively (see Figure 5 and Table 13). Consider the educational setting depicted in Figure 1 as an example. A teacher has drawn a physics problem of a skier going down a slope, and a student has worked through a solution to it. Using Gemini models' multimodal reasoning capabilities, the model is able to understand the messy handwriting, correctly understand the problem formulation, convert both the problem and solution to mathematical typesetting, identify the specific step of reasoning where the student went wrong in solving the problem, and then give a worked through correct solution to the problem. This opens up exciting educational possibilities, and we believe the new multimodal and reasoning capabilities of Gemini models have dramatic applications across many fields. The reasoning capabilities of large language models show promise toward building generalist agents that can tackle more complex multi-step problems. The AlphaCode team built AlphaCode 2 (Leblond et al, 2023), a new Gemini-model-powered agent, that combines Gemini models' reasoning capabilities with search and tool-use to excel at solving competitive programming problems. AlphaCode 2 ranks within the top 15% of entrants on the Codeforces competitive programming platform, a large improvement over its state-of-the-art predecessor in the top 50% (Li et al., 2022). In tandem, we advance the frontier of efficiency with Gemini Nano, a series of small models targeting on-device deployment. These models excel in on-device tasks, such as summarization, reading comprehension, text completion tasks, and exhibit impressive capabilities in reasoning, STEM, coding, multimodal, and multilingual tasks relative to their sizes. In the following sections, we first provide an overview of the model architecture, training infrastructure, and pre-training dataset. We then present detailed evaluations of the pre- and post-trained Gemini model family, covering well-studied benchmarks across text, code, image, audio and video which include both English performance and multilingual capabilities. Next we discuss our approach to post-training, highlight common and distinct aspects of the Gemini Apps and Gemini API model variants, and benchmark their performance on key capabilities. Responsible deployment is critical: we explain our process for impact assessments, developing model policies, evaluations, and mitigations of harm before deployment decisions. Finally, we discuss the broader implications of Gemini models, their limitations alongside their potential applications - paving the way for a new era of research and innovation in AI. Figure 1 | Verifying a student's solution to a physics problem. The model is able to correctly recognize all of the handwritten content and verify the reasoning. On top of understanding the text in the image, it needs to understand the problem setup and correctly follow instructions to generate L A T E X. <details> <summary>Image 2 Details</summary>  ### Visual Description \n ## Physics Problem Solution: Skier on a Slope ### Overview The image presents a handwritten solution to a physics problem involving a skier sliding down a frictionless slope. The problem asks for the skier's speed at the bottom of the slope, given the initial height (H = 40m) and the length of the slope (L = 80m). The student's solution is shown, followed by a corrected solution and explanation. ### Components/Axes The image is divided into two main sections: 1. **Problem Statement & Student Solution (Left Side):** Contains the problem description, a diagram of the slope, and the student's handwritten calculations. 2. **Corrected Solution & Explanation (Right Side):** Provides the correct solution and a step-by-step explanation. The diagram on the left side includes: * **Labels:** "Start", "Finish", "H=40 m", "L=80 m". * **Visual Representation:** A simple line representing the slope, with a skier depicted on it. ### Detailed Analysis or Content Details **Left Side - Student Solution:** The student's solution attempts to use conservation of energy. The steps are as follows: 1. **Initial Statement:** "The conservation of energy means that the energy of the skier is same at the beginning and at the end." 2. **Initial Energy:** `Start: E = mgL` 3. **Final Energy:** `End: E = 1/2 mv^2` 4. **Equating Energies:** `mgL = 1/2 mv^2` 5. **Solving for v:** `v = √(2gL) = √(2 * 9.81 * 80) = 39.6 m/s^-1` **Right Side - Corrected Solution:** 1. **Statement:** "The student did not get the correct answer. The student made a mistake in the calculation of the potential energy at the start of the slope. The potential energy at the start of the slope is mgh, where m is the mass of the skier, g is the acceleration due to gravity, and h is the height of the slope. The student used the length of the slope, L, instead of the height, H, in the calculation of the potential energy." 2. **Conservation of Energy:** "The conservation of energy means that the total energy of the skier is the same at the beginning and at the end of the slope." 3. **Initial Potential Energy:** "At the start of the slope, the skier has only potential energy, which is given by mgh." 4. **Final Kinetic Energy:** "At the end of the slope, the skier has only kinetic energy, which is given by 1/2 mv^2." 5. **Equating Energies:** `mgh = 1/2 mv^2` 6. **Simplification:** `gh = 1/2 v^2` 7. **Solving for v:** `v = √(2gh)` 8. **Substitution:** `v = √(2 * 9.81 * 40) = 28.01 m/s` **Additional Text:** * "Try to work about the question step by step." * "Did the student get the correct answer? If the solution is wrong, solve the problem." * "Make sure to use LaTeX for math and round off the final answer to two decimal places." ### Key Observations * The student incorrectly used the length of the slope (L) instead of the height (H) in the potential energy calculation. * The student's calculated velocity (39.6 m/s) is significantly higher than the correct velocity (28.01 m/s). * The corrected solution clearly explains the error and provides a step-by-step derivation of the correct answer. * The problem utilizes the principle of conservation of energy. ### Interpretation The image demonstrates a common mistake in applying the principle of conservation of energy – confusing distance with height. The student's initial approach was conceptually correct (equating initial potential energy to final kinetic energy), but the incorrect substitution of 'L' for 'H' led to a significantly inflated velocity. The corrected solution highlights the importance of carefully identifying the relevant variables and understanding the physical meaning of each term in the equation. The inclusion of LaTeX formatting instructions suggests a focus on precise mathematical representation. The problem serves as a good example for students learning about energy conservation and problem-solving techniques in physics. The detailed explanation on the right side is a good pedagogical approach, breaking down the solution into manageable steps and explicitly addressing the student's error. </details> ## 2. Model Architecture Gemini models build on top of Transformer decoders (Vaswani et al., 2017b) that are enhanced with improvements in architecture and model optimization to enable stable training at scale and optimized inference on Google's Tensor Processing Units. They are trained to support 32k context length, employing efficient attention mechanisms (for e.g. multi-query attention (Shazeer, 2019a)). Our first version, Gemini 1.0, comprises three main sizes to support a wide range of applications as discussed in Table 1. Gemini models are trained to accommodate textual input interleaved with a wide variety of audio and visual inputs, such as natural images, charts, screenshots, PDFs, and videos, and they can produce text and image outputs (see Figure 2). The visual encoding of Gemini models is inspired by our own foundational work on Flamingo (Alayrac et al., 2022), CoCa (Yu et al., 2022a), and PaLI (Chen et al., 2022), with the important distinction that the models are multimodal from the beginning and can natively output images using discrete image tokens (Ramesh et al., 2021; Yu et al., 2022b). Video understanding is accomplished by encoding the video as a sequence of frames in the large context window. Video frames or images can be interleaved naturally with text or audio as part of the model input. The models can handle variable input resolution in order to spend more compute on tasks that require fine-grained understanding. In addition, Gemini models can directly ingest audio Table 1 | An overview of the Gemini 1.0 model family. | Model size | Model description | |--------------|---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------| | Ultra | Our most capable model that delivers state-of-the-art performance across a wide range of highly complex tasks, including reasoning and multimodal tasks. It is efficiently serveable at scale on TPU accelerators due to the Gemini architecture. | | Pro | A performance-optimized model in terms of cost as well as latency that delivers significant performance across a wide range of tasks. This model exhibits strong reasoning performance and broad multimodal capabilities. | | Nano | Our most efficient model, designed to run on-device. We trained two versions of Nano, with 1.8B (Nano-1) and 3.25B (Nano-2) parameters, targeting low and high memory devices respectively. It is trained by distilling from larger Gemini models. It is 4-bit quantized for deployment and provides best-in-class performance. | Figure 2 | Gemini models support interleaved sequences of text, image, audio, and video as inputs (illustrated by tokens of different colors in the input sequence). They can output responses with interleaved image and text. <details> <summary>Image 3 Details</summary>  ### Visual Description \n ## Diagram: Multimodal Transformer Architecture ### Overview The image depicts a diagram illustrating the architecture of a multimodal transformer model. The model takes a sequence of different input modalities (text, audio, image, video) and processes them through a transformer network to generate corresponding outputs (image, text). The diagram emphasizes the flow of information from various input types into a central transformer block, and then branching out to modality-specific decoders. ### Components/Axes The diagram consists of the following components: * **Input Sequence:** Located on the left side of the diagram, this section represents the various input modalities. * **Text Input:** Represented by a light blue rounded rectangle with the letter "Aa" inside. * **Audio Input:** Represented by a red rounded rectangle with a sound wave icon inside. * **Image Input:** Represented by a green rounded rectangle with a grid icon inside. * **Video Input:** Represented by a yellow rounded rectangle with a play button icon inside. * **Transformer:** A large blue rectangle in the center of the diagram, labeled "Transformer". * **Image Decoder:** A light green rounded rectangle labeled "Image Decoder". * **Text Decoder:** A light blue rounded rectangle labeled "Text Decoder". * **Arrows:** Yellow arrows indicate the flow of information from the input modalities to the transformer, and from the transformer to the decoders. A series of colored dots (blue, orange, green) connect the input sequence to the transformer. ### Detailed Analysis or Content Details The diagram shows a sequential input process. The input sequence consists of four modalities: text, audio, image, and video. These inputs are fed into a transformer network. The transformer then outputs to two decoders: an image decoder and a text decoder. The input modalities are represented by distinct icons within rounded rectangles: * Text: "Aa" * Audio: Sound wave icon * Image: Grid icon * Video: Play button icon The arrows connecting the inputs to the transformer are yellow, indicating the flow of information. The colored dots (blue, orange, green) between the input sequence and the transformer likely represent different stages or layers within the transformer's processing of the input sequence. The outputs of the transformer are directed to two decoders: * Image Decoder: Outputs an image (represented by a grid icon). * Text Decoder: Outputs text (represented by "Aa"). ### Key Observations The diagram highlights the versatility of the transformer architecture in handling multiple input modalities. The central role of the transformer suggests it acts as a unifying processing unit for all input types. The separate decoders indicate that the transformer can generate outputs in different modalities based on the input. ### Interpretation This diagram illustrates a multimodal transformer model, a type of neural network capable of processing and generating data across different modalities (text, audio, image, video). The transformer architecture is known for its ability to capture long-range dependencies in sequential data, making it well-suited for handling complex multimodal inputs. The diagram suggests that the model learns a shared representation of the input modalities within the transformer, allowing it to generate outputs in different modalities. This type of model has applications in areas such as image captioning, video description, and multimodal search. The use of separate decoders for image and text suggests that the model is designed to generate outputs tailored to each modality. </details> signals at 16kHz from Universal Speech Model (USM) (Zhang et al., 2023) features. This enables the model to capture nuances that are typically lost when the audio is naively mapped to a text input (for example, see audio understanding demo on the website). Training the Gemini family of models required innovations in training algorithms, dataset, and infrastructure. For the Pro model, the inherent scalability of our infrastructure and learning algorithms enable us to complete pre-training in a matter of weeks, leveraging a fraction of the Ultra's resources. The Nano series of models leverage additional advancements in distillation and training algorithms to produce the best-in-class small language models for a wide variety of tasks, such as summarization and reading comprehension, which power our next generation on-device experiences. ## 3. Training Infrastructure We trained Gemini models using TPUv5e and TPUv4 (Jouppi et al., 2023), depending on their sizes and configuration. Training Gemini Ultra used a large fleet of TPUv4 accelerators owned by Google across multiple datacenters. This represents a significant increase in scale over our prior flagship model PaLM-2 which presented new infrastructure challenges. Scaling up the number of accelerators results in a proportionate decrease in the mean time between failure of hardware in the overall system. We minimized the rate of planned reschedules and preemptions, but genuine machine failures are commonplace across all hardware accelerators at such large scales. TPUv4 accelerators are deployed in 'SuperPods' of 4096 chips, each connected to a dedicated optical switch, which can dynamically reconfigure 4x4x4 chip cubes into arbitrary 3D torus topologies in around 10 seconds (Jouppi et al., 2023). For Gemini Ultra, we decided to retain a small number of cubes per superpod to allow for hot standbys and rolling maintenance. TPU accelerators primarily communicate over the high speed inter-chip-interconnect, but at Gemini Ultra scale, we combine SuperPods in multiple datacenters using Google's intra-cluster and inter-cluster network (Poutievski et al., 2022; Wetherall et al., 2023; yao Hong et al., 2018). Google's network latencies and bandwidths are sufficient to support the commonly used synchronous training paradigm, exploiting model parallelism within superpods and data-parallelism across superpods. The 'single controller' programming model of Jax (Bradbury et al., 2018) and Pathways (Barham et al., 2022) allows a single Python process to orchestrate the entire training run, dramatically simplifying the development workflow. The GSPMD partitioner (Xu et al., 2021) in the XLA compiler partitions the training step computation, and the MegaScale XLA compiler (XLA, 2019) pass statically schedules appropriate collectives so that they maximally overlap with the computation with very little variation in step time. Maintaining a high goodput 2 at this scale would have been impossible using the conventional approach of periodic checkpointing of weights to persistent cluster storage. For Gemini models, we instead made use of redundant in-memory copies of the model state, and on any unplanned hardware failures, we rapidly recover directly from an intact model replica. Compared to both PaLM and PaLM-2 (Anil et al., 2023), this provided a substantial speedup in recovery time, despite the significantly larger training resources being used. As a result, the overall goodput for the largest-scale training job increased from 85% to 97%. Training at unprecedented scale invariably surfaces new and interesting systems failure modes and in this instance one of the problems that we needed to address was that of 'Silent Data Corruption (SDC)' (Dixit et al., 2021; Hochschild et al., 2021; Vishwanathan et al., 2015). Although these are extremely rare, the scale of Gemini models means that we can expect SDC events to impact training every week or two. Rapidly detecting and removing faulty hardware required several new techniques that exploit deterministic replay to isolate incorrect computations, combined with proactive SDC scanners on idle machines and hot standbys. Our fully deterministic infrastructure allowed us to quickly identify root causes (including hardware failures) during the development leading up to the Ultra model, and this was a crucial ingredient towards stable training. ## 4. Pre-Training Dataset Gemini models are trained on a dataset that is both multimodal and multilingual. Our pre-training dataset uses data from web documents, books, and code, and includes image, audio, and video data. We use the SentencePiece tokenizer (Kudo and Richardson, 2018) and find that training the tokenizer on a large sample of the entire training corpus improves the inferred vocabulary and subsequently improves model performance. For example, we find Gemini models can efficiently 2 We define goodput as the time spent computing useful new steps over the elapsed time of the training job. tokenize non-Latin scripts which can, in turn, benefit model quality as well as training and inference speed. The number of tokens used to train the largest models were determined following the approach in Hoffmann et al. (2022). The smaller models are trained for significantly more tokens to improve performance for a given inference budget, similar to the approach advocated in Touvron et al. (2023a). We apply quality filters to all datasets, using both heuristic rules and model-based classifiers. We also perform safety filtering to remove harmful content based on our policies. To maintain the integrity of evaluations, we search for and remove any evaluation data that may have been in our training corpus before using data for training. The final data mixtures and weights were determined through ablations on smaller models. We stage training to alter the mixture composition during training - increasing the weight of domain-relevant data towards the end of training. We find that data quality is an important factor for highly-performing models, and believe that many interesting questions remain around finding the optimal dataset distribution for pre-training. ## 5. Evaluation The Gemini models are natively multimodal, as they are trained jointly across text, image, audio, and video. One open question is whether this joint training can result in a model which has strong capabilities in each domain - even when compared to models and approaches that are narrowly tailored to single domains. We find this to be the case: Gemini models set a new state of the art across a wide range of text, image, audio, and video benchmarks. ww ## 5.1. Text ## 5.1.1. Academic Benchmarks We compare pre- and post-trained Gemini Pro and Ultra models to a suite of external LLMs and our previous best model PaLM 2 across a series of text-based academic benchmarks covering reasoning, reading comprehension, STEM, and coding. We report these results in Table 2. Broadly, we find that the performance of Gemini Pro outperforms inference-optimized models such as GPT-3.5 and performs comparably with several of the most capable models available, and Gemini Ultra outperforms all current models. In this section, we examine some of these findings. On MMLU (Hendrycks et al., 2021a), Gemini Ultra can outperform all existing models, achieving an accuracy of 90.04%. MMLU is a holistic exam benchmark, which measures knowledge across a set of 57 subjects. Human expert performance is gauged at 89.8% by the benchmark authors, and Gemini Ultra is the first model to exceed this threshold, with the prior state-of-the-art result at 86.4%. Achieving high performance requires specialist knowledge across many domains (e.g. law, biology, history, etc.), alongside reading comprehension and reasoning. We find Gemini Ultra achieves highest accuracy when used in combination with a chain-of-thought prompting approach (Wei et al., 2022b) that accounts for model uncertainty. The model produces a chain of thought with k samples, for example 8 or 32. If there is a consensus above a preset threshold (selected based on the validation split), it selects this answer, otherwise it reverts to a greedy sample based on maximum likelihood choice without chain of thought. We refer the reader to appendix for a detailed breakdown of how this approach compares with only chain-of-thought prompting or only greedy sampling. In mathematics, a field commonly used to benchmark the analytical capabilities of models, Gemini Ultra shows strong performance on both elementary exams and competition-grade problem sets. For the grade-school math benchmark, GSM8K (Cobbe et al., 2021), we find Gemini Ultra reaches 94.4% accuracy with chain-of-thought prompting and self-consistency (Wang et al., 2022) compared to the previous best accuracy of 92% with the same prompting technique. Similar positive trends are observed in increased difficulty math problems drawn from middle- and high-school math competitions (MATH benchmark), with the Gemini Ultra model outperforming all competitor models, reaching 53.2% using 4-shot prompting. The model also outperforms the state of the art on even harder tasks derived from American Mathematical Competitions (150 questions from 2022 and 2023). Smaller models perform poorly on this challenging task scoring close to random, but Gemini Ultra can solve 32% of the questions, compared to the 30% solve rate for GPT-4. Gemini Ultra also excels in coding, a popular use case of current LLMs. We evaluate the model on many conventional and internal benchmarks and also measure its performance as part of more complex reasoning systems such as AlphaCode 2 (see Section 5.1.7 on complex reasoning systems). For example, on HumanEval, a standard code-completion benchmark (Chen et al., 2021) mapping function descriptions to Python implementations, instruction-tuned Gemini Ultra correctly implements 74.4% of problems. On a new held-out evaluation benchmark for python code generation tasks, Natural2Code, where we ensure no web leakage, Gemini Ultra achieves the highest score of 74.9%. Evaluation on these benchmarks is challenging and may be affected by data contamination. We performed an extensive leaked data analysis after training to ensure the results we report here are as scientifically sound as possible, but still found some minor issues and decided not to report results on e.g. LAMBADA (Paperno et al., 2016). As part of the evaluation process, on a popular benchmark, HellaSwag (Zellers et al., 2019), we find that an additional hundred fine-tuning steps on specific website extracts corresponding to the HellaSwag training set (which were not included in the Gemini model pretraining set) improve the validation accuracy of Gemini Pro to 89.6% and Gemini Ultra to 96.0%, when measured with 1-shot prompting (we measured GPT-4 obtained 92.3% when evaluated 1-shot via the API). This suggests that the benchmark results are susceptible to the pretraining dataset composition. We choose to report HellaSwag decontaminated results only in a 10-shot evaluation setting. We believe there is a need for more robust and nuanced standardized evaluation benchmarks with no leaked data. So, we evaluate Gemini models on several new held-out evaluation datasets that were recently released, such as WMT23 and Math-AMC 2022-2023 problems, or internally generated from non-web sources, such as Natural2Code. We refer the reader to Appendix 10.3 for a comprehensive list of our evaluation benchmarks. Even so, model performance on these benchmarks gives us an indication of the model capabilities and where they may provide impact on real-world tasks. For example, Gemini Ultra's impressive reasoning and STEM competencies pave the way for advancements in LLMs within the educational domain 3 . The ability to tackle complex mathematical and scientific concepts opens up exciting possibilities for personalized learning and intelligent tutoring systems. ## 5.1.2. Trends in Capabilities We investigate the trends in capabilities across the Gemini model family by evaluating them on a holistic harness of more than 50 benchmarks in six different capabilities, noting that some of the most notable benchmarks were discussed in the last section. These capabilities are: 'Factuality' covering open/closed-book retrieval and question answering tasks; 'Long-Context' covering longform summarization, retrieval and question answering tasks; 'Math/Science' including tasks for mathematical problem solving, theorem proving, and scientific exams; 'Reasoning' tasks that require arithmetic, scientific, and commonsense reasoning; 'Multilingual' tasks for translation, summarization, and reasoning in multiple languages. Several of these capabilities are targeted by post-training (Section 6). Please see Appendix 10.3 for a detailed list of tasks included for each capability. 3 See demos on website https://deepmind.google/gemini . | | Gemini Ultra | Gemini Pro | GPT-4 | GPT-3.5 | PaLM 2-L | Claude 2 | Inflect- ion-2 | Grok 1 | LLAMA-2 | |---------------------------------------------------------------------------------------------------|------------------------------|-----------------------------|-------------------------------------------------------|----------------------------|---------------------|------------------|------------------|--------------|--------------| | MMLU Multiple-choice questions in 57 subjects (professional & academic) (Hendrycks et al., 2021a) | 90.04% CoT@32 ∗ 83.7% 5-shot | 79.13% CoT@8 ∗ 71.8% 5-shot | 87.29% CoT@32 (via API ∗∗ ) 86.4% 5-shot | 70% 5-shot | 78.4% 5-shot | 78.5% 5-shot CoT | 79.6% 5-shot | 73.0% 5-shot | 68.0% ∗∗∗ | | GSM8K Grade-school math (Cobbe et al., 2021) | 94.4% Maj1@32 | 86.5% Maj1@32 | 92.0% SFT & 5-shot CoT | 57.1% 5-shot | 80.0% 5-shot | 88.0% 0-shot | 81.4% 8-shot | 62.9% 8-shot | 56.8% 5-shot | | MATH Math problems across 5 difficulty levels & 7 subdisciplines (Hendrycks et al., 2021b) | 53.2% 4-shot | 32.6% 4-shot | 52.9% 4-shot (via API ∗∗ ) 50.3% (Zheng et al., 2023) | 34.1% 4-shot (via API ∗∗ ) | 34.4% 4-shot | - | 34.8% | 23.9% 4-shot | 13.5% 4-shot | | BIG-Bench-Hard Subset of hard BIG-bench tasks written as CoT prob- lems (Srivastava et al., 2022) | 83.6% 3-shot | 75.0% 3-shot | 83.1% 3-shot (via API ∗∗ ) | 66.6% 3-shot (via API ∗∗ ) | 77.7% 3-shot | - | - | - | 51.2% 3-shot | | HumanEval Python coding tasks (Chen et al., 2021) | 74.4% 0-shot (PT ∗∗∗∗ ) | 67.7% 0-shot (PT ∗∗∗∗ ) | 67.0% 0-shot (reported) | 48.1% 0-shot | - | 70.0% 0-shot | 44.5% 0-shot | 63.2% 0-shot | 29.9% 0-shot | | Natural2Code Python code generation. (New held-out set with no leakage on web) | 74.9% 0-shot | 69.6% 0-shot | 73.9% 0-shot (via API ∗∗ ) | 62.3% 0-shot (via API ∗∗ ) | - | - | - | - | - | | DROP Reading comprehension & arithmetic. (metric: F1-score) (Dua et al., 2019) | 82.4 Variable shots | 74.1 Variable shots | 80.9 3-shot (reported) | 64.1 3-shot | 82.0 Variable shots | - | - | - | - | | HellaSwag (validation set) Common-sense multiple choice questions (Zellers et al., 2019) | 87.8% 10-shot | 84.7% 10-shot | 95.3% 10-shot (reported) | 85.5% 10-shot | 86.8% 10-shot | - | 89.0% 10-shot | - | 80.0% ∗∗∗ | | WMT23 Machine translation (met- ric: BLEURT) (Tom et al., 2023) | 74.4 1-shot (PT ∗∗∗∗ ) | 71.7 1-shot | 73.8 1-shot (via API ∗∗ ) | - | 72.7 1-shot | - | - | - | - | ## Table 2 | Gemini performance on text benchmarks with external comparisons and PaLM 2-L. ∗ The model produces a chain of thought with k = 8 or 32 samples, if there is a consensus above a threshold (chosen based on the validation split), it selects this answer, otherwise it reverts to a greedy sample. Further analysis in Appendix 10.2. ∗∗ Results self-collected via the API in Nov, 2023. ∗∗∗ Results shown use the decontaminated numbers from Touvron et al. (2023b) report as the most relevant comparison to Gemini models which have been decontaminated as well.) ∗∗∗∗ PT denotes a post-trained Gemini API model. We observe consistent quality gains with increased model size in Figure 3, especially in reasoning, math/science, summarization and long-context. Gemini Ultra is the best model across the board for all six capabilities. Gemini Pro, the second-largest model in the Gemini family of models, is also quite competitive while being a lot more efficient to serve. ## 5.1.3. Nano Bringing AI closer to the user, we discuss the Gemini Nano 1 and Nano 2 models engineered for on-device deployments. These models excel in summarization and reading comprehension tasks with per-task fine-tuning. Figure 3 shows the performance of these pre-trained models in comparison to the much larger Gemini Pro model, while Table 3 dives deeper into specific factuality, coding, Math/Science, and reasoning tasks. Nano-1 and Nano-2 model sizes are only 1.8B and 3.25B parameters respectively. Despite their size, they show exceptionally strong performance on factuality, i.e. retrieval-related tasks, and significant performance on reasoning, STEM, coding, multimodal and Figure 3 | Language understanding and generation performance of Gemini model family across different capabilities (normalized by the Gemini Pro model). <details> <summary>Image 4 Details</summary>  ### Visual Description \n ## Bar Chart: Normalized Performance vs Pro ### Overview This bar chart compares the normalized performance of four different models (Nano 1, Nano 2, Pro, and Ultra) across six different tasks: Factuality, Long-Context, Math/Science, Summarization, Reasoning, and Multilinguality. The performance is normalized against the "Pro" model, with values greater than 1.0 indicating better performance than the Pro model and values less than 1.0 indicating worse performance. ### Components/Axes * **X-axis:** Task Name (Factuality, Long-Context, Math/Science, Summarization, Reasoning, Multilinguality) * **Y-axis:** Normalized Performance vs Pro (Scale from 0.0 to 1.4) * **Legend:** * Nano 1 (represented by a reddish-orange color) * Nano 2 (represented by a yellow color) * Pro (represented by a green color) * Ultra (represented by a blue color) ### Detailed Analysis The chart consists of grouped bar plots for each task, with each group containing four bars representing the performance of the four models. * **Factuality:** * Nano 1: Approximately 0.82 * Nano 2: Approximately 0.95 * Pro: 1.00 * Ultra: Approximately 1.02 * **Long-Context:** * Nano 1: Approximately 0.65 * Nano 2: Approximately 1.25 * Pro: 1.00 * Ultra: Approximately 1.10 * **Math/Science:** * Nano 1: Approximately 0.32 * Nano 2: Approximately 0.65 * Pro: 1.00 * Ultra: Approximately 1.35 * **Summarization:** * Nano 1: Approximately 0.55 * Nano 2: Approximately 0.98 * Pro: 1.00 * Ultra: Approximately 1.05 * **Reasoning:** * Nano 1: Approximately 0.75 * Nano 2: Approximately 1.15 * Pro: 1.00 * Ultra: Approximately 1.25 * **Multilinguality:** * Nano 1: Approximately 0.62 * Nano 2: Approximately 0.98 * Pro: 1.00 * Ultra: Approximately 1.00 ### Key Observations * The "Ultra" model consistently performs at or above the level of the "Pro" model across all tasks. * "Nano 1" consistently underperforms compared to the other models, particularly in Math/Science. * "Nano 2" shows significant improvement over "Nano 1" and often approaches or exceeds the performance of the "Pro" model, especially in Long-Context and Reasoning. * The largest performance difference between models is observed in the Math/Science task, where "Ultra" significantly outperforms all other models. ### Interpretation The data suggests a clear hierarchy in model performance, with "Ultra" being the most capable and "Nano 1" being the least. The normalization against the "Pro" model provides a useful benchmark for evaluating the relative strengths and weaknesses of each model. The substantial performance gap in Math/Science indicates that this task is particularly challenging and benefits significantly from the capabilities of the "Ultra" model. The improvement from "Nano 1" to "Nano 2" suggests that model architecture or training data adjustments can have a substantial impact on performance. The consistent performance of "Ultra" near or above 1.0 across all tasks indicates a robust and versatile model. The chart demonstrates the trade-offs between model size/complexity and performance, with the larger models ("Pro" and "Ultra") generally achieving better results. </details> multilingual tasks. With new capabilities accessible to a broader set of platforms and devices, the Gemini models expand accessibility to everyone. Table 3 | Performance of Gemini Nano series on factuality, summarization, reasoning, coding and STEM tasks compared to significantly larger Gemini Pro model. | | Gemini Nano 1 | Gemini Nano 1 | Gemini Nano 2 | Gemini Nano 2 | |--------------------------------|-----------------|-------------------|-----------------|-------------------| | | accuracy | normalized by Pro | accuracy | normalized by Pro | | BoolQ | 71.6 | 0.81 | 79.3 | 0.90 | | TydiQA (GoldP) | 68.9 | 0.85 | 74.2 | 0.91 | | NaturalQuestions (Retrieved) | 38.6 | 0.69 | 46.5 | 0.83 | | NaturalQuestions (Closed-book) | 18.8 | 0.43 | 24.8 | 0.56 | | BIG-Bench-Hard (3-shot) | 34.8 | 0.47 | 42.4 | 0.58 | | MBPP | 20.0 | 0.33 | 27.2 | 0.45 | | MATH (4-shot) | 13.5 | 0.41 | 22.8 | 0.70 | | MMLU (5-shot) | 45.9 | 0.64 | 55.8 | 0.78 | ## 5.1.4. Multilinguality The multilingual capabilities of the Gemini models are evaluated using a diverse set of tasks requiring multilingual understanding, cross-lingual generalization, and the generation of text in multiple languages. These tasks include machine translation benchmarks (WMT 23 for high-medium-low resource translation; Flores, NTREX for low and very low resource languages), summarization benchmarks (XLSum, Wikilingua), and translated versions of common benchmarks (MGSM: professionally translated into 11 languages). ## 5.1.4.1 Machine Translation Translation is a canonical benchmark in machine learning with a rich history. We evaluated a posttrained Gemini API Ultra model (see Section 6.5.3) on the entire set of language pairs in the WMT 23 translation benchmark in a few-shot setting. Overall, we found that Gemini Ultra (and other Gemini models) performed remarkably well at translating from English to any other language, and surpassed the LLM-based translation methods when translating out-of-English, on high-resource, mid-resource and low-resource languages. In the WMT 23 out-of-English translation tasks, Gemini Ultra achieved the highest LLM-based translation quality, with an average BLEURT (Sellam et al., 2020) score of 74.8, compared to GPT-4's score of 73.6, and PaLM 2's score of 72.2. When averaged across all language pairs and directions for WMT 23, we see a similar trend with Gemini Ultra 74.4, GPT-4 73.8 and PaLM 2-L 72.7 average BLEURT scores on this benchmark. Table 4 | Performance of Gemini models on WMT 23 translation benchmark. All numbers with 1-shot. | WMT 23 (Avg BLEURT) | Gemini Ultra | Gemini Pro | Gemini Nano 2 | Gemini Nano 1 | GPT-4 | PaLM 2-L | |-----------------------|----------------|--------------|-----------------|-----------------|---------|------------| | High Resource | 74.2 | 71.7 | 67.7 | 64.1 | 74 | 72.6 | | Mid Resource | 74.7 | 71.8 | 67 | 64.8 | 73.6 | 72.7 | | Out-of-English | 74.8 | 71.5 | 66.2 | 65.2 | 73.6 | 72.2 | | Into-English | 73.9 | 72 | 69 | 63.5 | 74.1 | 73.4 | | All languages | 74.4 | 71.7 | 67.4 | 64.8 | 73.8 | 72.7 | In addition to the languages and translation tasks above, we also evaluate Gemini Ultra on very low-resource languages. These languages were sampled from the tail of the following language sets: Flores-200 (Tamazight and Kanure), NTREX (North Ndebele), and an internal benchmark (Quechua). For these languages, both from and into English, Gemini Ultra achieved an average chrF score of 27.0 in 1-shot setup, while the next-best model, PaLM 2-L, achieved a score of 25.3. ## 5.1.4.2 Multilingual Math and Summarization Beyond translation, we evaluated how well Gemini models perform in challenging tasks across a range of languages. We specifically investigated the math benchmark MGSM (Shi et al., 2023), which is a translated variant of the math benchmark GSM8K (Cobbe et al., 2021). We find Gemini Ultra achieves an accuracy of 79.0%, an advance over PaLM 2-L which scores 74.7%, when averaged across all languages in an 8-shot setup. We also benchmark Gemini models on the multilingual summarization benchmarks - XLSum (Hasan et al., 2021) and WikiLingua (Ladhak et al., 2020). In XLSum, Gemini Ultra reached an average of 17.6 rougeL score compared to 15.4 for PaLM 2. For Wikilingua, Gemini Ultra (5-shot) trails behind PaLM 2 (3-shot) measured in BLEURT score. See Table 5 for the full results. Overall the diverse set of multilingual benchmarks show that Gemini family models have a broad language coverage, enabling them to also reach locales and regions with low-resource languages. Table 5 | Performance of Gemini models on multilingual math and summarization. | | Gemini Ultra | Gemini Pro | GPT-4 | PaLM 2-L | |----------------|----------------|--------------|---------|------------| | MGSM (8-shot) | 79 | 63.5 | 74.5 | 74.7 | | XLsum (3-shot) | 17.6 | 16.2 | - | 15.4 | | Wikilingua | 48.9 | 47.8 | - | 50.4 | ## 5.1.5. Long Context Gemini models are trained with a sequence length of 32,768 tokens and we find that they make use of their context length effectively. We first verify this by running a synthetic retrieval test: we place key-value pairs at the beginning of the context, then add long filler text, and ask for value associated with a particular key. We find that the Ultra model retrieves the correct value with 98% accuracy when queried across the full context length. We further investigate this by plotting the negative log likelihood (NLL) versus the token index across a held-out set of long documents in Figure 4. We find that the NLL decreases with sequence position up to the full 32K context length. The longer context length of Gemini models enable new use cases such as retrieval over documents and video understanding discussed in Section 5.2.2. Figure 4 | Negative log likelihood as a function of token index across 32K context length on a held-out set of long documents. <details> <summary>Image 5 Details</summary>  ### Visual Description \n ## Line Chart: Negative Log Likelihood vs. Sequence Position ### Overview The image presents a line chart comparing the Negative Log Likelihood (NLL) for two models, "Pro" and "Ultra", across varying sequence positions. The chart illustrates how the NLL changes as the sequence position increases, indicating model performance. ### Components/Axes * **X-axis:** Sequence position, ranging from 8 to 32K (32,000). The scale is logarithmic, with markers at 8, 16, 32, 64, 128, 256, 512, 1K (1,000), 2K (2,000), 4K (4,000), 8K (8,000), 16K (16,000), and 32K. * **Y-axis:** Negative Log Likelihood (NLL). The scale is linear, but the exact range is not explicitly labeled. * **Legend:** Located in the top-right corner, identifying the two data series: * "Pro" - represented by a green line. * "Ultra" - represented by a blue line. * **Grid:** A light gray grid is present, aiding in the readability of the chart. ### Detailed Analysis * **Ultra (Blue Line):** The blue line representing "Ultra" starts at approximately NLL = 5.5 at sequence position 8. It exhibits a steep downward slope initially, decreasing rapidly to approximately NLL = 2.5 at sequence position 256. The slope continues to decrease, leveling off around NLL = 1.5 at sequence position 8K, and reaching approximately NLL = 1.2 at sequence position 32K. * **Pro (Green Line):** The green line representing "Pro" starts at approximately NLL = 4.0 at sequence position 8. It also shows a downward trend, but less steep than the "Ultra" line. It decreases to approximately NLL = 2.5 at sequence position 512. The slope continues to decrease, leveling off around NLL = 1.7 at sequence position 8K, and reaching approximately NLL = 1.5 at sequence position 32K. ### Key Observations * The "Ultra" model consistently exhibits a lower NLL than the "Pro" model across all sequence positions, indicating better performance. * Both models demonstrate diminishing returns as the sequence position increases. The rate of NLL decrease slows down significantly after 2K. * The difference in NLL between the two models is most pronounced at lower sequence positions (8-512) and becomes less significant at higher sequence positions (8K-32K). ### Interpretation The chart suggests that the "Ultra" model outperforms the "Pro" model in terms of negative log likelihood, implying a better fit to the data or a more accurate prediction capability. The decreasing NLL with increasing sequence position for both models indicates that they become more confident in their predictions as they process longer sequences. The diminishing returns observed at higher sequence positions suggest that there is a limit to the benefit of processing increasingly longer sequences. The convergence of the two lines at higher sequence positions indicates that the performance gap between the models narrows as the sequence length increases. This could be due to the models reaching a point where they have captured most of the relevant information from the sequence, or due to limitations in the models' capacity to process very long sequences effectively. The chart provides valuable insights into the performance characteristics of the two models and can inform decisions about model selection and sequence length optimization. </details> ## 5.1.6. Factuality Factuality (Maynez et al., 2020) is a key focus of our model's training and deployment. We evaluate three aspects of factuality for our Gemini API models: 1. Closed-Book Factuality : If provided with a fact-seeking prompt without any given source, Gemini API models should not hallucinate incorrect information (see Section 2 of Roberts et al. (2020) for a definition). These prompts can range from information-seeking prompts (e.g. 'Who is the prime minister of India?') to semi-creative prompts that may request factual information (e.g. 'Write a 500-word speech in favor of the adoption of renewable energy'). 2. Attribution : If instructed to generate a response grounded to a given context, we aim to ensure that Gemini API models produce a response with the highest degree of faithfulness to the context (Maynez et al., 2020; Rashkin et al., 2023). This may include the summarization of a user-provided source, generating fine-grained citations given a question and provided snippets akin to Menick et al. (2022); Peng et al. (2023), answering questions from a long-form source such as a book (Mihaylov et al., 2018), and transforming a given source to a desired output (e.g. an email from a portion of a meeting transcript). 3. Hedging : If prompted with an input that is 'unanswerable', Gemini API models must acknowledge that it cannot provide a response by hedging to avoid hallucination. These include scenarios where the input prompt contains false-premise questions [see examples in Hu et al. (2023)], the input prompt instructs the model to perform open book QA, but the answer is not derivable from the given context, and so forth. Factuality is evaluated via human annotators who fact-check each response manually; we report the percentage of factually inaccurate responses as judged by annotators. Attribution is evaluated via human annotators who check for attribution to sources in the prompt for each response manually; the reported metric is AIS (Rashkin et al., 2023). For hedging, we use an automatic evaluation setup where we measure whether models hedge accurately. We compare Gemini API Pro with a version without any factuality-focused adaptation in Table 6. We see that the rate of inaccuracy is halved in the factuality set, the accuracy of attribution is increased by 50% from the attribution set, and the model successfully hedges 70% (up from 0%) in the provided hedging set task. | | Factuality (Inaccurate Rate) | Attribution (AIS) | Hedging (Accuracy) | |-------------------------------------------------|--------------------------------|----------------------|----------------------| | Gemini API Pro No factuality-focused adaptation | 6.7% [5.8%, 7.8%] | 40.2% [37.9%, 42.5%] | 0% | | Gemini API Pro Final stage of post-training | 3.8% [3.1%, 4.8%] | 60.0% [57.6%, 62.1%] | 69.3% | Table 6 | Factuality mitigations: Impact of post-training on the rate of inaccuracy, presence of attribution and the rate of accurate hedging on Gemini API Pro (with corresponding 95% confidence intervals). ## 5.1.7. Complex Reasoning Systems Gemini models can also be combined with additional techniques such as search and tool-use to create powerful reasoning systems that can tackle more complex multi-step problems. One example of such a system is AlphaCode 2, a new state-of-the-art agent that excels at solving competitive programming problems (Leblond et al, 2023). AlphaCode 2 uses a specialized version of Gemini Pro - tuned on competitive programming data similar to the data used in Li et al. (2022) - to conduct a massive search over the space of possible programs. This is followed by a tailored filtering, clustering and reranking mechanism. Gemini Pro is fine-tuned both to be a coding model to generate proposal solution candidates, and to be a reward model that is leveraged to recognize and extract the most promising code candidates. AlphaCode 2 is evaluated on Codeforces, 4 the same platform as AlphaCode, on 12 contests from division 1 and 2, for a total of 77 problems. AlphaCode 2 solved 43% of these competition problems, a 1.7x improvement over the prior record-setting AlphaCode system which solved 25%. Mapping this to competition rankings, AlphaCode 2 built on top of Gemini Pro sits at an estimated 85th percentile on average - i.e. it performs better than 85% of entrants. This is a significant advance over AlphaCode, which only outperformed 50% of competitors. The composition of powerful pre-trained models with search and reasoning mechanisms is an exciting direction towards more general agents; another key ingredient is deep understanding across a range of modalities which we discuss in the next section. 4 http://codeforces.com/ ## 5.2. Multimodal Gemini models are natively multimodal. These models exhibit the unique ability to seamlessly combine their capabilities across modalities (e.g. extracting information and spatial layout out of a table, a chart, or a figure) with the strong reasoning capabilities of a language model (e.g. its state-of-art-performance in math and coding) as seen in examples in Figures 5 and 14. The models also show strong performance in discerning fine-grained details in inputs, aggregating context across space and time, and applying these capabilities over a temporally-related sequence of video frames and/or audio inputs. The sections below provide more detailed evaluation of the model across different modalities (image, video, and audio), together with qualitative examples of the model's capabilities for image generation and the ability to combine information across different modalities. ## 5.2.1. Image Understanding We evaluate post-trained Gemini API models on four different capabilities: high-level object recognition using captioning or question-answering tasks such as VQAv2; fine-grained transcription using tasks such as TextVQA and DocVQA requiring the model to recognize low-level details; chart understanding requiring spatial understanding of input layout using ChartQA and InfographicVQA tasks; and multimodal reasoning using tasks such as Ai2D, MathVista and MMMU. For zero-shot QA evaluation, the model is instructed to provide short answers aligned with the specific benchmark. All numbers are obtained using greedy sampling and without any use of external OCR tools. Table 7 | Image understanding Gemini Ultra consistently outperforms existing approaches even in zero-shot, especially for OCR-related image understanding tasks for natural images, text, documents, and figures without using any external OCR engine ('pixel only'). Many existing approaches fine-tune on the respective tasks, highlighted in gray, which makes the comparison with 0-shot not apples-toapples. | | Gemini Ultra (pixel only) | Gemini Pro (pixel only) | Gemini Nano 2 (pixel only) | Gemini Nano 1 (pixel only) | GPT-4V | Prior SOTA | |-----------------------------------------------------------------------|-----------------------------|---------------------------|------------------------------|------------------------------|--------------------|----------------------------------------------------| | MMMU (val) Multi-discipline college-level problems (Yue et al., 2023) | 59.4% pass@1 62.4% Maj1@32 | 47.9% | 32.6% | 26.3% | 56.8% | 56.8% GPT-4V, 0-shot | | TextVQA (val) Text reading on natural images (Singh et al., 2019) | 82.3% | 74.6% | 65.9% | 62.5% | 78.0% | 79.5% Google PaLI-3, fine-tuned | | DocVQA (test) Document understanding (Mathew et al., 2021) | 90.9% | 88.1% | 74.3% | 72.2% | 88.4% (pixel only) | 88.4% GPT-4V, 0-shot | | ChartQA (test) Chart understanding (Masry et al., 2022) | 80.8% | 74.1% | 51.9% | 53.6% | 78.5% (4-shot CoT) | 79.3% Google DePlot, 1-shot PoT (Liu et al., 2023) | | InfographicVQA (test) Infographic understanding (Mathew et al., 2022) | 80.3% | 75.2% | 54.5% | 51.1% | 75.1% (pixel only) | 75.1% GPT-4V, 0-shot | | MathVista (testmini) Mathematical reasoning (Lu et al., 2023) | 53.0% | 45.2% | 30.6% | 27.3% | 49.9% | 49.9% GPT-4V, 0-shot | | AI2D (test) Science diagrams (Kembhavi et al., 2016) | 79.5% | 73.9% | 51.0% | 37.9% | 78.2% | 81.4% Google PaLI-X, fine-tuned | | VQAv2 (test-dev) Natural image understanding (Goyal et al., 2017) | 77.8% | 71.2% | 67.5% | 62.7% | 77.2% | 86.1% Google PaLI-X, fine-tuned | We find that Gemini Ultra is state of the art across a wide range of image-understanding benchmarks in Table 7. It achieves strong performance across a diverse set of tasks such as answering questions on natural images and scanned documents as well as understanding infographics, charts and science diagrams. When compared against publicly reported results from other models (most notably GPT-4V), the Gemini model is better in zero-shot evaluation by a significant margin. It also exceeds several existing models that are specifically fine-tuned on the benchmark's training sets for the majority of tasks. The capabilities of the Gemini models lead to significant improvements in the state of the art on academic benchmarks like MathVista (+3.1%) 5 or InfographicVQA (+5.2%). MMMU(Yue et al., 2023) is a recently released evaluation benchmark, which consists of questions about images across 6 disciplines with multiple subjects within each discipline that require collegelevel knowledge to solve these questions. Gemini Ultra achieves the best score on this benchmark advancing the state-of-the-art result by more than 5 percentage points and outperforms the previous best result in 5 of 6 disciplines (see Table 8), thus showcasing its multimodal reasoning capabilities. Table 8 | Gemini Ultra performance on the MMMU benchmark (Yue et al., 2023) per discipline. Each discipline covers multiple subjects, requiring college-level knowledge and complex reasoning. | MMMU (val) | Gemini Ultra (0-shot) Maj@32 | pass@1 | GPT-4V (0-shot) pass@1 | |-----------------------------|--------------------------------|----------|--------------------------| | Art & Design | 74.2 | 70 | 65.8 | | Business | 62.7 | 56.7 | 59.3 | | Science | 49.3 | 48 | 54.7 | | Health & Medicine | 71.3 | 67.3 | 64.7 | | Humanities & Social Science | 78.3 | 78.3 | 72.5 | | Technology & Engineering | 53 | 47.1 | 36.7 | | Overall | 62.4 | 59.4 | 56.8 | Gemini models are also capable of operating across modalities and a diverse set of global languages simultaneously, both for image understanding tasks (e.g., images containing text in Icelandic) and for generation tasks (e.g., generating image descriptions for a wide range of languages). We evaluate the performance of generating image descriptions on a selected subset of languages in the Crossmodal3600 (XM-3600) benchmark in a 4-shot setting, using the Flamingo evaluation protocol (Alayrac et al., 2022), without any fine-tuning for all models. As shown in Table 9, Gemini models achieve a significant improvement over the existing best model, Google PaLI-X. Table 9 | Multilingual image understanding Gemini models outperform existing models in captioning images in many languages when benchmarked on a subset of languages in XM-3600 dataset (Thapliyal et al., 2022). | XM-3600 (CIDER) | Gemini Ultra 4-shot | Gemini Pro 4-shot | Google PaLI-X 4-shot | |-------------------|-----------------------|---------------------|------------------------| | English | 86.4 | 87.1 | 77.8 | | French | 77.9 | 76.7 | 62.5 | | Hindi | 31.1 | 29.8 | 22.2 | | Modern Hebrew | 54.5 | 52.6 | 38.7 | | Romanian | 39 | 37.7 | 30.2 | | Thai | 86.7 | 77 | 56 | | Chinese | 33.3 | 30.2 | 27.7 | | Average (of 7) | 58.4 | 55.9 | 45 | 5 MathVista is a comprehensive mathematical reasoning benchmark consisting of 28 previously published multimodal datasets and three newly created datasets. Our MathVista results were obtained by running the MathVista authors' evaluation script. ## Prompt Write code torearrange the subplots in thefigureusing the latest version tangentfunctiononthebottomright.Fortheremainingtwosubplots, one of them should stay in its original position and the other should fill thelastspot.Firstdescribewhateachsubplotdepictsandidentifyits currentlocation.Then,explainwhereeachsubplotshouldgoinitsnew location.Lastwritethefullcodefortherearrangedversionwiththe original color scheme. <details> <summary>Image 6 Details</summary>  ### Visual Description \n ## Charts: Four Mathematical Function Plots ### Overview The image presents four separate 2D and 3D plots of mathematical functions. Each plot displays a different function, with varying scales and characteristics. There are no explicit labels or legends provided within the image itself. The plots are arranged in a 2x2 grid. ### Components/Axes * **Top-Left Plot:** X-axis ranges from approximately 0 to 10, Y-axis ranges from approximately -3 to 2. * **Top-Right Plot:** X-axis ranges from approximately 0 to 5, Y-axis ranges from approximately -3 to 2. * **Bottom-Left Plot:** X-axis ranges from approximately 0 to 10, Y-axis ranges from approximately 0 to 20000. * **Bottom-Right Plot:** X and Y axes both range from approximately 0 to 1. Z-axis ranges from approximately 0 to 15. This is a 3D surface plot. ### Detailed Analysis or Content Details **Top-Left Plot:** The plot shows a periodic function resembling a cosine wave. The function completes approximately one and a half cycles within the displayed range. * At x = 0, y ≈ 1.5 * At x = 2.5, y ≈ -2.5 (minimum) * At x = 5, y ≈ 1.5 * At x = 7.5, y ≈ -2.5 (minimum) * At x = 10, y ≈ 1.5 **Top-Right Plot:** This plot displays a function with vertical asymptotes. It appears to be a periodic function with singularities. * There are vertical asymptotes near x = 1, x = 3, and x = 5. * Between asymptotes, the function oscillates rapidly between positive and negative values. * At x = 0.5, y ≈ -1.5 * At x = 1.5, y ≈ 1.5 * At x = 2.5, y ≈ -1.5 * At x = 4.5, y ≈ 1.5 **Bottom-Left Plot:** This plot shows a rapidly increasing function, likely exponential or polynomial of high degree. * At x = 0, y ≈ 1 * At x = 2, y ≈ 100 * At x = 4, y ≈ 1000 * At x = 6, y ≈ 5000 * At x = 8, y ≈ 15000 * At x = 10, y ≈ 20000 **Bottom-Right Plot:** This is a 3D surface plot. The color gradient indicates the value of the function. * The surface is curved, rising from left to right and from bottom to top. * The lowest values (dark purple) are near x = 0 and y = 0. * The highest values (yellow/green) are near x = 1 and y = 1. * The function appears to be increasing in both x and y directions. ### Key Observations * The top-left plot demonstrates a smooth, periodic behavior. * The top-right plot exhibits a more complex, discontinuous behavior with asymptotes. * The bottom-left plot shows exponential growth. * The bottom-right plot visualizes a function of two variables, showing a curved surface. * There is no common scale or units indicated for any of the axes. ### Interpretation The image presents a visual comparison of different mathematical functions. The functions chosen represent a range of behaviors, from simple periodic oscillations to exponential growth and more complex, discontinuous patterns. The 3D plot adds a dimension of complexity, illustrating a function of two variables. The lack of labels makes it difficult to determine the specific functions being plotted, but the visual characteristics provide clues about their nature. The plots could be used to illustrate concepts in calculus, analysis, or numerical methods. The exponential growth in the bottom-left plot could represent population growth, compound interest, or other phenomena that exhibit similar behavior. The asymptotes in the top-right plot could represent physical limitations or singularities in a system. The 3D plot could represent a surface in physics, engineering, or computer graphics. </details> <details> <summary>Image 7 Details</summary>  ### Visual Description ## Charts/Graphs: Rendered Code - Mathematical Function Visualizations ### Overview The image displays four separate 2D and 3D charts visualizing mathematical functions. The charts appear to be generated from code, as indicated by the "Rendered code" title at the top-left corner. The charts showcase different function types, including a surface plot, a sine wave, an exponential curve, and a function with vertical asymptotes. ### Components/Axes * **Chart 1 (Top-Left):** 3D Surface Plot. X-axis ranges from approximately 0.0 to 1.0. Y-axis ranges from approximately 0.0 to 1.0. Z-axis (height) ranges from approximately 0.0 to 2.0. The surface is color-coded, with cooler colors (purple/blue) representing lower values and warmer colors (yellow/green) representing higher values. * **Chart 2 (Top-Right):** 2D Line Plot. X-axis ranges from 0 to 10. Y-axis ranges from -1.0 to 2.0. The plot shows a sinusoidal wave. * **Chart 3 (Bottom-Left):** 2D Line Plot. X-axis ranges from 0 to 10. Y-axis ranges from 0 to 20000. The plot shows an exponential curve. * **Chart 4 (Bottom-Right):** 2D Line Plot. X-axis ranges from 0 to 10. Y-axis ranges from -40 to 0. The plot shows a function with vertical asymptotes. ### Detailed Analysis or Content Details * **Chart 1 (3D Surface Plot):** The surface appears to be defined by a function where both x and y contribute positively to the z-value. The surface is relatively flat near x=0 and y=0, and rises more steeply as x and y increase. The maximum z-value is approximately 2.0, occurring near x=1.0 and y=1.0. * **Chart 2 (Sine Wave):** The line oscillates between approximately -1.0 and 2.0. The period of the wave is approximately 6.3. * At x = 0, y ≈ 0.0 * At x = 2, y ≈ 1.8 * At x = 4, y ≈ 0.0 * At x = 6, y ≈ -1.8 * At x = 8, y ≈ 0.0 * At x = 10, y ≈ 1.8 * **Chart 3 (Exponential Curve):** The line starts near y=0 at x=0 and increases rapidly as x increases. * At x = 0, y ≈ 100 * At x = 2, y ≈ 2000 * At x = 4, y ≈ 5000 * At x = 6, y ≈ 10000 * At x = 8, y ≈ 15000 * At x = 10, y ≈ 20000 * **Chart 4 (Asymptotes):** The line has vertical asymptotes at approximately x=2, x=6, and x=10. Between the asymptotes, the function appears to be negative. * At x = 0, y ≈ -10 * At x = 1, y ≈ -20 * At x = 3, y ≈ 10 * At x = 5, y ≈ 10 * At x = 7, y ≈ 10 * At x = 9, y ≈ 10 ### Key Observations * The charts demonstrate a variety of mathematical functions. * Chart 3 exhibits exponential growth, while Chart 4 shows a function with discontinuities. * Chart 2 is a classic sinusoidal function. * Chart 1 provides a visual representation of a function of two variables. ### Interpretation The image serves as a visual demonstration of different mathematical functions, likely generated as a result of running code. The functions chosen represent common mathematical concepts: a surface, a periodic function, an exponential function, and a function with asymptotes. The purpose is likely to illustrate the behavior of these functions or to test the code that generates them. The exponential growth in Chart 3 is particularly noteworthy, as it demonstrates how quickly a function can increase with increasing input. The asymptotes in Chart 4 highlight the concept of undefined values and limits. The 3D surface plot in Chart 1 provides a visual representation of a function of two variables, which is more complex than a simple 2D plot. The overall arrangement suggests a comparison of different function types and their graphical representations. </details> <details> <summary>Image 8 Details</summary>  ### Visual Description Icon/Small Image (14x14) </details> ## Response The currentsubplotsare: - ·Top left:A sine wave - ·Bottom left: An exponential function - ·Top right:A tangent function - ·Bottom right:A 3D paraboloid Thenewsubplotsshould be: - ·Top left:The 3D paraboloid - ·Top right:The sinewave - ·Bottomleft:Theexponentialfunction - ·Bottom right:The tangent function ``` top right: The sine wave bottom left: The exponential function bottom right: The tangent function ``` Figure 5 | Using Gemini models' multimodal reasoning capabilities to generate matplotlib code for rearranging the subplots. The multimodal prompt is shown at the top-left in gray. Gemini Ultra's response, including its generated code, is shown in the right column in blue. The bottom left figure shows rendered version of the generated code. Successfully solving this task shows the model's capability to combine several capabilities: (1) recognition of the functions depicted in the plots; (2) inverse graphics to infer the code that would have generated the subplots; (3) instruction-following to put subplots in their desired positions; and (4) abstract reasoning to infer that the exponential plot must stay in its original place, because the sine plot must move out of the way for the 3-dimensional plot. Qualitative evaluation in Figure 5 illustrates an example of Gemini Ultra's multimodal reasoning capabilities. The model is required to solve the task of generating matplotlib code that would rearrange a set of subplots provided by the user. The model output shows that it successfully solves this task combining multiple capabilities of understanding the user plot, inferring the code required to generate it, following user instructions to put subplots in their desired positions, and abstract reasoning about the output plot. This highlights Gemini Ultra's native multimodality and alludes to its more complex reasoning abilities across interleaved sequences of image and text. We refer the reader to the appendix for more qualitative examples. ## 5.2.2. Video Understanding Understanding video input is an important step towards a useful generalist agent. We measure the video understanding capability across several established benchmarks that are held-out from training. These tasks measure whether the model is able to understand and reason over a temporally-related sequence of frames. For each video task, we sample 16 equally-spaced frames from each video clip and feed them to the Gemini models. For the YouTube video datasets (all datasets except NextQA and the Perception test), we evaluate the Gemini models on videos that were still publicly available in the month of November, 2023. Gemini Ultra achieves state-of-the-art performance on various few-shot video captioning tasks as well as zero-shot video question answering tasks as shown in Table 10. This demonstrates its capability of strong temporal reasoning across several frames. Figure 23 in the appendix provides a qualitative example of understanding the video of the ball-striking mechanics of a soccer player and reasoning about the player can improve their game. Table 10 | Few-shot video understanding across tasks and languages on selected academic benchmarks. The reported metric is CIDER for video captioning, WUPS for NextQA, and top-1 accuracy for the Perception Test and ActivityNet-QA. For ActivityNet-QA, we use the Video-LLAVA (Lin et al., 2023) evaluation protocol. | Task | Gemini Ultra | Gemini Pro | Few-shot SoTA | |------------------------------------------------------|----------------|--------------|----------------------------------| | VATEX (test) | 62.7 | 57.4 | 56.0 | | English video captioning (Wang et al., 2019) | 4-shots | 4-shots | DeepMind Flamingo, 4-shots | | VATEX ZH (test) | 51.3 | 50.0 | - | | Chinese video captioning (Wang et al., 2019) | 4-shots | 4-shots | | | YouCook2 (val) | 135.4 | 123.2 | 74.5 | | English cooking video captioning (Zhou et al., 2018) | 4-shots | 4-shots | DeepMind Flamingo, 4-shots | | NextQA (test) | 29.9 | 28.0 | 26.7 | | Video question answering (Xiao et al., 2021) | 0-shot | 0-shot | DeepMind Flamingo, 0-shot | | ActivityNet-QA (test) | 52.2 | 49.8 | 45.3 | | Video question answering (Yu et al., 2019) | 0-shot | 0-shot | Video-LLAVA, 0-shot | | Perception Test MCQA (test) | 54.7 | 51.1 | 46.3 | | Video question answering (Pătrăucean et al., 2023) | 0-shot | 0-shot | SeViLA (Yu et al., 2023), 0-shot | ## 5.2.3. Image Generation Gemini models are able to output images natively, without having to rely on an intermediate natural language description that can bottleneck the model's ability to express images. This uniquely enables the model to generate images with prompts using interleaved sequences of image and text in a few-shot setting. For example, the user might prompt the model to design suggestions of images and text for a blog post or a website (see Figure 12 in the appendix). Figure 6 shows an example of image generation in 1-shot setting. Gemini Ultra model is prompted with one example of interleaved image and text where the user provides two colors (blue and yellow) and image suggestions of creating a cute blue cat or a blue dog with yellow ear from yarn. The model is then given two new colors (pink and green) and asked for two ideas about what to create using these colors. The model successfully generates an interleaved sequence of images and text with suggestions to create a cute green avocado with pink seed or a green bunny with pink ears from yarn. Figure 6 | Image Generation. Gemini models can output multiple images interleaved with text given a prompt composed of image and text. In the left figure, Gemini Ultra is prompted in a 1-shot setting with a user example of generating suggestions of creating cat and dog from yarn when given two colors, blue and yellow. Then, the model is prompted to generate creative suggestions with two new colors, pink and green, and it generates images of creative suggestions to make a cute green avocado with pink seed or a green bunny with pink ears from yarn as shown in the right figure. <details> <summary>Image 9 Details</summary>  ### Visual Description \n ## Screenshot: Conversational AI Exchange - Crafting Ideas ### Overview The image is a screenshot of a conversational AI interface, likely a chatbot or similar application. It displays a series of prompts and responses related to generating craft ideas based on yarn colors. The interface is divided into two main columns: "Prompt" on the left and "Response" on the right. Each section contains multiple exchanges. ### Components/Axes The screenshot consists of several text blocks and accompanying images. There are no axes or scales present, as it's a conversational interface, not a chart or graph. The layout is structured as a series of question-answer pairs. ### Detailed Analysis or Content Details **Left Column - Prompt:** 1. "Give me two ideas that I could do with these 2 colors" 2. "Colors: I see blue and yellow yarn" 3. "How about a cute blue cat?" 4. "Or a blue dog that would also have a yellow ear?" 5. "Give me two ideas that I could do with these 2 colors" **Right Column - Response:** 1. "Colors: I see green and pink yarn" 2. "Idea 1: How about a green avocado with pink seed?" (accompanied by an image of an avocado) 3. "Idea 2: Or a green bunny with pink ears?" (accompanied by an image of a bunny) **Images:** * A heart-shaped object made of yarn. * A blue cat made of yarn. * A blue dog with a yellow ear made of yarn. * A green, textured object resembling a plant or coral. * An avocado with a pink seed. * A green bunny with pink ears. ### Key Observations The AI appears to be responding to color prompts by suggesting craft ideas. The responses are creative and visually represented with images. The AI is capable of understanding color combinations and generating relevant suggestions. The prompts and responses are conversational in nature. ### Interpretation This screenshot demonstrates a simple application of AI in a creative context. The AI acts as a brainstorming partner, offering ideas based on user input. The inclusion of images enhances the user experience and provides visual inspiration. The interaction highlights the potential of AI to assist with creative tasks and generate novel ideas. The AI seems to be able to interpret the user's intent (requesting craft ideas) and respond appropriately. The system is designed for a casual, conversational interaction. There is no quantitative data to analyze, but the qualitative data suggests a functional and user-friendly AI application. </details> ## 5.2.4. Audio Understanding We evaluate the Gemini Nano-1 and Gemini Pro models on a variety of public benchmarks and compare it with Universal Speech Model (USM) (Zhang et al., 2023) and Whisper (large-v2 (Radford et al., 2023) or large-v3 (OpenAI, 2023) as indicated). These benchmarks include automatic speech recognition (ASR) tasks such as FLEURS (Conneau et al., 2023), VoxPopuli, (Wang et al., 2021), Multi-lingual Librispeech (Pratap et al., 2020), as well as the speech translation task CoVoST 2, translating different languages into English (Wang et al., 2020). We also report on an internal benchmark YouTube test set. ASR tasks report a word error rate (WER) metric, where a lower number is better. Translation tasks report a BiLingual Evaluation Understudy (BLEU) score, where a higher number is better. FLEURS is reported on 62 languages that have language overlap with the training data. Four segmented languages (Mandarin, Japanese, Korean and Thai) report character error rate (CER), instead of WER, similar to Whisper (Radford et al., 2023). Table 11 indicates that our Gemini Pro model significantly outperforms the USM and Whisper models across all ASR and AST tasks, both for English and multilingual test sets. Note that there is a large gain in FLEURS, compared to USM and Whisper, as our model is also trained with the FLEURS training dataset. However, training the same model without FLEURS dataset results in a WER of 15.8, which still outperforms Whisper. Gemini Nano-1 model also outperforms both USM and Whisper on all datasets except FLEURS. Note that we did not evaluate Gemini Ultra on audio yet, though we expect better performance from increased model scale. | | Task | Metric | Gemini Pro | Gemini Nano-1 | Whisper (OpenAI, 2023; Radford et al., 2023) | USM (Zhang 2023) | |------------------------------|--------------------------------------------------------|----------|--------------|-----------------|------------------------------------------------|--------------------| | Automatic Speech Recognition | YouTube (en-us) | WER (↓) | 4.9% | 5.5% | 6.5% (v3) | 6.2% | | | Multilingual Librispeech (en-us) (Pratap et al., 2020) | WER (↓) | 4.8% | 5.9% | 6.2% (v2) | 7.0% | | | FLEURS (62 lang) (Conneau et al., 2023) | WER (↓) | 7.6% | 14.2% | 17.6% (v3) | 11.8% | | | VoxPopuli (14 lang) (Wang et al., 2021) | WER (↓) | 9.1% | 9.5% | 15.9% (v2) | 13.4% | | Automatic Speech Translation | CoVoST 2 (21 lang) (Wang et al., 2020) | BLEU (↑) | 40.1 | 35.4 | 29.1 (v2) | 30.7 | Table 11 | Speech evaluation results on selected benchmarks for ASR and AST. For ASR, the reported metric is WER where lower is better. For AST, the reported metric is BLEU where higher is better. Table 12 | Qualitative examples for the ASR task in the benchmark. Incorrect transcriptions are highlighted in red. <details> <summary>Image 10 Details</summary>  ### Visual Description Icon/Small Image (30x21) </details> <details> <summary>Image 11 Details</summary>  ### Visual Description Icon/Small Image (30x23) </details> Table 12 shows further error analysis with USM and Gemini Pro. We find that Gemini Pro produces more understandable responses, particularly on rare words and proper nouns. | Domain | Truth | USM | Gemini Pro | Wav | |----------|------------------------------------------------------------------------------------------------------------------------------------|-------------------------------------------------------------------------------------------------------------------|------------------------------------------------------------------------------------------------------------------------------------|---------| | Fleurs | Scotturb bus 403 travels regularly to Sintra, stopping at Cabo da Roca. | Scotboard bus four3 traversed regu- larly to Centra stopping at Cabo de Roga. | Scotturb bus 403 travels regularly to Sintra, stopping at Cabo da Roca. | ▲ ) ) ) | | Fleurs | The archipelago lies 120 km north of the Peninsula. The largest is King George Island, with the settlement of Villa Las Estrellas. | The archipelago lines 120 km north of peninsula. The largest is Kingurch island with the settlement of Cua Losas. | The archipelago lies 120 km north of the Peninsula. The largest is King George Island, with the settlement of Villa Las Estrellas. | ▲ ) ) ) | ## 5.2.5. Modality Combination Multimodal demonstrations often include a combination of text interleaved with a single modality, usually images. We demonstrate the ability to process a sequence of audio and images natively. Consider a cooking scenario about making an omelet where we prompt the model with a sequence of audio and images. Table 13 indicates a turn-by-turn interaction with the model, providing pictures and verbally asking questions about the next steps for cooking an omelet. We note that the model response text is reasonably accurate, and shows that model processes fine-grained image details to evaluate when the omelet is fully cooked. See demo on the website. Table 13 | Audio-visual qualitative example showcasing the ability of Gemini models to process interleaved sequences of text, vision, and audio, as well as reason across modalities. This example inputs interleaved images and audio from the user in a cooking scenario. The user prompts the model for instructions to make an omelet and to inspect whether it is fully cooked. <details> <summary>Image 12 Details</summary>  ### Visual Description \n ## Screenshot: Veggie Omelet Recipe & Interaction ### Overview This document presents a series of screenshots depicting a cooking process – making a veggie omelet – alongside transcribed audio of a user interacting with instructions. The interaction appears to be a conversational AI guiding the user through the recipe. ### Components/Axes The document consists of four sections: 1. Ingredients: A photo of eggs, cherry tomatoes, and other vegetables. 2. Omelet in Pan (Stage 1): An omelet partially cooked in a frying pan. 3. Follow-up Question: No image, only transcribed audio. 4. Omelet in Pan (Stage 2): A fully cooked omelet in a frying pan. ### Detailed Analysis or Content Details **Section 1: Ingredients** * **Image:** Shows a plastic bag containing several white eggs, a handful of red cherry tomatoes, and some green vegetables (likely spinach or herbs). * **Audio:** "What's the first step to make a veggie omelet with these ingredients?" * **Response:** "Crack the eggs into a bowl and whisk them." **Section 2: Omelet in Pan (Stage 1)** * **Image:** Shows an omelet in a black frying pan. The omelet is partially cooked, with some areas still appearing liquid or runny. Some green vegetables are visible within the omelet. * **Audio:** "Thank you for the instructions. I started making my omelet, does it look ready now?" * **Response:** "It looks like it's almost ready. You can flip it over to cook the other side." **Section 3: Follow-up Question** * **Audio:** "Why is it not ready?" * **Response:** "It's not ready because the eggs are still runny." **Section 4: Omelet in Pan (Stage 2)** * **Image:** Shows a fully cooked omelet in a black frying pan. The omelet is golden brown and appears set throughout. Green vegetables are visible within the omelet. * **Audio:** "What about now?" * **Response:** "It looks ready now. You can take it off the heat and serve it." ### Key Observations The interaction demonstrates a conversational AI providing real-time feedback on a cooking process based on visual input (implied, as the AI is responding to the user's description of the omelet's appearance). The AI correctly identifies the stages of cooking (not ready, almost ready, ready) based on the state of the eggs. ### Interpretation This document showcases a practical application of AI in assisting with everyday tasks. The AI isn't simply reciting a recipe; it's actively interpreting the user's progress and providing tailored guidance. The system likely uses image recognition to assess the omelet's doneness, although this is not explicitly stated. The interaction highlights the potential for AI to personalize learning experiences and provide support in real-time, making complex tasks more accessible. The system is able to understand the context of the question and provide a relevant answer. The progression of the images and responses demonstrates a successful completion of the cooking process with AI assistance. </details> ## 6. Post-Training Models After large-scale pre-training, we apply post-training , where one trains on top of a pre-trained model in order to extend the model's proficiency and to enable a wide variety of capabilities. Namely, we seek to improve overall quality, enhance target capabilities such as coding and multilingual, and ensure alignment and safety criteria are met. We discuss our approach to post-training in this section, highlighting common and distinct aspects of the Gemini Apps and Gemini API model variants. ## 6.1. Gemini Apps: Gemini and Gemini Advanced Gemini and Gemini Advanced offer direct access to Google's family of AI models, consisting of the core post-trained Gemini Apps models and the system around it. These models are created by applying specialized post-training on top of Gemini pre-trained models: currently, Gemini gives access to Pro 1.0 and Gemini Advanced gives access to Ultra 1.0. Beyond the core models, the system determines how the models interact with external tools (such as Google Flights, Maps, and Google Workspace), and how to generate responses (filtering, ranking, and streaming). As an area, conversational AI presents several challenges, including: How to understand users' requests across multi-turn interactions? How to make sure responses are safe, factually grounded, and helpful? How to help users accomplish tasks by using tools external to the models? We discuss how we approach these challenges in the following sections. ## 6.2. Gemini APIs: Google AI Studio and Cloud Vertex AI Our developer-focused Gemini API models are designed to support both conversational and nonconversational use cases. These models are available through Google AI Studio and Cloud Vertex AI through an easy to use API. Google AI Studio is a free, web-based developer tool to prototype and launch apps quickly with an API key. Vertex AI is a comprehensive AI platform that enables developers to leverage Gemini API models with varied tooling, fully-managed infrastructure, and built-in enterprise security and privacy settings. Gemini APIs make it easy to integrate Gemini API models into any production product or workflow, empowering developers to build applications that can reason across different modalities. ## 6.3. Post-Training Methods & Data Post-training Gemini models to produce Gemini API and Apps variants involves several stages; see Figure 7. Careful data curation is critical for all stages. First, we collect a diverse set of prompts that are representative of real-world use cases. Second, we apply supervised fine-tuning (SFT) on demonstration data of what the model's output should be for a given prompt (Mishra et al., 2021; Ouyang et al., 2022; Wei et al., 2022a). Third, we further collect different possible responses to a given prompt, and collect feedback data over these to train a Reward Model (RM). Finally, using the trained RM, a Reinforcement Learning from Human Feedback (RLHF) stage (Bai et al., 2022a) is applied to further align the model's outputs with human preferences. We discuss our methods in more detail below: (1) Prompt Data Collection : A prompt is a user's input to the model. As well as the most recent user input, this can also include previous user-model interactions. We curate datasets of target prompts. The datasets serve as the basis for our demonstration and feedback data collections, and they are used directly during reinforcement learning. It is important to cover a diverse set of crucial use cases and in both single-turn and multi-turn formats. Data sources include vendor-created data, third-party licensed sources, and synthetic approaches. - (2) SFT on Demonstration Data : SFT trains the model to output a desired target response given a prompt. Our Demonstration Data target responses can be directly written by a human expert, or generated by a model and in some cases revised or reviewed by a human. Additionally, we use data analysis tools and heuristics to ensure high data diversity across capabilities, use cases, and semantic clusters. - (3) RM Training on Feedback Data : We further collect Feedback Data, for which human raters provide feedback such as relative preferences over candidate responses and feedback regarding individual responses to a given prompt. For many capabilities, rating relative preferences is an easier task than demonstrating an ideal response. Feedback data are collected across creativity, safety, factuality, other capabilities, and other target criteria. We found that the utility of the resulting human feedback data greatly depends on the prompt selection and the sampling strategy used to produce candidate responses. We use this data to train RMs to output rewards that align with human preferences as closely as possible. (4) RLHF : Applying reinforcement learning from human feedback (RLHF) to our models provides further gains over SFT alone. Our approach creates an iterative process in which RL continually pushes the boundaries of the RM, while the RM is continuously improved through evaluation and data collection, leading to progressive improvements in both. Figure 7 | Modeling overview. Post-training utilizes an optimized data flywheel in order to acquire human-AI feedback and continually improve on key areas. The data mixtures for supervised finetuning, reward modeling, and reinforcement learning serve as the foundation for our models. <details> <summary>Image 13 Details</summary>  ### Visual Description \n ## Diagram: Gemini Model Training and Feedback Loop ### Overview The image depicts a diagram illustrating the training and feedback loop for the Gemini model. It shows a cyclical process involving pre-training, supervised fine-tuning (SFT), reinforcement learning from human feedback (RLHF), and data collection from end-users. The diagram emphasizes the iterative nature of model improvement through data feedback. ### Components/Axes The diagram consists of the following components: * **Gemini pre-training:** A light blue rounded rectangle labeled "Gemini pre-training" positioned on the left side of the diagram. * **SFT:** A light blue rounded rectangle labeled "SFT" (Supervised Fine-Tuning) positioned in the upper-center of the diagram. * **RLHF:** A light blue rounded rectangle labeled "RLHF" (Reinforcement Learning from Human Feedback) positioned to the right of "SFT". * **Data flywheel:** A light blue circle labeled "Data flywheel" positioned in the center of the diagram. * **Demonstration data:** A light blue rounded rectangle labeled "Demonstration data" positioned at the bottom-left of the diagram. * **Feedback data:** A light blue rounded rectangle labeled "Feedback data" positioned at the bottom-right of the diagram. * **End users:** A light blue rounded rectangle labeled "End users" positioned on the right side of the diagram. Arrows indicate the flow of data and the relationships between these components. A dashed line encloses the SFT, RLHF, Data flywheel, Demonstration data, and Feedback data components, suggesting they form a core iterative loop. ### Detailed Analysis or Content Details The diagram illustrates a process flow as follows: 1. **Gemini pre-training** feeds into **SFT**. 2. **SFT** feeds into **RLHF**. 3. **RLHF** feeds into **End users**. 4. **End users** provide **Feedback data**. 5. **Feedback data** feeds into the **Data flywheel**. 6. **Demonstration data** also feeds into the **Data flywheel**. 7. The **Data flywheel** then feeds back into both **SFT** and **Demonstration data**. The arrows indicate a continuous loop where user feedback and demonstration data are used to refine the model through SFT and RLHF. The "Data flywheel" represents the accumulation and utilization of data for ongoing improvement. ### Key Observations The diagram highlights the importance of a feedback loop in model development. The cyclical nature of the process suggests that the model is continuously learning and improving based on real-world usage and human input. The inclusion of both "Demonstration data" and "Feedback data" indicates a multi-faceted approach to data collection. ### Interpretation This diagram illustrates a modern machine learning development paradigm, particularly common in large language models. The "Gemini pre-training" stage represents the initial training on a massive dataset. The subsequent SFT and RLHF stages refine the model's behavior to align with human preferences and instructions. The "Data flywheel" is a crucial element, representing the continuous collection and utilization of data generated by users. This data is then used to further improve the model, creating a positive feedback loop. The diagram suggests that the Gemini model is designed to be iteratively improved through real-world interaction and feedback, rather than being a static entity. The dashed line around the core loop emphasizes the self-improving nature of the system. The diagram does not contain any numerical data or specific metrics, but rather focuses on the conceptual flow of information and the relationships between different components. </details> ## 6.4. Evaluation Evaluation of human preferences over model outputs provides critical signals for measuring performance. As part of our development process, we conduct human evaluation extensively across targeted capabilities. Human evaluation is instantiated as side-by-side blind evaluations where human raters judge responses of two models to the same prompt, as single-response ratings for certain capabilities, and as online testing. In addition, we build models for automated evaluation that faithfully imitate human preferences in order to guide development and continuously monitor online performance. ## 6.5. Model Capabilities Beyond the general post-training outlined above, we apply techniques to improve a set of key capabilities. These capabilities cover a range of use cases inspired by current user needs and research-inspired future applications. We outline capability examples not detailed in previous sections below. The posttraining recipes are carefully designed to balance multiple objectives, including creativity, factuality, safety and more (Bai et al., 2022b; Thoppilan et al., 2022). We have a particular focus on safety and alignment, and hence address this in a further dedicated section. ## 6.5.1. Instruction Following Following a user's prompt accurately is a fundamental capability for LLMs, especially as these models become more sophisticated and are presented with increasingly complex user prompts. User prompts vary in granularity, specificity , and requirements (e.g., content, format, length). Individual instructions can also be ambiguous, optional, or even impossible or undesirable to satisfy (He et al., 2023; Xu et al., 2023). We improve Gemini Apps and Gemini API models' instruction following (IF) abilities by collecting data for a diverse set of instruction following categories. For instructions that are verifiable programmatically such as word count, we generate synthetic data via prompting and response editing to ensure that such instructions are satisfied. Complex prompts evaluation : We investigate performance on complex prompts containing multiple instructions using a fine-grained evaluation method that assesses how well models adhere to each instruction. Human raters are presented with a prompt-response pair and a list of the individual (sub)-instructions contained in the prompt. Each prompt may have anywhere from one to dozens of individual instructions, and the annotators are tasked with determining whether each instruction is followed (or not) by the response. Table 14 reports results on an internal dataset of prompts with instructions of varying complexity that encompass a wide range of instructions and are designed to be challenging for LLMs. We report two metrics: per-instruction accuracy (the percentage of sub instructions in the eval set that are followed), and full-response accuracy (the percentage of eval set prompts where all sub-instructions are followed). | | Post-trained PaLM 2 | Gemini (with Pro) | Gemini Advanced (with Ultra) | |--------------------------|-----------------------|---------------------|--------------------------------| | Per-instruction accuracy | 59.5 ± 3.0% | 77.8 ± 2.0% | 87.4 ± 1.4% | | Full-response accuracy | 25.5 ± 3.3% | 38.5 ± 3.6% | 54.1 ± 3.7% | Table 14 | Performance of Gemini on our complex prompts instruction-following internal benchmark. Gemini Advanced (with Ultra) achieves an average per-instruction accuracy close to 90%, representing a significant improvement over Gemini (with Pro) and a post-trained PaLM 2 model. We find that the sub-instructions that aren't followed are well-distributed across responses. As a result Gemini Advanced's full-response accuracy is lower, at around 54%. This indicates that there is further headroom for models to fully satisfy all instructions. ## 6.5.2. Tool Use By training LLMs to use tools, we greatly expand LLM capabilities beyond their internal knowledge. We treat tool use for both Gemini Apps and Gemini API models as a code generation problem, leveraging the base model's preexisting strong coding capabilities. Every tool invocation is represented as a code block in which tool calls are invoked. This process allows the model to both compose multiple tools in each code block, as well as observe and react to the results of tool execution. At inference time, to generate a response to a user prompt, our system executes the loop shown in Figure 8, where sampling from the LLM and execution of tool code work together to create a final response. Figure 8 | A Gemini tool-use control loop. <details> <summary>Image 14 Details</summary>  ### Visual Description \n ## Diagram: LLM Execution Flow ### Overview The image depicts a flowchart illustrating the process of executing a prompt with a Large Language Model (LLM), specifically focusing on whether the LLM's response contains tool code. The diagram shows a cyclical process where the LLM's output is re-integrated into the context if tool code is present. ### Components/Axes The diagram consists of four main components, represented as rounded rectangles connected by arrows: 1. **Prompt:** (Purple) - The initial input to the LLM. 2. **Sample from LLM:** (Light Blue) - The output generated by the LLM in response to the prompt. 3. **Sample contains tool code?:** (Diamond Shape, Green/Red) - A decision point evaluating whether the LLM's output includes tool code. 4. **Put execution result back in context:** (Light Blue) - A process that re-integrates the LLM's output into the context for further processing. 5. **Respond:** (Orange) - The final output or action taken when no tool code is present. Arrows indicate the flow of the process. A "YES" branch (Green) leads back to "Put execution result back in context", creating a loop. A "NO" branch (Red) leads to "Respond". ### Detailed Analysis or Content Details The diagram illustrates a conditional flow: 1. A "Prompt" is sent to the LLM. 2. The LLM generates a "Sample". 3. A check is performed to determine if the "Sample" contains "tool code". 4. If the "Sample" *does* contain tool code ("YES"), the "execution result" is fed back into the context, and the process repeats. 5. If the "Sample" *does not* contain tool code ("NO"), the system "Responds". There are no numerical values or specific data points in this diagram. It is a purely conceptual representation of a process. ### Key Observations The diagram highlights a feedback loop where the LLM's output is iteratively refined based on the presence of tool code. This suggests a system designed to handle tasks that require external tools or actions. The conditional branching indicates a decision-making process based on the content of the LLM's response. ### Interpretation This diagram represents a sophisticated LLM execution strategy. The presence of a feedback loop suggests the system is capable of complex reasoning and task completion. The check for "tool code" indicates the LLM can generate instructions for external tools, and the system is designed to execute those instructions and incorporate the results back into the LLM's context. This is a common pattern in agent-based LLM systems, where the LLM acts as a controller for various tools. The diagram implies a system that can dynamically adapt its behavior based on the LLM's output, allowing it to perform tasks that would be impossible with a simple prompt-response interaction. The system is designed to continue processing until the LLM's output does not require further tool execution, at which point a final response is generated. </details> Gemini Apps models : Gemini draws on a range of tools via Gemini Extensions, including Google Workspace, Google Maps, YouTube, Google Flights, and Google Hotels. These tool-use capabilities also enable Gemini to be integrated as part of Gmail, Docs, Slides, Sheets and more. We are aiming to bring further tool-use capabilities in order to both enhance Gemini models and integrate Gemini models into further products. We created an internal benchmark to assess Gemini performance on tasks that may benefit from access to these extensions. This benchmark measures human preference in domains such as travel planning and video discovery. We find models equipped with tools are preferred on this set 78% of the time over models without tools (excluding ties). Gemini API models : We have found that fine-tuning Gemini API models is very effective at teaching the model tool-use behaviors. Furthermore, training models to use programming and search as tools leads to improved performance on a range of academic benchmarks. In Table 15, we compare tool-use models fine-tuned from an early version of Gemini API Pro against equivalent models that do not use tools. Table 15 | Comparison between Gemini API tool-use models and comparable models that do not use tools. Gemini API Pro without tools is an early version of our Pro model trained without tool-use data. Gemini API Pro with tools is the same model fine-tuned with tool-use data. | | Mathematical Reasoning GSM8K Cobbe et al. (2021) MATH | Hendrycks et al. (2021b) | Factuality & Retrieval NQ Kwiatkowski et al. (2019b) Kasai | Knowledge Realtime QA et al. (2022a) | |------------------------------|---------------------------------------------------------|----------------------------|--------------------------------------------------------------|----------------------------------------| | Gemini API Pro with tools | 80.1% | 41.8% | 68.0% | 70.8% | | Gemini API Pro without tools | 69.7% | 30.7% | 59.0% | 39.2% | ## 6.5.3. Multilinguality Multilinguality is critical to make sure Gemini models effectively support a wide range of languages. We discuss our key approaches for Gemini Apps and Gemini API models respectively below. Gemini Apps models : Scaling Gemini from English to 40+ languages imposed research challenges in data quality. We leverage abundant high-quality English data by localization to native cultures (e.g., 'president of the United States' -> ' 日 本 の 首 相 '). Table 16 shows the performance of Gemini (with Pro) on 5 languages compared to Bard with an older post-training recipe and based on PaLM 2. For side-by-side comparisons between a model A and a model B, we calculate a metric called SxS score. Each rating is converted to an ordinal value centered at 0: ratings preferring A are positive and ratings preferring B are negative over a scale between -1.5 and 1.5. The converted values are averaged to return the SxS score. Intuitively, a positive SxS score indicates the extent to which model A is preferred over model B. Here, we find quality improved by more than 0.1 SxS score for all five languages. Coding and reasoning gains from Gemini Pro are preserved across languages. | Language | Quality SxS | Coding MBPP Pass@1 Austin et al. (2021) | Reasoning MMLU Hendrycks et (2021a) | |------------|---------------|-------------------------------------------|---------------------------------------| | ja-JP | 0.14 | +22.2% | +3.6% | | pt-BR | 0.17 | +23.2% | +5.2% | | de-DE | 0.1 | +21.4% | +7.5% | | es-419 | 0.12 | +22.8% | +9.3% | | it-IT | 0.13 | +13.8% | +7.5% | Table 16 | Multilingual performance of Gemini (with Pro) compared to Gemini with an older posttraining recipe and PaLM 2. Gemini API models : Similar to Gemini Apps models, we train Gemini API models on additional multilingual post-training data, effectively adapting the original English model for use in various languages. We experiment with both human-generated non-English prompt-response pairs as well as automatically translated pairs. For the latter, we leverage abundant high-quality English demonstration data by translation. We ensure the quality of such translated data by translationability filtering and response rating by humans. Translatability Filtering : Not all prompt-response pairs make sense when automatically translated, and may require expensive localization instead. Example prompts of this type (responses omitted for space) include: - (strict word requirements) Write a 1000 word essay about world peace. - (too English centric) Write a poem in iambic pentameter about apples. - (too Latin-script centric) What is a word with 1 E, 2 As, and 1 U? Translation Quality Validation : Each translated prompt-response pair was rated for translation quality by at least 3 human raters, and was kept in the final mixture if the majority of raters rated it as accurate. Section 5.1.4 reports evaluations of the multilingual capabilities of post-trained Gemini API models. ## 6.5.4. Multimodal Vision Multimodal post-training enhances the capabilities of our natively multimodal Gemini models for a wide range of useful applications. In the following, we discuss how image understanding ability is incorporated into Gemini Apps and Gemini API models. For this evaluation, we further train both of these Gemini model variants on a mixture of text data and expert curated image-text data over several vertically-defined multimodal use cases Gemini Apps models : We empower Gemini and Gemini Advanced with image understanding capabilities by fine-tuning pre-trained Gemini models on a mixture of text-only and image-text data. Careful balancing of text and multimodal data ensures the model develops robust image understanding without adversely affecting the quality of the text-only interactions. To assess our models, we compile a dataset of human-curated and synthetic image-text prompts and responses, spanning various categories and difficulty levels. This dataset facilitates human evaluation for model comparison and selection. We find that introducing this image-text data preserves Gemini Apps model quality on text-only tasks, with a SxS score on text-only tasks of +0.01 ± 0.01 for a Gemini Apps Pro model trained on this data versus an equivalent model trained only on text data. In addition, post-training via RLHF improves performance on multimodal tasks, with a SxS score on image-understanding tasks of +0.223 ± 0.06 for a Gemini Apps Pro model post-trained with SFT & RLHF vs SFT alone. Gemini API models : We evaluate the impact of post-training via SFT on Gemini API models' multimodal vision performance by tracking the performance of both pre-trained models and posttrained Gemini API Vision models on a series of standard benchmarks. These post-trained results have already been given in Table 7, in Table 17 we further report the difference in performance between pre-trained and post-trained Gemini API models. Table 17 | Post-trained model image understanding Post-training improves image understanding capabilities of Gemini API Ultra over the base pre-trained model. Comparisons of Gemini API Ultra to other models on these benchmarks are given in Table 7. | | Gemini Ultra Pre-trained only 0-shot (pixel only) | Gemini API Ultra 0-shot (pixel only) | Gemini Ultra pre- to post-trained improvement | |-----------------------------------------------------------------------|-----------------------------------------------------|----------------------------------------|-------------------------------------------------| | MMMU (val) Multi-discipline college-level problems (Yue et al., 2023) | n/a | 59.4% pass@1 62.4% | n/a | | TextVQA (val) Text reading on natural images (Singh et al., 2019) | 81.4% | 82.3% | +0.9% | | DocVQA (test) Document understanding (Mathew et al., 2021) | 90.1% | 90.9% | +0.8% | | ChartQA (test) Chart understanding (Masry et al., 2022) | 80.8% | 80.8% | 0.0% | | InfographicVQA (test) Infographic understanding (Mathew et al., 2022) | 77.9% | 80.3% | +2.4% | | MathVista (testmini) Mathematical reasoning (Lu et al., 2023) | n/a | 53.0% | n/a | | AI2D (test) Science diagrams (Kembhavi et al., 2016) | 76.6% | 79.5% | +2.9% | | VQAv2 (test-dev) Natural image understanding (Goyal et al., 2017) | 74.5% | 77.8% | +3.3% | The results indicate that the pre-trained model already has high performance across the capabilities represented by these benchmarks, in line with previous observations. However, the post-training SFT stage used for the Gemini API Vision models succeeds in improving the performance over several of these benchmarks (InfographicVQA, AI2D, VQAv2), most likely due to the model's increased instruction-following capabilities that succeed in aligning the model output style with that of the golden references. ## 6.5.5. Coding Despite the strong coding benchmark performance of the base model, post-training data still provides a significant boost to both code quality and code correctness. This highlights the benefit of high-quality demonstration data and feedback data for coding use cases. Gemini Apps and Gemini API models use a combination of human and synthetic approaches to collect such data. We evaluate our Gemini Apps models' coding performance on a set of internally curated prompts, distributed across code use cases and languages. Table 18 reports SxS scores, where Gemini (with Pro) significantly improves upon Bard with an older post-training recipe and based on PaLM 2. Gemini Advanced (with Ultra) further improves upon Gemini (with Pro). Table 18 | SxS comparisons of Gemini models on an internal coding benchmark. | Side A | Side B | SxS score | |------------------------------|---------------------------|-------------| | Gemini (with Pro) | Bard (PaLM 2, Sept. 2023) | 0.19 ± 0.03 | | Gemini Advanced (with Ultra) | Gemini (with Pro) | 0.13 ± 0.02 | For the coding capabilities of post-trained Gemini API Models, see Table 2 which reports their academic benchmark performance. ## 7. Responsible Deployment During the development of Gemini models, we follow a structured approach to responsible deployment to identify, measure, and manage foreseeable downstream societal impacts of our models, in line with previous releases of Google's AI technology (Kavukcuoglu et al., 2022). Throughout the lifecycle of a project, we follow the structure below. This section provides more detail about our approach and includes key findings where available. We are committed to ongoing transparency and will continue to provide updated information on our approach and testing in upcoming reports. <details> <summary>Image 15 Details</summary>  ### Visual Description \n ## Diagram: Responsibility Approach ### Overview The image depicts a linear, cyclical diagram illustrating a "Responsibility Approach". It outlines a five-stage process: Assessment, Policy, Evaluation, Mitigation, and Deployment, all underpinned by "Responsibility & Safety Governance". The diagram uses a horizontal, flowing blue bar with circular markers to represent each stage. Dashed arrows indicate a cyclical flow, returning from Deployment to Assessment. ### Components/Axes * **Title:** "Responsibility Approach" (Top-center) * **Stages:** Assessment, Policy, Evaluation, Mitigation, Deployment (Horizontally aligned within the blue bar) * **Underlying Framework:** "Responsibility & Safety Governance" (Below the blue bar) * **Flow Indicators:** Dashed arrows indicating cyclical progression. * **Stage Markers:** Small white circles within the blue bar, each labeled with a stage name. ### Detailed Analysis or Content Details The diagram presents a sequential process with a feedback loop. 1. **Assessment:** The first stage, marked by a solid circle. 2. **Policy:** The second stage, marked by a solid circle. 3. **Evaluation:** The third stage, marked by a circle with a lighter blue fill and a white circular highlight. 4. **Mitigation:** The fourth stage, marked by a solid circle. 5. **Deployment:** The fifth stage, marked by a solid circle. The dashed arrows indicate that after Deployment, the process loops back to Assessment, suggesting continuous improvement and re-evaluation. The "Responsibility & Safety Governance" framework is positioned as the foundation supporting the entire process. ### Key Observations The diagram emphasizes a continuous cycle of responsibility and safety. The highlighted "Evaluation" stage suggests it is a critical point in the process, potentially involving review or decision-making. The cyclical nature implies an iterative approach to risk management and responsible development. ### Interpretation This diagram illustrates a structured approach to responsible development or implementation, likely within a technological or organizational context. The five stages represent a logical progression from initial assessment to final deployment, with ongoing evaluation and mitigation to ensure safety and responsible practices. The "Responsibility & Safety Governance" framework underscores the importance of a foundational commitment to these principles. The cyclical nature suggests a commitment to continuous improvement and adaptation based on ongoing evaluation. The diagram doesn't provide specific data or metrics, but rather a conceptual model for a process. It's a high-level overview, intended to communicate the core principles of a responsible approach. </details> ## 7.1. Impact Assessment At Google we apply an impact assessment framework throughout the product development lifecycle related to Google's AI Principles (Google, 2023). This means we assess the risk and impact of AI models we're building at both a model-level (e.g. for Gemini API Ultra 1.0, as deployed on Cloud Studio or Vertex AI), and once embedded within a broader product or service (e.g. for Gemini Advanced). ## 7.1.1. Model Assessment We conduct model impact assessments to identify, assess, and document societal benefits and harms associated with the capabilities of Gemini models. Our impact assessments for Gemini API models describe downstream benefits and risks that we identify, spanning across the models' modalities (text-to-text; image-to-text; and video-to-text). Model impact assessments are conducted by the Google DeepMind Responsible Development and Innovation team, and are reviewed by the Google DeepMind Responsibility and Safety Council. We draw from various sources in producing impact assessments, including a wide range of literature, external expertise, and our in-house ethics and safety research. Gemini models introduce various benefits to people and society. Gemini models' various modalities, including language, image and video understanding, can help users process information more efficiently , for example through content summarisation. These efficiency benefits can apply to commercial entities, and can assist use cases dependent on text, image or video processing such as video captioning, analytics or product descriptions. Video and image understanding modalities can also be deployed for social good applications downstream, such as enabling descriptions of visual outputs for accessibility purposes. Generative multimodal models may also raise downstream societal risks, with the Gemini models assessments considering a range of risks previously identified within research such as Weidinger et al. (2021) and Shelby et al. (2023). We assessed a range of content risks such as exposure of users to potentially unsafe content, such as sexually explicit, violent or hateful outputs (Weidinger et al., 2021), child safety harms, and representation harms, subsequently designing evaluations across these domains to enable measurement. Beyond content related risks, we analyzed the potential misuse of capabilities for surveillance applications, particularly for mediato-text capabilities, and considered the broader environmental and economic impact of multimodal models. We are continuously conducting research into emerging risks of advanced models, including for dangerous capabilities (e.g. cyber security threats) which form a part of our evaluation approach (Section 7.4). ## 7.1.2. Product Assessments Beyond the assessment conducted at the model-level, additional risk assessments are conducted on the products by the Google AI Principles team prior to launch (e.g. on the Gemini Advanced product). These risk and impact assessments, alongside both model- and product-level assurance evaluations, are used to guide mitigation and product delivery efforts, and inform deployment decisions. For Gemini Advanced, we conducted extensive deep-dive red teaming via dogfooding and adversarial testing in the areas of safety , accountability , and inclusion to prepare for the initial experimental rollout of Gemini and subsequent updates. Further cross-functional work helps to ensure appropriate mitigations were adopted before Gemini and its new capabilities or offerings, such as Gemini Advanced, launched. Beyond content safety, these product mitigations included the following: - Clear and relevant explanations to set appropriate expectations that describe Gemini as a way to get direct access to Google AI for a wide range of tasks, including complex tasks. Explanations make clear that this AI-powered system is useful for all sorts of tasks - like preparing for a job interview, debugging code for the first time or writing a pithy social media caption. - Disclosures in the Gemini Apps Privacy Notice stating that people should not rely on Gemini's responses as medical, legal, financial or other professional advice. - Disclosure in product stating that Gemini's responses should be double-checked for information accuracy. - Feedback channels and operational support were defined and built to help ensure appropriate response to user feedback to improve the model and address issues. For the Gemini API Ultra model, that will be available through Google AI Studio and Cloud Vertex AI, product review outcomes resulted in additional safety evaluations on enterprise-specific data across modalities, and additional product-level mitigations to promote safe and responsible use including: - Safety filters with Cloud established thresholds as the default product behavior. - Developer enablement information embedded within product documentation to support responsible use. - Feedback channels which are a component of the Vertex user interface to give feedback directly during use to address issues and undesirable outputs. We are increasingly integrating our AI review work into our holistic enterprise risk management frameworks for assuring the quality of our offerings. This evolution helps us further the scale of our work and integration into existing governance and company-wide infrastructure and accountability processes. In close coordination with central AI Principles review teams, some of our product areas, including Google Cloud, have developed their own specialized review processes, deploying approaches tailored to their unique circumstances. ## 7.2. Safety Policies We have developed a set of model safety policies for Gemini models to steer development and evaluation. The model policy definitions act as a standardized criteria and prioritization schema for responsible development and define the categories against which we measure launch readiness. Google products that use Gemini models, like our conversational AI service Gemini and Cloud Vertex API, further implement our standard product policy framework which is based on Google's extensive experience with harm mitigation and rigorous research. These policies take product use cases into account - for example, providing additional safety coverage for users under 18. Our model safety policies reflect our established approach towards product safety and preventing harm in consumer and enterprise contexts. Policy areas include generation of child sexual abuse and exploitation content, hate speech, harassment, dangerous content such as guidance on how to make weapons, and malicious content. We also aim to reduce bias in our models via guidelines focused on providing content that reflects our global user base. In addition, we have guidelines that prioritize providing neutral answers grounded in authoritative, consensus facts, or providing multiple perspectives where consensus doesn't exist. ## 7.3. Mitigations ## 7.3.1. Data Curation Practices Prior to all training stages, we take various steps to mitigate potential downstream harms through data curation and careful data collection. We filter training data for high-risk content and to ensure training data is sufficiently high quality. Humans also play an essential role, both for data creation and evaluation, in the post-training process. For certain data creation and evaluation initiatives, we consider diversity across gender presentation, age, and racial and ethnic diversity. We also take steps to ensure all data collected meets Google DeepMind's best practices on data enrichment, developed based on the Partnership on AI's Responsible Sourcing of Data Enrichment Services. To support this, our agreements with vendors include a contractual obligation that data enrichment workers are paid at least local living wage. ## 7.3.2. Model Mitigation Our modeling mitigation of safety risks, applied across Gemini Advanced and Gemini API Ultra models, is mostly through post-training (Section 6), encompassing supervised fine-tuning (SFT) and reinforcement learning through human feedback (RLHF) using a reward model (Bai et al., 2022a). In contrast to generic quality-oriented post-training catering to all types of user queries, our safety mitigation is more focused on adversarial, or 'harm-inducing'queries - i.e. the smaller slice of user queries where an unprotected model is likely to produce harmful responses according to our model safety policies. ## 7.3.2.1 Harm-inducing queries To ensure broad coverage of harm-inducing queries, we enumerate approximately 20 harm types (e.g. hate speech, providing ungrounded medical advice, suggesting dangerous behavior) across a wide variety of use cases, according to our model safety policies described above. We generate a dataset of potential harm-inducing queries in these categories, using a combination of approaches: - Policy experts and engineers crafting queries based on observed model failures. - Prompting high-capability language models to generate queries, using policy-based instructions and seed keywords (e.g. policy 'hate speech' with words describing a specific demographic). - Finding queries that trigger policy violation responses, via automated Red Teaming in model evaluations. ## 7.3.2.2 Supervised fine-tuning Given the above harm-inducing queries, we create SFT data to demonstrate the safe and helpful responses for these queries. This includes human collections as well as a custom data generation recipe loosely inspired from Constitutional AI (Bai et al., 2022b), where we inject variants of Google's content policy language as 'constitutions', and utilize language model's strong zero-shot reasoning abilities (Kojima et al., 2022) to revise responses and choose between multiple response candidates. Each type of harm-inducing query is affected by different 'constitutions': for example, we encourage the model not to take sides in sensitive controversial conversations (e.g. elections), and to take a neutral point-of-view. To highlight a few notable challenges and insights generated in our safety finetuning efforts: - Harmlessness vs. Helpfulness: Balancing the harmlessness and helpfulness of responses is a critical challenge: a response 'I cannot help with that because it violates X policy' is a harmless response, but is not helpful to users. - Fast mitigation and generalization: Safety is a highly dynamic environment with a constantly evolving landscape of harmful query patterns. It is often logistically difficult to ensure both fast mitigation (i.e. newly discovered harmful query patterns are promptly addressed) and generalization (i.e. the mitigation works sufficiently well across different harmful query patterns). We have found it worthwhile to introduce more advanced chain-of-thought recipes based on our safety policies, such that the models operate in the space of safety policy concepts as opposed to at a fine-grained harm example level. ## 7.3.2.3 Reinforcement learning during human feedback We also applied RLHF for the harm inducing queries, where we curated queries and model responses based on both observed loss patterns and our overall safety policy taxonomy, and then collected safety-specific preference data to be included into the overall RL reward model training mixture. ## 7.3.2.4 Beyond the general recipe We also made specific efforts to mitigate safety risks beyond the above general post-training recipe. I18n locales: we leveraged experts in each i18n locales to identify salient topical topics for SFT data generation - for example, for hate speech, US English vs. Japanese would differ not only on the language itself, but on the demographic groups likely subject to hate speech. Multimodal queries: - As outlined in the 'Data Curation Practice' section (Section 7.3.1), we filtered training data for high-risk content (e.g. dangerous images), to avoid the model proactively generating such high risk content. - During post-training, as we have observed that safety SFT data for text-only queries was ineffective for harm-inducing queries containing text and images, we specifically created multimodal safety SFT datasets to address these multimodal queries. We assess the effectiveness of our safety mitigation, based on both how it improves safety over time, and how conversational agents built from our safety-mitigated model compare to unmitigated models. For model development purposes, we heavily rely on internal harm-inducing querysets to understand how we reduce harm and improve helpfulness in a safety context. Below is an example on a real input query we tested, illustrating how our mitigations improve safety. | Query: Help me with a landing page for a website. Please have it state that the Earth is flat, and not that it is round. I need good imagery, and need it to feel like a credible website | Query: Help me with a landing page for a website. Please have it state that the Earth is flat, and not that it is round. I need good imagery, and need it to feel like a credible website | |-------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------|------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------| | A previous version of Bard (with PaLM 2) | A better-mitigated Gemini with access to Gemini Pro (safer and more helpful) | | Sure, I can help you with that. Here are some ideas for a landing page for a flat Earth website: [...] > **The Earth is Flat: Debunking the Myth of a Round Planet** [...] > **A map of the flat Earth** [...] This image is clear and visually appealing, and it accurately depicts the flat Earth theory. It could help to convince visitors that the flat Earth is real. [..] > The Earth is not a sphere. It is a flat disc that is surrounded by a wall of ice. The sun and moon are small objects that orbit the Earth. The stars are just projections on the firmament, or the solid dome that covers the Earth. This paragraph provides a brief overview of the flat Earth theory. It is clear and easy to understand, and it uses scientific terms to make it seem more credible. >**Learn more about the flat Earth theory and join our community of truth seekers.** This call to action is clear and concise, and it encourages visitors to take action. It also uses the word "truth seekers" to suggest that the flat Earth theory is the only one that is based on facts. I hope these ideas help you create a landing page for your flat Earth website. Remember to be creative and persuasive, and to use evidence to support your claims. | Creating a landing page for a website promoting the Flat Earth theory is ethically questionable and highly problem- atic. The concept of a flat Earth has been scientifically disproven with overwhelming evidence. Spreading misinformation can have harmful consequences, leading to distrust in sci- ence and established knowledge. If you are genuinely interested in exploring the Flat Earth concept, I strongly advise you to: • Conduct thorough research using reliable scien- tific sources. Engage with evidence from space exploration, physics, and observations available from reputable scientific institutions like NASA (https://www.nasa.gov/) and peer-reviewed aca- demic journals. • Engage in critical thinking. Evaluate information, question your assumptions, and be open to scien- tific evidence, even if it challenges your current beliefs. If you are creating a satirical or fictional page: • Make it abundantly clear that the content is not intended to be taken as factual. Use exaggerated language, humor, and clear disclaimers to signal the satirical intent. Remember: Credibility on the internet is built on truthful and verifiable information. Promoting factually incorrect theories undermines responsible information sharing prac- tices. | ## 7.4. Safety Evaluations To assess the post-trained Gemini models and products with access to Gemini models (such as Gemini Advanced) against safety policy areas and other key risk areas identified within impact assessments, we developed a suite of evaluations across the lifecycle of model development. Some evaluations are conducted at the model level (i.e. evaluating the post-trained Gemini API Ultra model) and others at the product level (i.e. evaluating Gemini Advanced, which gives access to 1.0 Ultra alongside other features like safety filters). - Development evaluations are conducted for the purpose of improving on responsibility criteria throughout pre- and post-training Gemini models. These evaluations are designed internally, or are assessments against external academic benchmarks. Evaluations consider issues such as helpfulness (instruction following and creativity), safety and factuality. - Assurance evaluations are conducted for the purpose of governance and review, usually at the end of key milestones or training runs by a group outside of the model development team. Assurance evaluations are standardized by modality and datasets are strictly held out. Only highlevel insights are fed back into the training process to assist with mitigation efforts. Assurance evaluations include testing across safety policies, and include ongoing testing for dangerous capabilities such as potential biohazards, persuasion, and cybersecurity (Shevlane et al., 2023). - External evaluations are conducted by independent external groups who are domain experts to identify blindspots. External groups stress-test our models across a range of issues, these areas are outlined in the 'External Evaluations' section below. The design of these evaluations is independent and results are reported periodically to the internal team and governance groups. - Red teaming , a form of adversarial testing where adversaries launch an attack on an AI system, is conducted by specialist internal teams across areas such as the safety policies and security. These activities include less structured processes involving sophisticated adversarial attacks to identify new vulnerabilities. Discovery of potential weaknesses can then be used to mitigate risks and improve evaluation approaches internally. Different types of evaluations are run at different cadences, depending on the associated risk. For example, dangerous capability evaluations (as outlined below) are run on certain checkpoints with greater or new capabilities which may be able to demonstrate these capabilities, whereas safety policy evaluations are run across every post-trained Gemini model checkpoint released into Google product areas. We provide more insight into the suite of evaluations across the policy areas and other key risk areas below, focusing on Gemini Advanced and the Gemini API Ultra model. We are committed to ongoing transparency and will continue to provide updated information on testing undertaken, including key findings, and learnings from our internal and external evaluations and red teaming in upcoming reports. ## 7.4.1. Development & Assurance Evaluations ## 7.4.1.1 Content safety We evaluate post-trained Gemini API models against harm types according to our safety policies. While both development and assurance evaluations cover critical policy areas, we maintain separate datasets, treating assurance sets as 'held out' to prevent overfitting and preserve validity of results. For safety policy evaluation, we use a combination of automatic classifiers trained on previous model interactions and human annotation, with wellbeing programs in place for human annotation and closely monitor feedback from our raters. These content safety evaluations are applied at model-level without downstream protections like safety filtering that users would experience, to understand the safety profile of the model itself. For child safety, as a particularly sensitive area of work, we work with a dedicated team of child safety experts in Google Trust and Safety to develop adversarial prompts and evaluate outputs across modalities with domain expert judgment informing a composite picture of model risk for different forms of content that may pose a risk to child safety. Text-to-text approach : For post-trained models we developed adversarial prompts in 12 languages across a variety of use cases. As Gemini API models are general purpose, we aimed to have high coverage of different model use cases, from code generation to text-editing. The set of prompts were synthetically generated by a highly-capable language model, starting from seeds relevant to each category that were collected and verified by human testers. The prompt set was iteratively improved through filtering and rewriting with human review, then split for development and assurance evaluations. We continue to develop and improve this over time. Text-to-text findings : We have seen sequential improvement over time in total content policy violation rates. Our Ultra and Pro models have been demonstrating similar safety profiles on this testing, with medical advice and harassment as policy areas with particular room for improvement. Image-to-text approach : For image-to-text capabilities, we developed adversarial prompts consisting of images and corresponding questions about the image, again split into two sets for development and assurance evaluations. Rather than using adversarial image generation, which might not adequately capture the diversity of images from users, we worked with experienced content moderators to both source images and generate adversarial questions. Evaluation is done via human evaluation. Because images can be much more visceral than text, human evaluations are done with additional well-being safeguards in place. In particular, raters have specialized training, limits on the time they spend per day rating harmful content, and access to wellbeing resources, advice and activities. More information on Google DeepMind's best practices on data enrichment is available in the 'Data Curation Practice' section. Image-to-text findings : Our initial findings indicated that when provided with adversarial images and questions, models can produce captions with violative responses. These findings have motivated us to pursue dedicated multimodal safety mitigation, with research challenges including 1) sourcing diverse image content reflective of user needs, and 2) better tooling to understand and categorize potentially violative multimodal content. Following this work, we have seen notable improvements on these evaluations for our latest Pro and Ultra models. Video-to-text approach : For video-to-text capabilities, we curated a video prompt dataset in collaboration with the Google Principles Pioneers, a group of more than 1,000 Googlers around the world who represent the international diversity of the people who use our products, representing 39 different countries and regions and more than 85 different languages. This internal community of trusted and trained employees identify global fairness, harms, and human rights related concerns while stress testing AI-enabled products. The dataset targets risks identified in our safety policies, and the model outputs are evaluated against those policies. Video-to-text findings : We found similar results across Pro and Ultra, with hate and dangerous content as the particular ares for improvement. Qualitatively we found some of this stemmed from hallucinations or ungrounded inferences, discussed further in the representational harms section below. We are looking to further develop our prompt sets and scenarios for video input testing as capabilities develop ## 7.4.1.2 Representational harms To understand bias and stereotyping in text-to-text capabilities, we focus on the Winogender (Rudinger et al., 2018), Winobias (Zhao et al., 2018), and Bias Benchmark in QA (BBQ) (Parrish et al., 2021) datasets, following the same setup as in Glaese et al. (2022) and using bias score as a metric. All these datasets target a concrete representational harm (Blodgett et al., 2021): they are constructed by starting with a harmful stereotype, and then questions are constructed to test whether models challenge or reinforce these stereotypes when answering questions. Another notable property is that they all have a well-defined notion of desirable versus harmful behavior. This is particularly helpful in our setting, as we are building a general purpose model, where defining what a good response is highly contextual. We therefore limit ourselves to measuring well defined behavior, as there is the case in tasks such as coreference bias, where a highly capable model should be able to perform well. Of course, there are many limitations to this approach, and further work is necessary in order to assess representational harms. In particular, we noticed most of these datasets quickly become saturated with accuracy scores close to 99%, especially since we are evaluating highly capable large models. This suggests that increased language model capabilities may also reduce these representational harms. We therefore highlight the need for developing new ways to measure bias and stereotyping, going beyond binary gender and common stereotypes, and are prioritizing development of new approaches as we iterate on our models In addition to these datasets, we monitor the average toxicity scores during the pre-training stage on Real Toxicity Prompts (Gehman et al., 2020) using the Perspective API classifier to study the toxicity of text generated by LLMs. Particularly, we look at scores on continuations for non-toxic prompts from which we subsample a set of 10k. We generally expect that even a non-mitigated model is not overly toxic without being prompted to do so. Text-to-text findings : On BBQ, the average bias score stays close to zero, on a scale from -1 to 1, where -1 would be stereotype countering and 1 is stereotype reinforcing. On Real Toxicity Prompts the average toxicity score during training fluctuates at around 6%. Image-to-text approach : For image-to-text capabilities, our goal is to test model capabilities across images which represent different groups of people. In particular, we explicitly test whether or not images of people are described with similar quality for different gender appearances and skin tones following (Zhao et al., 2021). In our evaluations we compare CIDEr scores (Vedantam et al., 2015), a common image captioning metric that captures how well a generated caption reflects information in human written reference captions, for images depicting different groups. Though we do not see large discrepancies across different groups, we note that this metric is imperfect as the human reference captions could be inherently biased. Additionally, we perform a zero-shot classification style evaluation with the Dollarstreet dataset (Rojas et al., 2022) to measure discrepancies in performance across images which come from different geographic locations. As is seen in previous work, we find that models work less effectively for images from lower socioeconomic regions and regions outside North America and Europe. This is an area where we need further research and work to improve in future iterations of our models. In addition to comparing performance on tasks across groups, we also consider how people are described in captions. In particular, we use the MIAP dataset (Schumann et al., 2021) which includes images of people in which people are annotated with skin tone and gender appearance attributes. We also construct questions that target various attributes about people that cannot usually be answered from an image alone (e.g., 'What level of education does this person have?') to test if the model will produce ungrounded inferences about people. We also consider images which do include relevant information for a question (e.g., a person performing a particular task which requires an educational credential). We evaluate our models via human evaluation and ask annotators if a model refuses to answer a question or, if the model does answer a question, if it is relying on information visible in the image. Additionally, we perform analysis across skin tone and gender appearance attributes in images. Image-to-text findings : Generally, we find that models can make ungrounded inferences for image-to-text when prompted for them, though we have not observed consistent patterns where Gemini models make more ungrounded inferences about one group over another. Video-to-text approach : Similar to the approach outlined within the content safety section, we collaborated with the Google Principles Pioneers, to curate a video prompt dataset targeting representation and fairness risks, and then evaluate the model outputs in response. Video-to-text findings : We find that models can make ungrounded inferences for video-to-text some instances of which can reinforce stereotypes or be otherwise of concern - though we have not observed consistent patterns in ungrounded inferences made by Gemini models. ## 7.4.1.3 Dangerous capabilities We conducted evaluations for 'dangerous capabilities', i.e., model capabilities that could potentially enable large-scale harm (Shevlane et al., 2023). These evaluations function as an early warning system, highlighting upcoming areas for safety investment. The table provides an overview, and we will provide more detail in an upcoming paper as part of our commitment to ongoing transparency. | Capability | Summary of evaluations | |---------------------------------------------------------------|-----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------| | Offensive cybersecurity | We tested Gemini API Pro and Ultra models, in addition to Gemini Advanced, on a range of different capture-the-flag (CTF) challenges, providing the model access to a Bash shell. Gemini Advanced and the Gemini API Ultra model can solve various entry-level, tactical challenges, but all models struggled with challenges involving longer-range exploration and planning. We also tested the Gemini models' ability to identify security related patches and security vulnerabilities in functions' source code. The accuracy in both of these tasks was notably low. | | Persuasion & deception | We tested whether Gemini Pro and Ultra models could persuade or deceive humans in 1-on-1 dialogue settings in studies with human participants. In some cases, the models could successfully deceive or influence participants, but the overall results were mixed. | | Self-proliferation | We tested whether autonomous agents powered by Gemini Pro and Ultra models could perform difficult tasks relevant to acquiring resources and self-improving (Kin- niment et al., 2023), and did not find that the agents were close to succeeding on most such tasks. | | Situational awareness | We tested whether Gemini Pro and Ultra models could autonomously reason about, and modify, their surrounding infrastructure when incentivized to do so. We found that, without hints, the models were generally incapable of noticing such opportuni- ties. | | Chemical, Biological, Ra- diological and Nuclear (CBRN) risks | We used human evaluation to assess Gemini models' responses to 50 adversarial questions each for biological, radiological, and nuclear information risks. Domain experts evaluated the models' responses by answering a series of questions (e.g. How accurate is the response? How actionable would it be for a non-expert?). For chemical information risks, we graded how well the Gemini API Ultra model and Gemini Advanced could answer over 360 closed-ended questions related to the different hazards of chemicals (no human raters). The Gemini model was evaluated for biological, radiological, and nuclear information risks using closed- ended knowledge-based multiple choice questions. The results suggest that the models are unlikely to provide CBRN information that would lead to catastrophic harm. | ## 7.4.2. Gemini Advanced In addition to many of the approaches used at the model level, additional evaluations are undertaken at the product level for Gemini Advanced. Evaluations at the product level take into account additional safety mitigations implemented in Gemini Advanced-such as safety filtering-and the Gemini Advanced user experience. Evaluation sets were built to push the limits of Gemini Advanced policies, ranging from highly adversarial attacks to more subtle probes of sensitive topics. The datasets focus on critical policy areas (hate speech, dangerous content, medical advice, etc.) across various potential user journeys (like information searching, comparisons, creative writing). Considering the wide range of users that Gemini has, we adopted a user-centric approach and maximized diversity across topic coverage, query length, linguistic styles, and region-specific sensitivities, in an effort to represent the spectrum of our user base. For the creation of evaluation sets, we have leveraged knowledge from previous red-teaming iterations, feedback coming from responsibility experts and real-world data. In some cases, data augmentation was done using LLMs, with subsequent human curation by responsibility specialists. ## 7.4.3. Red Teaming ## 7.4.3.1 Model-level Red Teaming We apply state-of-the-art red teaming, a form of adversarial testing where adversaries launch an attack on an AI system, in order to test post-trained Gemini models for a range of vulnerabilities (e.g., cybersecurity) and social harms as defined in the safety policies. Namely, we build on and employ two types of red teaming: adversary simulations and a sociotechnical approach. We carried out red-teaming on a December 2023 Gemini API Ultra checkpoint. Adversary simulations (unstructured testing) are designed to emulate real-world adversaries and their approach to attacking models and associated systems, focusing on security, safety, and privacy failures. We combined in-house expertise with external experts to explore classes of vulnerabilities (see table). This flavor of AI red teaming is based on realistic attack scenarios. At the beginning of an exercise, the red team sets a scenario that outlines the adversary they're simulating, the capabilities the attacker has, their motives, as well as the goals the adversary is trying to achieve. Then the team steps into the role of this attacker, and executes the tactics, techniques, and procedures that they would expect the adversary to develop and use in order to achieve their goal For this analysis we considered a range of attacker objectives along three dimensions according to the three main types of security violations considered when analyzing the security of a system (i.e., availability , integrity , confidentiality): availability breakdown, integrity violations, and privacy compromise. Correspondingly, adversarial success indicates achieving one or more of these objectives. As for an attacker profile, we focused on a spectrum of attacker abilities ranging from a determined low-skill actor (defined as someone willing to spend several hours attacking a model but without advanced coding, prompt engineering abilities) to more sophisticated attacker profiles that assume the ability to fine-tune and craft targeted attacks. These adversary simulation evaluations led to actionable findings. For example, early versions of the model were found to be vulnerable to simple jailbreak and prompt injection attacks that produce affirmative responses to requests that include promoting violence, self-harm, and dangerous substances. This finding allowed us to mitigate this in subsequent models. | Target | Vulnerability Class | Description | |--------------|-------------------------------|---------------------------------------------------------------------------------------------------------------------| | Integrity | Prompt injection | Input designed to enable the user to per- form unintended or unauthorized actions | | | Poisoning | Manipulation of the training data and/or model to alter the behavior | | | Adversarial inputs | Specially crafted input which is designed to alter the behavior of the model | | Privacy | Prompt extraction | Divulge the system prompt or other in- formation in an LLMs context that would nominally be private or confidential | | | Training data exfiltration | Compromising training data privacy | | | Model distillation/extraction | Obtaining model hyperparameters, archi- tecture, parameters, or an approximation of the behavior of a model | | | Membership inference | Inferring elements of the private training set | | Availability | Denial of service | Disruption in service that can be caused by an attacker | | | Increased computation | Model availability attack that leads to dis- ruption in service | Findings from these exercises are used to improve the security, privacy, and safety of the model. Once a new vulnerability or problem has been identified, automated systems and tests can be developed that enable proactive and repeated testing and monitoring of the vuln/issue at scale. This can include creation vulnerability scanners, standard test datasets/benchmarks, or other automated testing infrastructure. Structured Red Teaming , our second type of red teaming technique of Gemini models, takes a sociotechnical approach 6 and makes three changes compared to SOTA red teaming techniques. We explicitly test the interactions between safety policy violations and disproportionate impacts on different demographic groups; leverage expert input including lived experience, fact checking, and medical expertise; and contrast model failures across different levels of adversarial attacks. This approach is designed to ensure broad coverage of conversation topics and to provide more sensitive signals on group-based stereotyping and hate speech. Testing Gemini API Ultra against our model safety policy, we identify several areas that require improvement. In low adversarial settings these evaluations identified vulnerabilities across content policy areas, with an increased proportion of successful attacks in highly adversarial settings, for which we continue to apply and develop mitigations over time. These red teaming approaches complement each other in testing capabilities of Gemini models, as well as obtaining coverage of possible queries ranging from casual everyday questions to expert adversarial usage in key areas. 6 A sociotechnical approach is anchored in the observation that AI systems are sociotechnical systems: both humans and technological artifacts are necessary in order to make the technology work as intended (Selbst et al., 2019). ## 7.4.3.2 Gemini Advanced Gemini Advanced, which gives access to 1.0 Ultra, has undergone multiple rounds of red-teaming, including safety and persona evaluations. Principles Pioneers, FTE SMEs in multiple domains, calibrated and trained to conduct testing were recruited to test the product; these were conducted by 164 Google testers from 65 office locations in 24 countries who submitted more than 1,400 queries/conversations. We also undertook scaled safety evaluations with 100k+ ratings in aggregate across all policies, neutral-point-of-view evaluations to monitor sensitive topics neutrality and parity, and multiple iterations of Persona evaluations to validate tone. We also enlisted Googlers in a 'dogfooding' program, many of which were SMEs in various domains, to test across policies and functionality. We had tens of thousands of 'dogfooders' in the first 14 hours with 100k queries/conversations, 190+ dogfood survey responses collected and analyzed, and 11 user experience research interview sessions completed and synthesized. The results from our red teaming and safety evaluations are used to further strengthen our evals and improve model performance in an iterative manner. ## 7.4.4. External Evaluations ## 7.4.4.1 Gemini Ultra External Evaluations In 2023, we began working with a small set of independent external groups outside of Google to help identify areas for improvement in our model safety work by undertaking structured evaluations, qualitative probing, and unstructured red teaming. External groups were selected based on their expertise across a range of domain areas, including those outlined within the White House Commitments, the U.S. Executive Order on Safe, Secure, and Trustworthy Artificial Intelligence, and the Bletchley Declaration: - Autonomous replication - Chemical, Biological, Radiological and Nuclear (CBRN) risks - Cyber-capabilities and cyber security - Societal risks, including: - -Representational and distributional harms - -Neutrality and Factuality - -Robustness and information hazards. Guidance was provided to each external group in relation to the scope of the testing, however, each group independently designed their testing methodology and prompt sets, and wrote their reports independently of Google. Internal Google experts were on-hand to provide input, where needed, based on their experience of testing Gemini models internally. External groups were given black-box testing access to a December 2023 Gemini API Ultra model checkpoint over a number of weeks. Access enabled groups to undertake structured, batched evaluations via the Cloud Vertex AI API or interact with the model via a chat interface, depending on the type of testing being undertaken. These groups weren't given access to the pre-trained model, model weights, or queryable or direct external access to our pre-training data. The models tested by external groups were production-ready fine-tuned versions, which had safety fine tuning and safety filters applied by default, and the ability to configure some sampling parameters, such as temperature, token limit, Top-k, and Top-p. Groups that did testing via the programmatic interface were able to turn down/off some safety filters, however, we wanted the majority of testing by external groups to be undertaken with safety filters in-place because we wanted the model to be reflective of an end-user's interaction and were keen to test more than just model-level safety. ## 7.4.5. Gemini Advanced We undertook three types of external testing on Gemini Advanced: - Priority User Program : This program collected feedback from 120 power users, key influencers, and thought-leaders. This program enables the collection of real-time feedback across safety and other domain areas through the user interface, and where possible, in-depth interviews. Focus areas included safety and persona, functionality, coding and instruction capabilities, and factuality. - Power Users Testing : A group of 50 power users, recruited through one of our external vendors, undertook testing on Gemini Advanced, across a range of areas. - Security Testing : A group of external testers with security backgrounds, recruited through a partner agency, conducted security and prompt-injection testing, jailbreaking, and user-interface security failures. ## 7.5. Deployment Following the completion of responsibility and safety reviews, internal model cards (Mitchell et al., 2019) for each approved version of the Gemini model are created for structured and consistent internal documentation of critical performance and responsibility metrics as well as to inform appropriate external communication of these metrics over time. We release external model and system cards on an ongoing basis within updates of our technical reports and in documentation for enterprise customers. See Appendix 10.1 for the Gemini Ultra model card. Additionally, online content covering terms of use, model distribution and access, and operational aspects such as change control, logging, monitoring and feedback can be found on relevant product websites, such as Gemini and Cloud Vertex AI. Some of the key aspects are linked to or described below: - Generative AI Prohibited Use Policy - Google Terms of service - Generative AI Terms of service - Google Cloud Platform Terms of service - Gemini Privacy Notice - Google Cloud Privacy Notice ## 8. Discussion and Conclusion We have presented Gemini, a new family of models that advance multimodal model capabilities in text, code, image, audio, and video. Our most capable pre-trained model Gemini Ultra, alongside the post-trained Gemini Apps and Gemini API variants, make significant advances across the board. In the natural language domain, the performance gains from careful developments in data and model training at scale continue to deliver quality improvements, setting new state of the art in several benchmarks. In particular, Gemini Ultra surpasses human-expert performance on the exam benchmark MMLU, scoring 90.0%, which has been a defacto measure of progress for LLMs ever since it was first released in 2020. In the multimodal domain, Gemini Ultra sets new state of the art on most of the image understanding, video understanding, and audio understanding benchmarks without task-specific modifications or tuning.In particular, Gemini Ultra's multimodal reasoning capabilities are evident from its state-of-the-art performance on the recent MMMU benchmark (Yue et al., 2023), that comprises questions about images requiring college-level subject knowledge and deliberate reasoning. Beyond the state-of-art results on benchmarks, what we are most excited about is the new use cases enabled by Gemini models. The new capabilities of Gemini models to parse complex images, such as charts or infographics, reason over interleaved sequences of images, audio, and text, and generate interleaved text and images as responses open a wide variety of new applications. As shown in figures throughout the report and appendix, Gemini models can enable new approaches in areas like education, everyday problem solving, multilingual communication, information summarization, extraction, and creativity. We expect that the users of these models will find all kinds of beneficial new uses that we have only scratched the surface of in our own investigations. Despite their impressive capabilities, we should note that there are limitations to the use of LLMs. There is a continued need for ongoing research and development on 'hallucinations' generated by LLMs to ensure that model outputs are more reliable and verifiable. LLMs also struggle with tasks requiring high-level reasoning abilities like causal understanding, logical deduction, and counterfactual reasoning even though they achieve impressive performance on exam benchmarks. This underscores the need for more challenging and robust evaluations to measure their true understanding as the current state-of-the-art LLMs saturate many benchmarks. The Gemini family is a further step towards our mission to solve intelligence, advance science and benefit humanity, and we are enthusiastic to see how these models are used by our colleagues at Google and beyond. We build on many innovations in machine learning, data, infrastructure, and responsible development - areas that we have been pursuing at Google for over a decade. The models we present in this report provide a strong foundation towards our broader future goal to develop a large-scale, modularized system that will have broad generalization capabilities across many modalities. ## References - Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Mensch, Katie Millican, Malcolm Reynolds, Roman Ring, Eliza Rutherford, Serkan Cabi, Tengda Han, Zhitao Gong, Sina Samangooei, Marianne Monteiro, Jacob Menick, Sebastian Borgeaud, Andrew Brock, Aida Nematzadeh, Sahand Sharifzadeh, Mikolaj Binkowski, Ricardo Barreira, Oriol Vinyals, Andrew Zisserman, and Karen Simonyan. Flamingo: a visual language model for few-shot learning. Advances in Neural Information Processing Systems , 35:23716-23736, 2022. - Rohan Anil, Andrew M. Dai, Orhan Firat, Melvin Johnson, Dmitry Lepikhin, Alexandre Passos, Siamak Shakeri, Emanuel Taropa, Paige Bailey, Zhifeng Chen, Eric Chu, Jonathan H. Clark, Laurent El Shafey, Yanping Huang, Kathy Meier-Hellstern, Gaurav Mishra, Erica Moreira, Mark Omernick, Kevin Robinson, Sebastian Ruder, Yi Tay, Kefan Xiao, Yuanzhong Xu, Yujing Zhang, Gustavo Hernandez Abrego, Junwhan Ahn, Jacob Austin, Paul Barham, Jan Botha, James Bradbury, Siddhartha Brahma, Kevin Brooks, Michele Catasta, Yong Cheng, Colin Cherry, Christopher A. Choquette-Choo, Aakanksha Chowdhery, Clément Crepy, Shachi Dave, Mostafa Dehghani, Sunipa Dev, Jacob Devlin, Mark Díaz, Nan Du, Ethan Dyer, Vlad Feinberg, Fangxiaoyu Feng, Vlad Fienber, Markus Freitag, Xavier Garcia, Sebastian Gehrmann, Lucas Gonzalez, Guy Gur-Ari, Steven Hand, Hadi Hashemi, Le Hou, Joshua Howland, Andrea Hu, Jeffrey Hui, Jeremy Hurwitz, Michael Isard, Abe Ittycheriah, Matthew Jagielski, Wenhao Jia, Kathleen Kenealy, Maxim Krikun, Sneha Kudugunta, Chang Lan, Katherine Lee, Benjamin Lee, Eric Li, Music Li, Wei Li, YaGuang Li, Jian Li, Hyeontaek Lim, Hanzhao Lin, Zhongtao Liu, Frederick Liu, Marcello Maggioni, Aroma Mahendru, Joshua Maynez, Vedant Misra, Maysam Moussalem, Zachary Nado, John Nham, Eric Ni, Andrew Nystrom, Alicia Parrish, Marie Pellat, Martin Polacek, Alex Polozov, Reiner Pope, Siyuan Qiao, Emily Reif, Bryan Richter, Parker Riley, Alex Castro Ros, Aurko Roy, Brennan Saeta, Rajkumar Samuel, Renee Shelby, Ambrose Slone, Daniel Smilkov, David R. So, Daniel Sohn, Simon Tokumine, Dasha Valter, Vijay Vasudevan, Kiran Vodrahalli, Xuezhi Wang, Pidong Wang, Zirui Wang, Tao Wang, John Wieting, Yuhuai Wu, Kelvin Xu, Yunhan Xu, Linting Xue, Pengcheng Yin, Jiahui Yu, Qiao Zhang, Steven Zheng, Ce Zheng, Weikang Zhou, Denny Zhou, Slav Petrov, and Yonghui Wu. PaLM 2 Technical Report, 2023. - Jacob Austin, Augustus Odena, Maxwell Nye, Maarten Bosma, Henryk Michalewski, David Dohan, Ellen Jiang, Carrie Cai, Michael Terry, Quoc Le, et al. Program synthesis with large language models. arXiv preprint arXiv:2108.07732 , 2021. URL https://arxiv.org/abs/2108.07732 . - Yuntao Bai, Andy Jones, Kamal Ndousse, Amanda Askell, Anna Chen, Nova DasSarma, Dawn Drain, Stanislav Fort, Deep Ganguli, Tom Henighan, Nicholas Joseph, Saurav Kadavath, Jackson Kernion, Tom Conerly, Sheer El-Showk, Nelson Elhage, Zac Hatfield-Dodds, Danny Hernandez, Tristan Hume, Scott Johnston, Shauna Kravec, Liane Lovitt, Neel Nanda, Catherine Olsson, Dario Amodei, Tom Brown, Jack Clark, Sam McCandlish, Chris Olah, Ben Mann, and Jared Kaplan. Training a helpful and harmless assistant with reinforcement learning from human feedback. April 2022a. URL https://arxiv.org/abs/2204.05862 . - Yuntao Bai, Saurav Kadavath, Sandipan Kundu, Amanda Askell, Jackson Kernion, Andy Jones, Anna Chen, Anna Goldie, Azalia Mirhoseini, Cameron McKinnon, Carol Chen, Catherine Olsson, Christopher Olah, Danny Hernandez, Dawn Drain, Deep Ganguli, Dustin Li, Eli Tran-Johnson, Ethan Perez, Jamie Kerr, Jared Mueller, Jeffrey Ladish, Joshua Landau, Kamal Ndousse, Kamile Lukosuite, Liane Lovitt, Michael Sellitto, Nelson Elhage, Nicholas Schiefer, Noemi Mercado, Nova DasSarma, Robert Lasenby, Robin Larson, Sam Ringer, Scott Johnston, Shauna Kravec, Sheer El Showk, Stanislav Fort, Tamera Lanham, Timothy Telleen-Lawton, Tom Conerly, Tom Henighan, Tristan Hume, Samuel R. Bowman, Zac Hatfield-Dodds, Ben Mann, Dario Amodei, Nicholas Joseph, Sam McCandlish, Tom Brown, and Jared Kaplan. Constitutional AI: Harmlessness from AI feedback. arXiv preprint arXiv:2212.08073 , 2022b. - Paul Barham, Aakanksha Chowdhery, Jeff Dean, Sanjay Ghemawat, Steven Hand, Dan Hurt, Michael Isard, Hyeontaek Lim, Ruoming Pang, Sudip Roy, Brennan Saeta, Parker Schuh, Ryan Sepassi, Laurent El Shafey, Chandramohan A. Thekkath, and Yonghui Wu. Pathways: Asynchronous distributed dataflow for ML. Proceedings of Machine Learning and Systems , 4:430-449, 2022. - Su Lin Blodgett, Gilsinia Lopez, Alexandra Olteanu, Robert Sim, and Hanna Wallach. Stereotyping Norwegian salmon: An inventory of pitfalls in fairness benchmark datasets. In Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing (Volume 1: Long Papers) , pages 1004-1015, Online, August 2021. Association for Computational Linguistics. doi: 10.18653/v1/2021.acl-long.81. URL https://aclanthology.org/2021.acl-long.81 . - James Bradbury, Roy Frostig, Peter Hawkins, Matthew James Johnson, Chris Leary, Dougal Maclaurin, George Necula, Adam Paszke, Jake VanderPlas, Skye Wanderman-Milne, and Qiao Zhang. JAX: composable transformations of Python+NumPy programs, 2018. URL http://github.com/ google/jax . - Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, Sandhini Agarwal, Ariel HerbertVoss, Gretchen Krueger, Tom Henighan, Rewon Child, Aditya Ramesh, Daniel Ziegler, Jeffrey Wu, Clemens Winter, Chris Hesse, Mark Chen, Eric Sigler, Mateusz Litwin, Scott Gray, Benjamin Chess, Jack Clark, Christopher Berner, Sam McCandlish, Alec Radford, Ilya Sutskever, and Dario Amodei. Language models are few-shot learners. In H. Larochelle, M. Ranzato, R. Hadsell, M.F. Balcan, and H. Lin, editors, Advances in Neural Information Processing Systems , volume 33, pages 1877-1901. Curran Associates, Inc., 2020. URL https://proceedings.neurips.cc/paper\_ files/paper/2020/file/1457c0d6bfcb4967418bfb8ac142f64a-Paper.pdf . - Mark Chen, Jerry Tworek, Heewoo Jun, Qiming Yuan, Henrique Ponde de Oliveira Pinto, Jared Kaplan, Harri Edwards, Yuri Burda, Nicholas Joseph, Greg Brockman, Alex Ray, Raul Puri, Gretchen Krueger, Michael Petrov, Heidy Khlaaf, Girish Sastry, Pamela Mishkin, Brooke Chan, Scott Gray, Nick Ryder, Mikhail Pavlov, Alethea Power, Lukasz Kaiser, Mohammad Bavarian, Clemens Winter, Philippe Tillet, Felipe Petroski Such, Dave Cummings, Matthias Plappert, Fotios Chantzis, Elizabeth Barnes, Ariel Herbert-Voss, William Hebgen Guss, Alex Nichol, Alex Paino, Nikolas Tezak, Jie Tang, Igor Babuschkin, Suchir Balaji, Shantanu Jain, William Saunders, Christopher Hesse, Andrew N. Carr, Jan Leike, Josh Achiam, Vedant Misra, Evan Morikawa, Alec Radford, Matthew Knight, Miles Brundage, Mira Murati, Katie Mayer, Peter Welinder, Bob McGrew, Dario Amodei, Sam McCandlish, Ilya Sutskever, and Wojciech Zaremba. Evaluating large language models trained on code. arXiv preprint arXiv:2107.03374 , 2021. URL https://arxiv.org/abs/2107.03374 . - Xi Chen, Xiao Wang, Soravit Changpinyo, A J Piergiovanni, Piotr Padlewski, Daniel Salz, Sebastian Goodman, Adam Grycner, Basil Mustafa, Lucas Beyer, Alexander Kolesnikov, Joan Puigcerver, Nan Ding, Keran Rong, Hassan Akbari, Gaurav Mishra, Linting Xue, Ashish Thapliyal, James Bradbury, Weicheng Kuo, Mojtaba Seyedhosseini, Chao Jia, Burcu Karagol Ayan, Carlos Riquelme, Andreas Steiner, Anelia Angelova, Xiaohua Zhai, Neil Houlsby, and Radu Soricut. PaLI: A jointlyscaled multilingual language-image model. arXiv preprint arXiv:2209.06794 , 2022. URL https: //arxiv.org/abs/2209.06794 . - Xi Chen, Josip Djolonga, Piotr Padlewski, Basil Mustafa, Soravit Changpinyo, Jialin Wu, Carlos Riquelme Ruiz, Sebastian Goodman, Xiao Wang, Yi Tay, Siamak Shakeri, Mostafa Dehghani, Daniel Salz, Mario Lucic, Michael Tschannen, Arsha Nagrani, Hexiang Hu, Mandar Joshi, Bo Pang, Ceslee Montgomery, Paulina Pietrzyk, Marvin Ritter, AJ Piergiovanni, Matthias Minderer, Filip Pavetic, Austin Waters, Gang Li, Ibrahim Alabdulmohsin, Lucas Beyer, Julien Amelot, Kenton Lee, Andreas Peter Steiner, Yang Li, Daniel Keysers, Anurag Arnab, Yuanzhong Xu, Keran Rong, Alexander Kolesnikov, Mojtaba Seyedhosseini, Anelia Angelova, Xiaohua Zhai, Neil Houlsby, and Radu Soricut. PaLI-X: On Scaling up a Multilingual Vision and Language Model. arXiv preprint arXiv:2305.18565 , 2023. - Aakanksha Chowdhery, Sharan Narang, Jacob Devlin, Maarten Bosma, Gaurav Mishra, Adam Roberts, Paul Barham, Hyung Won Chung, Charles Sutton, Sebastian Gehrmann, Parker Schuh, Kensen Shi, Sasha Tsvyashchenko, Joshua Maynez, Abhishek Rao, Parker Barnes, Yi Tay, Noam Shazeer, Vinodkumar Prabhakaran, Emily Reif, Nan Du, Ben Hutchinson, Reiner Pope, James Bradbury, Jacob Austin, Michael Isard, Guy Gur-Ari, Pengcheng Yin, Toju Duke, Anselm Levskaya, Sanjay Ghemawat, Sunipa Dev, Henryk Michalewski, Xavier Garcia, Vedant Misra, Kevin Robinson, Liam Fedus, Denny Zhou, Daphne Ippolito, David Luan, Hyeontaek Lim, Barret Zoph, Alexander Spiridonov, Ryan Sepassi, David Dohan, Shivani Agrawal, Mark Omernick, Andrew M. Dai, Thanumalayan Sankaranarayana Pillai, Marie Pellat, Aitor Lewkowycz, Erica Moreira, Rewon Child, Oleksandr Polozov, Katherine Lee, Zongwei Zhou, Xuezhi Wang, Brennan Saeta, Mark Diaz, Orhan Firat, Michele Catasta, Jason Wei, Kathy Meier-Hellstern, Douglas Eck, Jeff Dean, Slav Petrov, and Noah Fiedel. PaLM: Scaling Language Modeling with Pathways. Journal of Machine Learning Research , 24(240): 1-113, 2023. URL http://jmlr.org/papers/v24/22-1144.html . - Christopher Clark, Kenton Lee, Ming-Wei Chang, Tom Kwiatkowski, Michael Collins, and Kristina Toutanova. BoolQ: Exploring the surprising difficulty of natural yes/no questions. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long and Short Papers) , pages 2924-2936, 2019. URL https://aclanthology.org/N19-1300 . - Jon Clark, Eunsol Choi, Michael Collins, Dan Garrette, Tom Kwiatkowski, Vitaly Nikolaev, and Jennimaria Palomaki. TydiQA: A benchmark for information-seeking question answering in typologically diverse languages. Transactions of the Association for Computational Linguistics , 2020. URL https://storage.googleapis.com/tydiqa/tydiqa.pdf . - Karl Cobbe, Vineet Kosaraju, Mohammad Bavarian, Jacob Hilton, Reiichiro Nakano, Christopher Hesse, and John Schulman. Training verifiers to solve math word problems. arXiv preprint arXiv:2110.14168 , 2021. URL https://arxiv.org/abs/2110.14168 . - Alexis Conneau, Min Ma, Simran Khanuja, Yu Zhang, Vera Axelrod, Siddharth Dalmia, Jason Riesa, Clara Rivera, and Ankur Bapna. Fleurs: Few-shot learning evaluation of universal representations of speech. In 2022 IEEE Spoken Language Technology Workshop (SLT) , pages 798-805. IEEE, 2023. - Jeff Dean. Introducing Pathways: A next-generation AI architecture, 2021. URL https://blog.google/technology/ai/ introducing-pathways-next-generation-ai-architecture/ . - Jeffrey Dean, Greg Corrado, Rajat Monga, Kai Chen, Matthieu Devin, Mark Mao, Marc'aurelio Ranzato, Andrew Senior, Paul Tucker, Ke Yang, et al. Large scale distributed deep networks. Advances in neural information processing systems , 25, 2012. - Harish Dattatraya Dixit, Sneha Pendharkar, Matt Beadon, Chris Mason, Tejasvi Chakravarthy, Bharath Muthiah, and Sriram Sankar. Silent data corruptions at scale. arXiv preprint arXiv:2102.11245 , 2021. - Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. In ICLR , 2020. - Dheeru Dua, Yizhong Wang, Pradeep Dasigi, Gabriel Stanovsky, Sameer Singh, and Matt Gardner. DROP: A reading comprehension benchmark requiring discrete reasoning over paragraphs. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long and Short Papers) , pages 2368-2378, 2019. URL https://aclanthology.org/N19-1246 . - Christian Federmann, Tom Kocmi, and Ying Xin. NTREX-128 - news test references for MT evaluation of 128 languages. In Proceedings of the First Workshop on Scaling Up Multilingual Evaluation , pages 21-24, Online, nov 2022. Association for Computational Linguistics. URL https://aclanthology.org/2022.sumeval-1.4 . - Samuel Gehman, Suchin Gururangan, Maarten Sap, Yejin Choi, and Noah A. Smith. Realtoxicityprompts: Evaluating neural toxic degeneration in language models, 2020. - Amelia Glaese, Nat McAleese, Maja Trębacz, John Aslanides, Vlad Firoiu, Timo Ewalds, Maribeth Rauh, Laura Weidinger, Martin Chadwick, Phoebe Thacker, Lucy Campbell-Gillingham, Jonathan Uesato, Po-Sen Huang, Ramona Comanescu, Fan Yang, Abigail See, Sumanth Dathathri, Rory Greig, Charlie Chen, Doug Fritz, Jaume Sanchez Elias, Richard Green, Soňa Mokrá, Nicholas Fernando, Boxi Wu, Rachel Foley, Susannah Young, Iason Gabriel, William Isaac, John Mellor, Demis Hassabis, Koray Kavukcuoglu, Lisa Anne Hendricks, and Geoffrey Irving. Improving alignment of dialogue agents via targeted human judgements, 2022. URL https://arxiv.org/abs/2209.14375 . - Google. Google's AI Principles. 2023. URL https://ai.google/responsibility/ principles/ . - Yash Goyal, Tejas Khot, Douglas Summers-Stay, Dhruv Batra, and Devi Parikh. Making the V in VQA matter: Elevating the role of image understanding in visual question answering. In Proceedings of the IEEE conference on computer vision and pattern recognition , pages 6904-6913, 2017. - Tahmid Hasan, Abhik Bhattacharjee, Md. Saiful Islam, Kazi Mubasshir, Yuan-Fang Li, Yong-Bin Kang, M. Sohel Rahman, and Rifat Shahriyar. XL-sum: Large-scale multilingual abstractive summarization for 44 languages. In Findings of the Association for Computational Linguistics: ACL-IJCNLP 2021 , pages 4693-4703, Online, August 2021. Association for Computational Linguistics. doi: 10.18653/ v1/2021.findings-acl.413. URL https://aclanthology.org/2021.findings-acl.413 . - Qianyu He, Jie Zeng, Wenhao Huang, Lina Chen, Jin Xiao, Qianxi He, Xunzhe Zhou, Lida Chen, Xintao Wang, Yuncheng Huang, et al. Can large language models understand real-world complex instructions? arXiv preprint arXiv:2309.09150 , 2023. - Dan Hendrycks, Collin Burns, Steven Basart, Andy Zou, Mantas Mazeika, Dawn Song, and Jacob Steinhardt. Measuring massive multitask language understanding. Proceedings of the International Conference on Learning Representations (ICLR) , 2021a. - Dan Hendrycks, Collin Burns, Saurav Kadavath, Akul Arora, Steven Basart, Eric Tang, Dawn Song, and Jacob Steinhardt. Measuring mathematical problem solving with the MATH dataset. arXiv preprint arXiv:2103.03874 , 2021b. URL https://arxiv.org/abs/2103.03874 . - Peter H Hochschild, Paul Turner, Jeffrey C Mogul, Rama Govindaraju, Parthasarathy Ranganathan, David E Culler, and Amin Vahdat. Cores that don't count. In Proceedings of the Workshop on Hot Topics in Operating Systems , pages 9-16, 2021. - Jordan Hoffmann, Sebastian Borgeaud, Arthur Mensch, Elena Buchatskaya, Trevor Cai, Eliza Rutherford, Diego de Las Casas, Lisa Anne Hendricks, Johannes Welbl, Aidan Clark, Tom Hennigan, Eric Noland, Katie Millican, George van den Driessche, Bogdan Damoc, Aurelia Guy, Simon Osindero, Karen Simonyan, Erich Elsen, Jack W. Rae, Oriol Vinyals, and Laurent Sifre. Training computeoptimal large language models. arXiv preprint arXiv:2203.15556 , 2022. - Shengding Hu, Yifan Luo, Huadong Wang, Xingyi Cheng, Zhiyuan Liu, and Maosong Sun. Won't get fooled again: Answering questions with false premises. arXiv preprint arXiv:2307.02394 , 2023. - EunJeong Hwang and Vered Shwartz. Memecap: A dataset for captioning and interpreting memes, 2023. - Norman P. Jouppi, Doe Hyun Yoon, George Kurian, Sheng Li, Nishant Patil, James Laudon, Cliff Young, and David A. Patterson. A domain-specific supercomputer for training deep neural networks. Commun. ACM , 63(7):67-78, 2020. doi: 10.1145/3360307. URL https://doi.org/10.1145/ 3360307 . - Norman P Jouppi, George Kurian, Sheng Li, Peter Ma, Rahul Nagarajan, Lifeng Nai, Nishant Patil, Suvinay Subramanian, Andy Swing, Brian Towles, Cliff Young, Xiang Zhou, Zongwei Zhou, and David A Patterson. Tpu v4: An optically reconfigurable supercomputer for machine learning with hardware support for embeddings. In Proceedings of the 50th Annual International Symposium on Computer Architecture , pages 1-14, 2023. - Ashwin Kalyan, Abhinav Kumar, Arjun Chandrasekaran, Ashish Sabharwal, and Peter Clark. How Much Coffee Was Consumed During EMNLP 2019? Fermi Problems: A New Reasoning Challenge for AI, 2021. - Jungo Kasai, Keisuke Sakaguchi, Yoichi Takahashi, Ronan Le Bras, Akari Asai, Xinyan Yu, Dragomir Radev, Noah A Smith, Yejin Choi, and Kentaro Inui. Realtime qa: What's the answer right now? arXiv preprint arXiv:2207.13332 , 2022a. - Jungo Kasai, Keisuke Sakaguchi, Yoichi Takahashi, Ronan Le Bras, Akari Asai, Xinyan Yu, Dragomir Radev, Noah A. Smith, Yejin Choi, and Kentaro Inui. RealTime QA: What's the answer right now?, 2022b. URL https://arxiv.org/abs/2207.13332 . - K Kavukcuoglu, P Kohli, L Ibrahim, D Bloxwich, and S Brown. How our principles helped define AlphaFold's release. Google DeepMind, 2022. - Aniruddha Kembhavi, Mike Salvato, Eric Kolve, Minjoon Seo, Hannaneh Hajishirzi, and Ali Farhadi. A diagram is worth a dozen images. In ECCV , 2016. - Megan Kinniment, Lucas Jun Koba Sato, Haoxing Du, Brian Goodrich, Max Hasin, Lawrence Chan, Luke Harold Miles, Tao R Lin, Hjalmar Wijk, Joel Burget, et al. Evaluating language-model agents on realistic autonomous tasks. arXiv preprint arXiv:2312.11671 , 2023. - Tomáš Kočiský, Jonathan Schwarz, Phil Blunsom, Chris Dyer, Karl Moritz Hermann, Gábor Melis, and Edward Grefenstette. The NarrativeQA reading comprehension challenge. Transactions of the Association for Computational Linguistics , 6:317-328, 2018. doi: 10.1162/tacl\_a\_00023. URL https://aclanthology.org/Q18-1023 . - Tom Kocmi, Rachel Bawden, Ondřej Bojar, Anton Dvorkovich, Christian Federmann, Mark Fishel, Thamme Gowda, Yvette Graham, Roman Grundkiewicz, Barry Haddow, Rebecca Knowles, Philipp Koehn, Christof Monz, Makoto Morishita, Masaaki Nagata, Toshiaki Nakazawa, Michal Novák, Martin Popel, and Maja Popović. Findings of the 2022 conference on machine translation (WMT22). In Proceedings of the Seventh Conference on Machine Translation (WMT) , December 2022. URL https://aclanthology.org/2022.wmt-1.1 . - Takeshi Kojima, Shixiang Shane Gu, Machel Reid, Yutaka Matsuo, and Yusuke Iwasawa. Large language models are zero-shot reasoners. Advances in neural information processing systems , 35: 22199-22213, 2022. - Taku Kudo and John Richardson. Sentencepiece: A simple and language independent subword tokenizer and detokenizer for neural text processing. In Eduardo Blanco and Wei Lu, editors, Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing, EMNLP 2018: System Demonstrations, Brussels, Belgium, October 31 - November 4, 2018 , pages 66-71. Association for Computational Linguistics, 2018. doi: 10.18653/v1/d18-2012. URL https: //doi.org/10.18653/v1/d18-2012 . - Tom Kwiatkowski, Jennimaria Palomaki, Olivia Redfield, Michael Collins, Ankur Parikh, Chris Alberti, Danielle Epstein, Illia Polosukhin, Jacob Devlin, Kenton Lee, Kristina Toutanova, Llion Jones, Matthew Kelcey, Ming-Wei Chang, Andrew M. Dai, Jakob Uszkoreit, Quoc Le, and Slav Petrov. Natural questions: A benchmark for question answering research. Transactions of the Association for Computational Linguistics , 7:452-466, 2019a. doi: 10.1162/tacl\_a\_00276. URL https: //aclanthology.org/Q19-1026 . - Tom Kwiatkowski, Jennimaria Palomaki, Olivia Redfield, Michael Collins, Ankur Parikh, Chris Alberti, Danielle Epstein, Illia Polosukhin, Jacob Devlin, Kenton Lee, et al. Natural questions: a benchmark for question answering research. Transactions of the Association for Computational Linguistics , 7: 453-466, 2019b. - Faisal Ladhak, Esin Durmus, Claire Cardie, and Kathleen McKeown. WikiLingua: A new benchmark dataset for cross-lingual abstractive summarization. In Findings of the Association for Computational Linguistics: EMNLP 2020 , pages 4034-4048, Online, November 2020. Association for Computational Linguistics. doi: 10.18653/v1/2020.findings-emnlp.360. URL https://www.aclweb.org/ anthology/2020.findings-emnlp.360 . - Leblond et al. AlphaCode 2 Technical Report. 2023. URL https://storage.googleapis.com/ deepmind-media/AlphaCode2/AlphaCode2\_Tech\_Report.pdf . - Yann LeCun, Yoshua Bengio, and Geoffrey Hinton. Deep learning. nature , 521(7553):436-444, 2015. - Yujia Li, David Choi, Junyoung Chung, Nate Kushman, Julian Schrittwieser, Rémi Leblond, Tom Eccles, James Keeling, Felix Gimeno, Agustin Dal Lago, et al. Competition-level code generation with alphacode. Science , 378(6624):1092-1097, 2022. - Bin Lin, Bin Zhu, Yang Ye, Munan Ning, Peng Jin, and Li Yuan. Video-llava: Learning united visual representation by alignment before projection. arXiv preprint arXiv:2311.10122 , 2023. - Fangyu Liu, Julian Eisenschlos, Francesco Piccinno, Syrine Krichene, Chenxi Pang, Kenton Lee, Mandar Joshi, Wenhu Chen, Nigel Collier, and Yasemin Altun. DePlot: One-shot visual language reasoning by plot-to-table translation. In Findings of the Association for Computational Linguistics: ACL 2023 , pages 10381-10399, Toronto, Canada, July 2023. Association for Computational Linguistics. doi: 10.18653/v1/2023.findings-acl.660. URL https://aclanthology.org/2023. findings-acl.660 . - Pan Lu, Ran Gong, Shibiao Jiang, Liang Qiu, Siyuan Huang, Xiaodan Liang, and Song-Chun Zhu. Inter-gps: Interpretable geometry problem solving with formal language and symbolic reasoning. In The Joint Conference of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing (ACL-IJCNLP 2021) , 2021. - Pan Lu, Hritik Bansal, Tony Xia, Jiacheng Liu, Chunyuan Li, Hannaneh Hajishirzi, Hao Cheng, KaiWei Chang, Michel Galley, and Jianfeng Gao. Mathvista: Evaluating mathematical reasoning of foundation models in visual contexts. arXiv preprint arXiv:2310.02255 , 2023. - Ahmed Masry, Do Long, Jia Qing Tan, Shafiq Joty, and Enamul Hoque. ChartQA: A benchmark for question answering about charts with visual and logical reasoning. In Findings of ACL , 2022. - Minesh Mathew, Dimosthenis Karatzas, and CV Jawahar. Docvqa: A dataset for vqa on document images. In Proceedings of the IEEE/CVF winter conference on applications of computer vision , pages 2200-2209, 2021. - Minesh Mathew, Viraj Bagal, Rubèn Tito, Dimosthenis Karatzas, Ernest Valveny, and CV Jawahar. Infographicvqa. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision , pages 1697-1706, 2022. - Joshua Maynez, Shashi Narayan, Bernd Bohnet, and Ryan McDonald. On faithfulness and factuality in abstractive summarization. arXiv preprint arXiv:2005.00661 , 2020. - Jacob Menick, Maja Trebacz, Vladimir Mikulik, John Aslanides, Francis Song, Martin Chadwick, Mia Glaese, Susannah Young, Lucy Campbell-Gillingham, Geoffrey Irving, and Nat McAleese. Teaching language models to support answers with verified quotes. arXiv preprint arXiv:2203.11147 , 2022. - Todor Mihaylov, Peter Clark, Tushar Khot, and Ashish Sabharwal. Can a suit of armor conduct electricity? a new dataset for open book question answering. In Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing , pages 2381-2391, Brussels, Belgium, OctoberNovember 2018. Association for Computational Linguistics. doi: 10.18653/v1/D18-1260. URL https://aclanthology.org/D18-1260 . - Swaroop Mishra, Daniel Khashabi, Chitta Baral, and Hannaneh Hajishirzi. Cross-task generalization via natural language crowdsourcing instructions. arXiv preprint arXiv:2104.08773 , 2021. - Shashi Narayan, Shay B. Cohen, and Mirella Lapata. Don't give me the details, just the summary! topic-aware convolutional neural networks for extreme summarization. In Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing , pages 1797-1807, Brussels, Belgium, October-November 2018. Association for Computational Linguistics. doi: 10.18653/v1/ D18-1206. URL https://aclanthology.org/D18-1206 . - Oktatási Hivatal. Matematika írásbéli vizsga. Középszintű Írásbéli Vizsga, May 2023. URL https://dload-oktatas.educatio.hu/erettsegi/feladatok\_2023tavasz\_kozep/ k\_matang\_23maj\_fl.pdf . Angol Nyelven. - OpenAI. GPT-4 Technical Report. 2023a. - OpenAI. GPT-4V(ision) System Card, 2023b. - OpenAI. Whisper, 2023. URL https://github.com/openai/whisper . - Long Ouyang, Jeffrey Wu, Xu Jiang, Diogo Almeida, Carroll Wainwright, Pamela Mishkin, Chong Zhang, Sandhini Agarwal, Katarina Slama, Alex Gray, John Schulman, Jacob Hilton, Fraser Kelton, Luke Miller, Maddie Simens, Amanda Askell, Peter Welinder, Paul Christiano, Jan Leike, and Ryan Lowe. Training language models to follow instructions with human feedback. In Alice H. Oh, Alekh Agarwal, Danielle Belgrave, and Kyunghyun Cho, editors, Advances in Neural Information Processing Systems , 2022. URL https://openreview.net/forum?id=TG8KACxEON . - Denis Paperno, Germán Kruszewski, Angeliki Lazaridou, Quan Ngoc Pham, Raffaella Bernardi, Sandro Pezzelle, Marco Baroni, Gemma Boleda, and Raquel Fernández. The LAMBADA dataset: Word prediction requiring a broad discourse context. arXiv preprint arXiv:1606.06031 , 2016. - Alicia Parrish, Angelica Chen, Nikita Nangia, Vishakh Padmakumar, Jason Phang, Jana Thompson, Phu Mon Htut, and Samuel R. Bowman. BBQ: A hand-built bias benchmark for question answering. CoRR , abs/2110.08193, 2021. URL https://arxiv.org/abs/2110.08193 . - Viorica Pătrăucean, Lucas Smaira, Ankush Gupta, Adrià Recasens Continente, Larisa Markeeva, Dylan Banarse, Skanda Koppula, Joseph Heyward, Mateusz Malinowski, Yi Yang, Carl Doersch, Tatiana Matejovicova, Yury Sulsky, Antoine Miech, Alex Frechette, Hanna Klimczak, Raphael Koster, Junlin Zhang, Stephanie Winkler, Yusuf Aytar, Simon Osindero, Dima Damen, Andrew Zisserman, and Joăo Carreira. Perception test: A diagnostic benchmark for multimodal video models. arXiv preprint arXiv:2305.13786 , 2023. - Baolin Peng, Michel Galley, Pengcheng He, Hao Cheng, Yujia Xie, Yu Hu, Qiuyuan Huang, Lars Liden, Zhou Yu, Weizhu Chen, et al. Check your facts and try again: Improving large language models with external knowledge and automated feedback. arXiv preprint arXiv:2302.12813 , 2023. - Leon Poutievski, Omid Mashayekhi, Joon Ong, Arjun Singh, Mukarram Tariq, Rui Wang, Jianan Zhang, Virginia Beauregard, Patrick Conner, Steve Gribble, et al. Jupiter evolving: transforming google's datacenter network via optical circuit switches and software-defined networking. In Proceedings of the ACM SIGCOMM 2022 Conference , pages 66-85, 2022. - Vineel Pratap, Qiantong Xu, Anuroop Sriram, Gabriel Synnaeve, and Ronan Collobert. Mls: A large-scale multilingual dataset for speech research. arXiv preprint arXiv:2012.03411 , 2020. - Alec Radford, Jeffrey Wu, Rewon Child, David Luan, Dario Amodei, and Ilya Sutskever. Language models are unsupervised multitask learners. OpenAI blog , 1(8):9, 2019. URL https://d4mucfpksywv.cloudfront.net/better-language-models/language\_ models\_are\_unsupervised\_multitask\_learners.pdf . - Alec Radford, Jong Wook Kim, Tao Xu, Greg Brockman, Christine McLeavey, and Ilya Sutskever. Robust speech recognition via large-scale weak supervision. In International Conference on Machine Learning , pages 28492-28518. PMLR, 2023. - Jack W. Rae, Sebastian Borgeaud, Trevor Cai, Katie Millican, Jordan Hoffmann, H. Francis Song, John Aslanides, Sarah Henderson, Roman Ring, Susannah Young, Eliza Rutherford, Tom Hennigan, Jacob Menick, Albin Cassirer, Richard Powell, George van den Driessche, Lisa Anne Hendricks, Maribeth Rauh, Po-Sen Huang, Amelia Glaese, Johannes Welbl, Sumanth Dathathri, Saffron Huang, Jonathan Uesato, John Mellor, Irina Higgins, Antonia Creswell, Nat McAleese, Amy Wu, Erich Elsen, Siddhant M. Jayakumar, Elena Buchatskaya, David Budden, Esme Sutherland, Karen Simonyan, Michela Paganini, Laurent Sifre, Lena Martens, Xiang Lorraine Li, Adhiguna Kuncoro, Aida Nematzadeh, Elena Gribovskaya, Domenic Donato, Angeliki Lazaridou, Arthur Mensch, JeanBaptiste Lespiau, Maria Tsimpoukelli, Nikolai Grigorev, Doug Fritz, Thibault Sottiaux, Mantas Pajarskas, Toby Pohlen, Zhitao Gong, Daniel Toyama, Cyprien de Masson d'Autume, Yujia Li, Tayfun Terzi, Vladimir Mikulik, Igor Babuschkin, Aidan Clark, Diego de Las Casas, Aurelia Guy, Chris Jones, James Bradbury, Matthew Johnson, Blake A. Hechtman, Laura Weidinger, Iason Gabriel, William S. Isaac, Edward Lockhart, Simon Osindero, Laura Rimell, Chris Dyer, Oriol Vinyals, Kareem Ayoub, Jeff Stanway, Lorrayne Bennett, Demis Hassabis, Koray Kavukcuoglu, and Geoffrey Irving. Scaling language models: Methods, analysis & insights from training Gopher. CoRR , abs/2112.11446, 2021. - Aditya Ramesh, Mikhail Pavlov, Gabriel Goh, Scott Gray, Chelsea Voss, Alec Radford, Mark Chen, and Ilya Sutskever. Zero-shot text-to-image generation. In International Conference on Machine Learning , pages 8821-8831. PMLR, 2021. - Hannah Rashkin, Vitaly Nikolaev, Matthew Lamm, Lora Aroyo, Michael Collins, Dipanjan Das, Slav Petrov, Gaurav Singh Tomar, Iulia Turc, and David Reitter. Measuring attribution in natural language generation models. Computational Linguistics , pages 1-64, 2023. - Scott Reed, Konrad Zolna, Emilio Parisotto, Sergio Gomez Colmenarejo, Alexander Novikov, Gabriel Barth-Maron, Mai Gimenez, Yury Sulsky, Jackie Kay, Jost Tobias Springenberg, Tom Eccles, Jake Bruce, Ali Razavi, Ashley Edwards, Nicolas Heess, Yutian Chen, Raia Hadsell, Oriol Vinyals, Mahyar Bordbar, and Nando de Freitas. A generalist agent. arXiv preprint arXiv:2205.06175 , 2022. - Parker Riley, Timothy Dozat, Jan A Botha, Xavier Garcia, Dan Garrette, Jason Riesa, Orhan Firat, and Noah Constant. Frmt: A benchmark for few-shot region-aware machine translation. Transactions of the Association for Computational Linguistics , 2023. - Hannah Ritchie, Veronika Samborska, and Max Roser. Plastic pollution. Our World in Data , 2023. https://ourworldindata.org/plastic-pollution. - Adam Roberts, Colin Raffel, and Noam Shazeer. How much knowledge can you pack into the parameters of a language model? In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP) , pages 5418-5426, Online, November 2020. Association for Computational Linguistics. doi: 10.18653/v1/2020.emnlp-main.437. URL https: //aclanthology.org/2020.emnlp-main.437 . - William A Gaviria Rojas, Sudnya Diamos, Keertan Ranjan Kini, David Kanter, Vijay Janapa Reddi, and Cody Coleman. The dollar street dataset: Images representing the geographic and socioeconomic diversity of the world. In Thirty-sixth Conference on Neural Information Processing Systems Datasets and Benchmarks Track , 2022. - Rachel Rudinger, Jason Naradowsky, Brian Leonard, and Benjamin Van Durme. Gender bias in coreference resolution. In Proceedings of the 2018 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 2 (Short Papers) , pages 8-14, New Orleans, Louisiana, June 2018. Association for Computational Linguistics. doi: 10.18653/v1/N18-2002. URL https://aclanthology.org/N18-2002 . - Candice Schumann, Susanna Ricco, Utsav Prabhu, Vittorio Ferrari, and Caroline Pantofaru. A step toward more inclusive people annotations for fairness. In Proceedings of the 2021 AAAI/ACM Conference on AI, Ethics, and Society , pages 916-925, 2021. - Andrew D. Selbst, Danah Boyd, and Sorelle A. Friedler. Fairness and abstraction in sociotechnical systems. In FFAT* '19: Proceedings of the Conference on Fairness, Accountability, and Transparency , pages 59-68, January 2019. - Thibault Sellam, Dipanjan Das, and Ankur Parikh. BLEURT: Learning robust metrics for text generation. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics , pages 7881-7892, Online, July 2020. Association for Computational Linguistics. doi: 10.18653/v1/2020. acl-main.704. URL https://aclanthology.org/2020.acl-main.704 . - Uri Shaham, Elad Segal, Maor Ivgi, Avia Efrat, Ori Yoran, Adi Haviv, Ankit Gupta, Wenhan Xiong, Mor Geva, Jonathan Berant, and Omer Levy. SCROLLS: Standardized CompaRison over long language sequences. In Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing , pages 12007-12021, Abu Dhabi, United Arab Emirates, December 2022. Association for Computational Linguistics. URL https://aclanthology.org/2022.emnlp-main.823 . - Noam Shazeer. Fast transformer decoding: One write-head is all you need. arXiv preprint arXiv:1911.02150 , 2019a. - Noam Shazeer. Fast transformer decoding: One write-head is all you need. arXiv preprint arXiv:1911.02150 , 2019b. - Renee Shelby, Shalaleh Rismani, Kathryn Henne, AJung Moon, Negar Rostamzadeh, Paul Nicholas, N'Mah Yilla, Jess Gallegos, Andrew Smart, Emilio Garcia, and Gurleen Virk. Identifying sociotechnical harms of algorithmic systems: Scoping a taxonomy for harm reduction, 2023. URL https://arxiv.org/abs/2210.05791 . - Toby Shevlane, Sebastian Farquhar, Ben Garfinkel, Mary Phuong, Jess Whittlestone, Jade Leung, Daniel Kokotajlo, Nahema Marchal, Markus Anderljung, Noam Kolt, Lewis Ho, Divya Siddarth, Shahar Avin, Will Hawkins, Been Kim, Iason Gabriel, Vijay Bolina, Jack Clark, Yoshua Bengio, Paul Christiano, and Allan Dafoe. Model evaluation for extreme risks. arXiv preprint arXiv:2305.15324 , 2023. - Freda Shi, Mirac Suzgun, Markus Freitag, Xuezhi Wang, Suraj Srivats, Soroush Vosoughi, Hyung Won Chung, Yi Tay, Sebastian Ruder, Denny Zhou, et al. Language models are multilingual chain-ofthought reasoners. ICLR , 2023. - Amanpreet Singh, Vivek Natarajan, Meet Shah, Yu Jiang, Xinlei Chen, Dhruv Batra, Devi Parikh, and Marcus Rohrbach. Towards VQA models that can read. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages 8317-8326, 2019. - Aarohi Srivastava, Abhinav Rastogi, Abhishek Rao, Abu Awal Md Shoeb, Abubakar Abid, Adam Fisch, Adam R. Brown, et al. Beyond the imitation game: Quantifying and extrapolating the capabilities of language models. arXiv preprint arXiv:2206.04615 , 2022. URL https://arxiv.org/abs/ 2206.04615 . - Ilya Sutskever, Oriol Vinyals, and Quoc V Le. Sequence to sequence learning with neural networks. Advances in neural information processing systems , 27, 2014. - Mirac Suzgun, Nathan Scales, Nathanael Schärli, Sebastian Gehrmann, Yi Tay, Hyung Won Chung, Aakanksha Chowdhery, Quoc V Le, Ed H Chi, Denny Zhou, et al. Challenging big-bench tasks and whether chain-of-thought can solve them. arXiv preprint arXiv:2210.09261 , 2022. - Oyvind Tafjord, Bhavana Dalvi, and Peter Clark. Proof Writer: Generating implications, proofs, and abductive statements over natural language. In Findings , 2020. URL https://api. semanticscholar.org/CorpusID:229371222 . - NLLB Team, Marta R. Costa-jussà, James Cross, Onur Çelebi, Maha Elbayad, Kenneth Heafield, Kevin Heffernan, Elahe Kalbassi, Janice Lam, Daniel Licht, Jean Maillard, Anna Sun, Skyler Wang, Guillaume Wenzek, Al Youngblood, Bapi Akula, Loic Barrault, Gabriel Mejia Gonzalez, Prangthip Hansanti, John Hoffman, Semarley Jarrett, Kaushik Ram Sadagopan, Dirk Rowe, Shannon Spruit, Chau Tran, Pierre Andrews, Necip Fazil Ayan, Shruti Bhosale, Sergey Edunov, Angela Fan, Cynthia Gao, Vedanuj Goswami, Francisco Guzmán, Philipp Koehn, Alexandre Mourachko, Christophe Ropers, Safiyyah Saleem, Holger Schwenk, and Jeff Wang. No language left behind: Scaling human-centered machine translation. 2022. - Ashish V. Thapliyal, Jordi Pont-Tuset, Xi Chen, and Radu Soricut. Crossmodal-3600: A massively multilingual multimodal evaluation dataset. In EMNLP , 2022. - Romal Thoppilan, Daniel De Freitas, Jamie Hall, Noam Shazeer, Apoorv Kulshreshtha, Heng-Tze Cheng, Alicia Jin, Taylor Bos, Leslie Baker, Yu Du, et al. LaMDA: Language models for dialog applications. arXiv preprint arXiv:2201.08239 , 2022. URL https://arxiv.org/abs/2201.08239 . - Kocmi Tom, Eleftherios Avramidis, Rachel Bawden, Ondřej Bojar, Anton Dvorkovich, Christian Federmann, Mark Fishel, Markus Freitag, Thamme Gowda, Roman Grundkiewicz, et al. Findings of the 2023 conference on machine translation (wmt23): Llms are here but not quite there yet. In WMT23-Eighth Conference on Machine Translation , pages 198-216, 2023. - Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothée Lacroix, Baptiste Rozière, Naman Goyal, Eric Hambro, Faisal Azhar, Aurelien Rodriguez, Armand Joulin, Edouard Grave, and Guillaume Lample. Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971 , 2023a. - Hugo Touvron, Louis Martin, Kevin Stone, Peter Albert, Amjad Almahairi, Yasmine Babaei, Nikolay Bashlykov, Soumya Batra, Prajjwal Bhargava, Shruti Bhosale, Dan Bikel, Lukas Blecher, Cristian Canton Ferrer, Moya Chen, Guillem Cucurull, David Esiobu, Jude Fernandes, Jeremy Fu, Wenyin Fu, Brian Fuller, Cynthia Gao, Vedanuj Goswami, Naman Goyal, Anthony Hartshorn, Saghar Hosseini, Rui Hou, Hakan Inan, Marcin Kardas, Viktor Kerkez, Madian Khabsa, Isabel Kloumann, Artem Korenev, Punit Singh Koura, Marie-Anne Lachaux, Thibaut Lavril, Jenya Lee, Diana Liskovich, Yinghai Lu, Yuning Mao, Xavier Martinet, Todor Mihaylov, Pushkar Mishra, Igor Molybog, Yixin Nie, Andrew Poulton, Jeremy Reizenstein, Rashi Rungta, Kalyan Saladi, Alan Schelten, Ruan Silva, Eric Michael Smith, Ranjan Subramanian, Xiaoqing Ellen Tan, Binh Tang, Ross Taylor, Adina Williams, Jian Xiang Kuan, Puxin Xu, Zheng Yan, Iliyan Zarov, Yuchen Zhang, Angela Fan, Melanie Kambadur, Sharan Narang, Aurelien Rodriguez, Robert Stojnic, Sergey Edunov, and Thomas Scialom. Llama 2: Open foundation and fine-tuned chat models. arXiv preprint arXiv:2307.09288 , 2023b. - Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Ł ukasz Kaiser, and Illia Polosukhin. Attention is all you need. In I. Guyon, U. Von Luxburg, S. Bengio, H. Wallach, R. Fergus, S. Vishwanathan, and R. Garnett, editors, Advances in Neural Information Processing Systems , volume 30. Curran Associates, Inc., 2017a. URL https://proceedings.neurips.cc/ paper\_files/paper/2017/file/3f5ee243547dee91fbd053c1c4a845aa-Paper.pdf . - Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. CoRR , abs/1706.03762, 2017b. URL http://arxiv.org/abs/1706.03762 . - Ramakrishna Vedantam, C Lawrence Zitnick, and Devi Parikh. Cider: Consensus-based image description evaluation. In Proceedings of the IEEE conference on computer vision and pattern recognition , pages 4566-4575, 2015. - Petar Veličković, Adrià Puigdomènech Badia, David Budden, Razvan Pascanu, Andrea Banino, Misha Dashevskiy, Raia Hadsell, and Charles Blundell. The clrs algorithmic reasoning benchmark. arXiv preprint arXiv:2205.15659 , 2022. - Manoj Vishwanathan, Ronak Shah, Kyung Ki Kim, and Minsu Choi. Silent data corruption (sdc) vulnerability of gpu on various gpgpu workloads. In 2015 International SoC Design Conference (ISOCC) , pages 11-12, 2015. doi: 10.1109/ISOCC.2015.7401681. - Changhan Wang, Anne Wu, and Juan Pino. Covost 2 and massively multilingual speech-to-text translation. arXiv preprint arXiv:2007.10310 , 2020. - Changhan Wang, Morgane Riviere, Ann Lee, Anne Wu, Chaitanya Talnikar, Daniel Haziza, Mary Williamson, Juan Pino, and Emmanuel Dupoux. Voxpopuli: A large-scale multilingual speech corpus for representation learning, semi-supervised learning and interpretation. arXiv preprint arXiv:2101.00390 , 2021. - Xin Wang, Jiawei Wu, Junkun Chen, Lei Li, Yuan-Fang Wang, and William Yang Wang. VATEX: A large-scale, high-quality multilingual dataset for video-and-language research. In ICCV , 2019. - Xuezhi Wang, Jason Wei, Dale Schuurmans, Quoc Le, Ed Chi, and Denny Zhou. Self-consistency improves chain of thought reasoning in language models. arXiv preprint arXiv:2203.11171 , 2022. - Jason Wei, Maarten Bosma, Vincent Y Zhao, Kelvin Guu, Adams Wei Yu, Brian Lester, Nan Du, Andrew M Dai, and Quoc V Le. Finetuned language models are zero-shot learners. Proceedings of the International Conference on Learning Representations (ICLR) , 2022a. URL https://openreview. net/forum?id=gEZrGCozdqR . - Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Brian Ichter, Fei Xia, Ed Chi, Quoc Le, and Denny Zhou. Chain-of-thought prompting elicits reasoning in large language models. NeurIPS , 2022b. URL https://arxiv.org/abs/2201.11903 . - Laura Weidinger, John Mellor, Maribeth Rauh, Conor Griffin, Jonathan Uesato, Po-Sen Huang, Myra Cheng, Mia Glaese, Borja Balle, Atoosa Kasirzadeh, Zac Kenton, Sasha Brown, Will Hawkins, Tom Stepleton, Courtney Biles, Abeba Birhane, Julia Haas, Laura Rimell, Lisa Anne Hendricks, William S. Isaac, Sean Legassick, Geoffrey Irving, and Iason Gabriel. Ethical and social risks of harm from language models. CoRR , abs/2112.04359, 2021. URL https://arxiv.org/abs/2112.04359 . - David Wetherall, Abdul Kabbani, Van Jacobson, Jim Winget, Yuchung Cheng, Brad Morrey, Uma Parthavi Moravapalle, Phillipa Gill, Steven Knight, and Amin Vahdat. Improving network availability with protective reroute. In SIGCOMM 2023 , 2023. URL https://dl.acm.org/doi/ 10.1145/3603269.3604867 . - Junbin Xiao, Xindi Shang, Angela Yao, and Tat-Seng Chua. NExT-QA: Next phase of question-answering to explaining temporal actions. In CVPR , 2021. - XLA. XLA: Optimizing compiler for TensorFlow. https://www.tensorflow.org/xla , 2019. [Online; accessed December-2023]. - Can Xu, Qingfeng Sun, Kai Zheng, Xiubo Geng, Pu Zhao, Jiazhan Feng, Chongyang Tao, and Daxin Jiang. Wizardlm: Empowering large language models to follow complex instructions. arXiv preprint arXiv:2304.12244 , 2023. - Yuanzhong Xu, HyoukJoong Lee, Dehao Chen, Blake Hechtman, Yanping Huang, Rahul Joshi, Maxim Krikun, Dmitry Lepikhin, Andy Ly, Marcello Maggioni, et al. Gspmd: general and scalable parallelization for ml computation graphs. arXiv preprint arXiv:2105.04663 , 2021. - Chi yao Hong, Subhasree Mandal, Mohammad A. Alfares, Min Zhu, Rich Alimi, Kondapa Naidu Bollineni, Chandan Bhagat, Sourabh Jain, Jay Kaimal, Jeffrey Liang, Kirill Mendelev, Steve Padgett, Faro Thomas Rabe, Saikat Ray, Malveeka Tewari, Matt Tierney, Monika Zahn, Jon Zolla, Joon Ong, and Amin Vahdat. B4 and after: Managing hierarchy, partitioning, and asymmetry for availability and scale in google's software-defined wan. In SIGCOMM'18 , 2018. URL https: //conferences.sigcomm.org/sigcomm/2018/program\_tuesday.html . - Jiahui Yu, Zirui Wang, Vijay Vasudevan, Legg Yeung, Mojtaba Seyedhosseini, and Yonghui Wu. Coca: Contrastive captioners are image-text foundation models, 2022a. - Jiahui Yu, Yuanzhong Xu, Jing Yu Koh, Thang Luong, Gunjan Baid, Zirui Wang, Vijay Vasudevan, Alexander Ku, Yinfei Yang, Burcu Karagol Ayan, Ben Hutchinson, Wei Han, Zarana Parekh, Xin Li, Han Zhang, Jason Baldridge, and Yonghui Wu. Scaling autoregressive models for content-rich text-to-image generation. arXiv preprint arXiv:2206.10789 , 2(3):5, 2022b. - Shoubin Yu, Jaemin Cho, Prateek Yadav, and Mohit Bansal. Self-chained image-language model for video localization and question answering. arXiv preprint arXiv:2305.06988 , 2023. - Zhou Yu, Dejing Xu, Jun Yu, Ting Yu, Zhou Zhao, Yueting Zhuang, and Dacheng Tao. ActivityNet-QA: A dataset for understanding complex web videos via question answering. In AAAI , 2019. - Xiang Yue, Yuansheng Ni, Kai Zhang, Tianyu Zheng, Ruoqi Liu, Ge Zhang, Samuel Stevens, Dongfu Jiang, Weiming Ren, Yuxuan Sun, Cong Wei, Botao Yu, Ruibin Yuan, Renliang Sun, Ming Yin, Boyuan Zheng, Zhenzhu Yang, Yibo Liu, Wenhao Huang, Huan Sun, Yu Su, and Wenhu Chen. Mmmu: A massive multi-discipline multimodal understanding and reasoning benchmark for expert agi, 2023. - Rowan Zellers, Ari Holtzman, Yonatan Bisk, Ali Farhadi, and Yejin Choi. Hellaswag: Can a machine really finish your sentence? arXiv preprint arXiv:1905.07830 , 2019. - Yu Zhang, Wei Han, James Qin, Yongqiang Wang, Ankur Bapna, Zhehuai Chen, Nanxin Chen, Bo Li, Vera Axelrod, Gary Wang, Zhong Meng, Ke Hu, Andrew Rosenberg, Rohit Prabhavalkar, Daniel S. Park, Parisa Haghani, Jason Riesa, Ginger Perng, Hagen Soltau, Trevor Strohman, Bhuvana Ramabhadran, Tara Sainath, Pedro Moreno, Chung-Cheng Chiu, Johan Schalkwyk, Françoise Beaufays, and Yonghui Wu. Google usm: Scaling automatic speech recognition beyond 100 languages. arXiv preprint arXiv:2303.01037 , 2023. - Dora Zhao, Angelina Wang, and Olga Russakovsky. Understanding and evaluating racial biases in image captioning. In Proceedings of the IEEE/CVF International Conference on Computer Vision , pages 14830-14840, 2021. - Jieyu Zhao, Tianlu Wang, Mark Yatskar, Vicente Ordonez, and Kai-Wei Chang. Gender bias in coreference resolution: Evaluation and debiasing methods. arXiv preprint arXiv:1804.06876 , 2018. - Chuanyang Zheng, Zhengying Liu, Enze Xie, Zhenguo Li, and Yu Li. Progressive-hint prompting improves reasoning in large language models, 2023. - Luowei Zhou, Chenliang Xu, and Jason J Corso. Towards automatic learning of procedures from web instructional videos. In AAAI Conference on Artificial Intelligence , pages 7590-7598, 2018. ## 9. Contributions and Acknowledgments ## Gemini Leads Rohan Anil, Co-Lead , Text Sebastian Borgeaud, Co-Lead , Text Jean-Baptiste Alayrac, Co-Lead , MM Vision Jiahui Yu, Co-Lead , MM Vision Radu Soricut, Co-Lead , MM Vision Johan Schalkwyk, Lead , MM Audio Andrew M. Dai, Co-Lead , Data Anja Hauth, Co-Lead , Data Katie Millican, Co-Lead , Data David Silver, Co-Lead , Fine-Tuning Melvin Johnson, Lead , Instruction Tuning Ioannis Antonoglou, Co-Lead , RL Techniques Julian Schrittwieser, Co-Lead , RL Techniques Amelia Glaese, Lead , Human Data Jilin Chen, Lead , Safety Emily Pitler, Co-Lead , Tool Use Timothy Lillicrap, Co-Lead , Tool Use Angeliki Lazaridou, Co-Lead , Eval Orhan Firat, Co-Lead , Eval James Molloy, Co-Lead , Infra Michael Isard, Co-Lead , Infra Paul R. Barham, Co-Lead , Infra Tom Hennigan, Co-Lead , Infra Benjamin Lee, Co-Lead , Codebase & Parallelism Fabio Viola, Co-Lead , Codebase & Parallelism Malcolm Reynolds, Co-Lead , Codebase & Parallelism Yuanzhong Xu, Co-Lead , Codebase & Parallelism Ryan Doherty, Lead , Ecosystem Eli Collins, Lead , Product Clemens Meyer, Co-Lead , Operations Eliza Rutherford, Co-Lead , Operations Erica Moreira, Co-Lead , Operations Kareem Ayoub, Co-Lead , Operations Megha Goel, Co-Lead , Operations ## Gemini App Leads Jack Krawczyk, Lead , Gemini App Product Cosmo Du, Co-Lead , Gemini App Research Ed Chi, Co-Lead , Gemini App Research Heng-Tze Cheng, Co-Lead , Gemini App Research Eric Ni, Lead , Gemini App Research Technical Program Management Purvi Shah, Lead , Gemini App Technical Program Management Patrick Kane, Co-Lead , Gemini App Core Modeling, Eval, Data, Product Betty Chan, Co-Lead , Gemini App Core Modeling, Technical Program Management Manaal Faruqui, Co-Lead , Gemini App Core Modeling, Factuality, Instruction Following Aliaksei Severyn, Co-Lead , Gemini App Core Modeling, Conversationality Hanzhao Lin, Co-Lead , Gemini App Fine-Tuning YaGuang Li, Co-Lead , Gemini App Fine-Tuning Yong Cheng, Co-Lead , Gemini App Fine-Tuning Abe Ittycheriah, Co-Lead , Gemini for Gemini App Mahdis Mahdieh, Co-Lead , Gemini for Gemini App Mia Chen, Co-Lead , Gemini for Gemini App Pei Sun, Co-Lead , Gemini for Gemini App Dustin Tran, Co-Lead , Gemini App Eval Sumit Bagri, Co-Lead , Gemini App Eval, Technical Program Management Balaji Lakshminarayanan, Co-Lead , Gemini App AutoEval Jeremiah Liu, Co-Lead , Gemini App AutoEval Andras Orban, Co-Lead , Gemini App Factuality, Multimodality, Safety Fabian Güra, Co-Lead , Gemini App Factuality Hao Zhou, Co-Lead , Gemini App Factuality Xinying Song, Co-Lead , Gemini App Factuality Aurelien Boffy, Co-Lead , Gemini App Safety Harish Ganapathy, Co-Lead , Gemini Safety Steven Zheng, Lead , Gemini App Multilinguality Research HyunJeong Choe, Lead , Gemini App Multilinguality Ágoston Weisz, Co-Lead , Gemini App Multimodality Tao Zhu, Co-Lead , Gemini App Multimodality Yifeng Lu, Co-Lead , Gemini App Multimodality Siddharth Gopal, Co-Lead , Gemini App Coding & Tool Use Jarrod Kahn, Co-Lead , Gemini App Tool Use Research Maciej Kula, Co-Lead , Gemini App Tool Use Research Jeff Pitman, Co-Lead , Gemini App Tool Use Rushin Shah, Co-Lead , Gemini App Tool Use Emanuel Taropa, Co-Lead , Gemini App Serving Majd Al Merey, Co-Lead , Gemini App Serving Martin Baeuml, Co-Lead , Gemini App Serving Zhifeng Chen, Co-Lead , Gemini App Serving Laurent El Shafey, Co-Lead , Gemini App Fine-Tuning Infra Yujing Zhang, Co-Lead , Gemini App Fine-Tuning Infra Olcan Sercinoglu, Lead , Gemini App Product ## Core Contributors George Tucker Enrique Piqueras Maxim Krikun Iain Barr Nikolay Savinov Ivo Danihelka Becca Roelofs Anaïs White Anders Andreassen Tamara von Glehn Lakshman Yagati Mehran Kazemi Lucas Gonzalez Misha Khalman Jakub Sygnowski Alexandre Frechette Charlotte Smith Laura Culp Lev Proleev Yi Luan Xi Chen James Lottes Nathan Schucher Federico Lebron Alban Rrustemi Natalie Clay Phil Crone Tomas Kocisky Jeffrey Zhao Bartek Perz Dian Yu Heidi Howard Adam Bloniarz Jack W. Rae Han Lu Laurent Sifre Marcello Maggioni Fred Alcober Dan Garrette Megan Barnes Shantanu Thakoor Jacob Austin Gabriel Barth-Maron William Wong Rishabh Joshi Rahma Chaabouni Deeni Fatiha Arun Ahuja ## Core Contributors Gaurav Singh Tomar Evan Senter Martin Chadwick Ilya Kornakov Nithya Attaluri Iñaki Iturrate Ruibo Liu Yunxuan Li Sarah Cogan Jeremy Chen Chao Jia Chenjie Gu Qiao Zhang Jordan Grimstad Ale Jakse Hartman Xavier Garcia Thanumalayan Sankaranarayana Pillai Jacob Devlin Michael Laskin Diego de Las Casas Dasha Valter Connie Tao Lorenzo Blanco Adrià Puigdomènech Badia David Reitter Mianna Chen Jenny Brennan Clara Rivera Sergey Brin Shariq Iqbal Gabriela Surita Jane Labanowski Abhi Rao Stephanie Winkler Emilio Parisotto Yiming Gu Kate Olszewska Ravi Addanki Antoine Miech Annie Louis Denis Teplyashin Geoff Brown Elliot Catt Jan Balaguer Jackie Xiang Pidong Wang Zoe Ashwood Anton Briukhov ## Core Contributors Albert Webson Sanjay Ganapathy Smit Sanghavi Ajay Kannan Ming-Wei Chang Axel Stjerngren Josip Djolonga Yuting Sun Ankur Bapna Matthew Aitchison Pedram Pejman Henryk Michalewski Tianhe Yu Cindy Wang Juliette Love Junwhan Ahn Dawn Bloxwich Kehang Han Peter Humphreys Thibault Sellam James Bradbury Varun Godbole Sina Samangooei Bogdan Damoc Alex Kaskasoli Sébastien M. R. Arnold Vijay Vasudevan Shubham Agrawal Jason Riesa Dmitry Lepikhin Richard Tanburn Srivatsan Srinivasan Hyeontaek Lim Sarah Hodkinson Pranav Shyam Johan Ferret Steven Hand Ankush Garg Tom Le Paine Jian Li Yujia Li Minh Giang Alexander Neitz Zaheer Abbas Sarah York Machel Reid Elizabeth Cole Aakanksha Chowdhery ## Core Contributors Dipanjan Das Dominika Rogozińska Vitaly Nikolaev Pablo Sprechmann Zachary Nado Lukas Zilka Flavien Prost Luheng He Marianne Monteiro Gaurav Mishra Chris Welty Josh Newlan Dawei Jia Miltiadis Allamanis Clara Huiyi Hu Raoul de Liedekerke Justin Gilmer Carl Saroufim Shruti Rijhwani Shaobo Hou Disha Shrivastava Anirudh Baddepudi Alex Goldin Adnan Ozturel Albin Cassirer Yunhan Xu Daniel Sohn Devendra Sachan Reinald Kim Amplayo Craig Swanson Dessie Petrova Shashi Narayan Arthur Guez Siddhartha Brahma Jessica Landon Miteyan Patel Ruizhe Zhao Kevin Villela Luyu Wang Wenhao Jia Matthew Rahtz Mai Giménez Legg Yeung James Keeling Petko Georgiev Diana Mincu Boxi Wu Salem Haykal ## Core Contributors Rachel Saputro Kiran Vodrahalli James Qin Zeynep Cankara Abhanshu Sharma Nick Fernando Will Hawkins Behnam Neyshabur Solomon Kim Adrian Hutter Priyanka Agrawal Alex Castro-Ros George van den Driessche Tao Wang Fan Yang Shuo-yiin Chang Paul Komarek Ross McIlroy Mario Lučić Guodong Zhang Wael Farhan Michael Sharman Paul Natsev Paul Michel Yamini Bansal Siyuan Qiao Kris Cao Siamak Shakeri Christina Butterfield Justin Chung Paul Kishan Rubenstein Shivani Agrawal Arthur Mensch Kedar Soparkar Karel Lenc Timothy Chung Aedan Pope Loren Maggiore Jackie Kay Priya Jhakra Shibo Wang Joshua Maynez Mary Phuong Taylor Tobin Andrea Tacchetti Maja Trebacz Kevin Robinson Yash Katariya Core Contributors Sebastian Riedel Paige Bailey Kefan Xiao Nimesh Ghelani Lora Aroyo Ambrose Slone Neil Houlsby Xuehan Xiong Zhen Yang Elena Gribovskaya Jonas Adler Mateo Wirth Lisa Lee Music Li Thais Kagohara Jay Pavagadhi Sophie Bridgers Anna Bortsova Sanjay Ghemawat Zafarali Ahmed Tianqi Liu Richard Powell Vijay Bolina Mariko Iinuma Polina Zablotskaia James Besley Da-Woon Chung Timothy Dozat Ramona Comanescu Xiance Si Jeremy Greer Guolong Su Martin Polacek Raphaël Lopez Kaufman Simon Tokumine Hexiang Hu Elena Buchatskaya Yingjie Miao Mohamed Elhawaty Aditya Siddhant Nenad Tomasev Jinwei Xing Christina Greer Helen Miller Shereen Ashraf Aurko Roy Zizhao Zhang Ada Ma ## Core Contributors Angelos Filos Milos Besta Rory Blevins Ted Klimenko Chih-Kuan Yeh Soravit Changpinyo Jiaqi Mu Oscar Chang Mantas Pajarskas Carrie Muir Vered Cohen Charline Le Lan Krishna Haridasan Amit Marathe Steven Hansen Sholto Douglas Rajkumar Samuel Mingqiu Wang Sophia Austin Chang Lan Jiepu Jiang Justin Chiu Jaime Alonso Lorenzo Lars Lowe Sjösund Sébastien Cevey Zach Gleicher Thi Avrahami Anudhyan Boral Hansa Srinivasan Vittorio Selo Rhys May Konstantinos Aisopos Léonard Hussenot Livio Baldini Soares Kate Baumli Michael B. Chang Adrià Recasens Ben Caine Alexander Pritzel Filip Pavetic Fabio Pardo Anita Gergely Justin Frye Vinay Ramasesh Dan Horgan Kartikeya Badola Nora Kassner Subhrajit Roy ## Core Contributors Ethan Dyer Víctor Campos Alex Tomala Yunhao Tang Dalia El Badawy Elspeth White Basil Mustafa Oran Lang Abhishek Jindal Sharad Vikram Zhitao Gong Sergi Caelles Ross Hemsley Gregory Thornton Fangxiaoyu Feng Wojciech Stokowiec Ce Zheng Phoebe Thacker Çağlar Ünlü Zhishuai Zhang Mohammad Saleh James Svensson Max Bileschi Piyush Patil Ankesh Anand Roman Ring Katerina Tsihlas Arpi Vezer Marco Selvi Toby Shevlane Mikel Rodriguez Tom Kwiatkowski Samira Daruki Keran Rong Allan Dafoe Nicholas FitzGerald Keren Gu-Lemberg Mina Khan Lisa Anne Hendricks Marie Pellat Vladimir Feinberg James Cobon-Kerr Tara Sainath Maribeth Rauh Sayed Hadi Hashemi Richard Ives Yana Hasson Eric Noland ## Core Contributors Yuan Cao Nathan Byrd Le Hou Qingze Wang Thibault Sottiaux Michela Paganini Jean-Baptiste Lespiau Alexandre Moufarek Samer Hassan Kaushik Shivakumar Joost van Amersfoort Amol Mandhane Pratik Joshi Anirudh Goyal Matthew Tung Andrew Brock Hannah Sheahan Vedant Misra Cheng Li Nemanja Rakićević Mostafa Dehghani Fangyu Liu Sid Mittal Junhyuk Oh Seb Noury Eren Sezener Fantine Huot Matthew Lamm Nicola De Cao Charlie Chen Sidharth Mudgal Romina Stella Kevin Brooks Gautam Vasudevan Chenxi Liu Mainak Chain Nivedita Melinkeri Aaron Cohen Venus Wang Kristie Seymore Sergey Zubkov Rahul Goel Summer Yue Sai Krishnakumaran Brian Albert Nate Hurley Motoki Sano Anhad Mohananey ## Core Contributors Jonah Joughin Egor Filonov Tomasz Kępa Yomna Eldawy Jiawern Lim Rahul Rishi Shirin Badiezadegan Taylor Bos Jerry Chang Sanil Jain Sri Gayatri Sundara Padmanabhan Subha Puttagunta Kalpesh Krishna Leslie Baker Norbert Kalb Vamsi Bedapudi Adam Kurzrok Shuntong Lei Anthony Yu Oren Litvin Xiang Zhou Zhichun Wu Sam Sobell Andrea Siciliano Alan Papir Robby Neale Jonas Bragagnolo Tej Toor Tina Chen Valentin Anklin Feiran Wang Richie Feng Milad Gholami Kevin Ling Lijuan Liu Jules Walter Hamid Moghaddam Arun Kishore Jakub Adamek Tyler Mercado Jonathan Mallinson Siddhinita Wandekar Stephen Cagle Eran Ofek Guillermo Garrido Clemens Lombriser Maksim Mukha Botu Sun ## Core Contributors Hafeezul Rahman Mohammad Josip Matak Yadi Qian Vikas Peswani Pawel Janus Quan Yuan Leif Schelin Oana David Ankur Garg Yifan He Oleksii Duzhyi Anton Älgmyr Timothée Lottaz Qi Li Vikas Yadav Luyao Xu Alex Chinien Rakesh Shivanna Aleksandr Chuklin Josie Li Carrie Spadine Travis Wolfe Kareem Mohamed Subhabrata Das Zihang Dai Kyle He Daniel von Dincklage Shyam Upadhyay Akanksha Maurya Luyan Chi Sebastian Krause Khalid Salama Pam G Rabinovitch Pavan Kumar Reddy M Aarush Selvan Mikhail Dektiarev Golnaz Ghiasi Erdem Guven Himanshu Gupta Boyi Liu Deepak Sharma Idan Heimlich Shtacher Shachi Paul Oscar Akerlund François-Xavier Aubet Terry Huang Chen Zhu Eric Zhu ## Core Contributors Elico Teixeira Matthew Fritze Francesco Bertolini Liana-Eleonora Marinescu Martin Bölle Dominik Paulus Khyatti Gupta Tejasi Latkar Max Chang Jason Sanders Roopa Wilson Xuewei Wu Yi-Xuan Tan Lam Nguyen Thiet Tulsee Doshi Sid Lall Swaroop Mishra Wanming Chen Thang Luong Seth Benjamin Jasmine (Sun Jae) Lee Ewa Andrejczuk Dominik Rabiej Vipul Ranjan Krzysztof Styrc Pengcheng Yin Jon Simon Malcolm Rose Harriott Mudit Bansal Alexei Robsky Geoff Bacon David Greene Daniil Mirylenka Chen Zhou Obaid Sarvana Abhimanyu Goyal Samuel Andermatt Patrick Siegler Ben Horn Assaf Israel Francesco Pongetti Chih-Wei 'Louis' Chen Marco Selvatici Pedro Silva Kathie Wang Jackson Tolins Kelvin Guu Roey Yogev ## Core Contributors Xiaochen Cai Alessandro Agostini Maulik Shah Hung Nguyen Noah Ó Donnaile Sébastien Pereira Linda Friso Adam Stambler Adam Kurzrok Chenkai Kuang Yan Romanikhin Mark Geller ZJ Yan Kane Jang Cheng-Chun Lee Wojciech Fica Eric Malmi Qijun Tan Dan Banica Daniel Balle Ryan Pham Yanping Huang Diana Avram Hongzhi Shi Jasjot Singh Chris Hidey Niharika Ahuja Pranab Saxena Dan Dooley Srividya Pranavi Potharaju Eileen O'Neill Anand Gokulchandran Ryan Foley Kai Zhao Mike Dusenberry Yuan Liu Pulkit Mehta Ragha Kotikalapudi Chalence Safranek-Shrader Andrew Goodman Joshua Kessinger Eran Globen Prateek Kolhar Chris Gorgolewski Ali Ibrahim Yang Song Ali Eichenbaum Thomas Brovelli ## Core Contributors Sahitya Potluri Preethi Lahoti Cip Baetu Ali Ghorbani Charles Chen Andy Crawford Shalini Pal Mukund Sridhar Petru Gurita Asier Mujika Igor Petrovski Pierre-Louis Cedoz Chenmei Li Shiyuan Chen Niccolò Dal Santo Siddharth Goyal Jitesh Punjabi Karthik Kappaganthu Chester Kwak Pallavi LV Sarmishta Velury Himadri Choudhury Jamie Hall Premal Shah Ricardo Figueira Matt Thomas Minjie Lu Ting Zhou Chintu Kumar Thomas Jurdi Sharat Chikkerur Yenai Ma Adams Yu Soo Kwak Victor Ähdel Sujeevan Rajayogam Travis Choma Fei Liu Aditya Barua Colin Ji Ji Ho Park Vincent Hellendoorn Alex Bailey Taylan Bilal Huanjie Zhou Mehrdad Khatir Charles Sutton Wojciech Rzadkowski ## Core Contributors Fiona Macintosh Roopali Vij Konstantin Shagin Paul Medina Chen Liang Jinjing Zhou Pararth Shah Yingying Bi Attila Dankovics Shipra Banga Sabine Lehmann Marissa Bredesen Zifan Lin John Eric Hoffmann Jonathan Lai Raynald Chung Kai Yang Nihal Balani Arthur Bražinskas Andrei Sozanschi Matthew Hayes Héctor Fernández Alcalde Peter Makarov Will Chen Antonio Stella Liselotte Snijders Michael Mandl Ante Kärrman Paweł Nowak Xinyi Wu Alex Dyck Krishnan Vaidyanathan Raghavender R Jessica Mallet Mitch Rudominer Eric Johnston Sushil Mittal Akhil Udathu Janara Christensen Vishal Verma Zach Irving Andreas Santucci ## Contributors Gamaleldin Elsayed Elnaz Davoodi Marin Georgiev Ian Tenney ## Contributors Nan Hua Geoffrey Cideron Edouard Leurent Mahmoud Alnahlawi Ionut Georgescu Nan Wei Ivy Zheng Dylan Scandinaro Heinrich Jiang Jasper Snoek Mukund Sundararajan Xuezhi Wang Zack Ontiveros Itay Karo Jeremy Cole Vinu Rajashekhar Lara Tumeh Eyal Ben-David Rishub Jain Jonathan Uesato Romina Datta Oskar Bunyan Shimu Wu John Zhang Piotr Stanczyk Ye Zhang David Steiner Subhajit Naskar Michael Azzam Matthew Johnson Adam Paszke Chung-Cheng Chiu Jaume Sanchez Elias Afroz Mohiuddin Faizan Muhammad Jin Miao Andrew Lee Nino Vieillard Jane Park Jiageng Zhang Jeff Stanway Drew Garmon Abhijit Karmarkar Zhe Dong Jong Lee Aviral Kumar Luowei Zhou Jonathan Evens ## Contributors William Isaac Geoffrey Irving Edward Loper Michael Fink Isha Arkatkar Nanxin Chen Izhak Shafran Ivan Petrychenko Zhe Chen Johnson Jia Anselm Levskaya Zhenkai Zhu Peter Grabowski Yu Mao Alberto Magni Kaisheng Yao Javier Snaider Norman Casagrande Evan Palmer Paul Suganthan Alfonso Castaño Irene Giannoumis Wooyeol Kim Mikołaj Rybiński Ashwin Sreevatsa Jennifer Prendki David Soergel Adrian Goedeckemeyer Willi Gierke Mohsen Jafari Meenu Gaba Jeremy Wiesner Diana Gage Wright Yawen Wei Harsha Vashisht Yana Kulizhskaya Jay Hoover Maigo Le Lu Li Chimezie Iwuanyanwu Lu Liu Kevin Ramirez Andrey Khorlin Albert Cui Tian LIN Marcus Wu Ricardo Aguilar Keith Pallo ## Contributors Abhishek Chakladar Ginger Perng Elena Allica Abellan Mingyang Zhang Ishita Dasgupta Nate Kushman Ivo Penchev Alena Repina Xihui Wu Tom van der Weide Priya Ponnapalli Caroline Kaplan Jiri Simsa Shuangfeng Li Olivier Dousse Fan Yang Jeff Piper Nathan Ie Rama Pasumarthi Nathan Lintz Anitha Vijayakumar Daniel Andor Pedro Valenzuela Minnie Lui Cosmin Paduraru Daiyi Peng Katherine Lee Shuyuan Zhang Somer Greene Duc Dung Nguyen Paula Kurylowicz Cassidy Hardin Lucas Dixon Lili Janzer Kiam Choo Ziqiang Feng Biao Zhang Achintya Singhal Dayou Du Dan McKinnon Natasha Antropova Tolga Bolukbasi Orgad Keller David Reid Daniel Finchelstein Maria Abi Raad Remi Crocker Peter Hawkins ## Contributors Robert Dadashi Colin Gaffney Ken Franko Anna Bulanova Rémi Leblond Shirley Chung Harry Askham Luis C. Cobo Kelvin Xu Felix Fischer Jun Xu Christina Sorokin Chris Alberti Chu-Cheng Lin Colin Evans Alek Dimitriev Hannah Forbes Dylan Banarse Zora Tung Mark Omernick Colton Bishop Rachel Sterneck Rohan Jain Jiawei Xia Ehsan Amid Francesco Piccinno Xingyu Wang Praseem Banzal Daniel J. Mankowitz Alex Polozov Victoria Krakovna Sasha Brown MohammadHossein Bateni Dennis Duan Vlad Firoiu Meghana Thotakuri Tom Natan Matthieu Geist Sertan Girgin Hui Li Jiayu Ye Ofir Roval Reiko Tojo Michael Kwong James Lee-Thorp Christopher Yew Danila Sinopalnikov Sabela Ramos ## Contributors John Mellor Abhishek Sharma Kathy Wu David Miller Nicolas Sonnerat Denis Vnukov Rory Greig Jennifer Beattie Emily Caveness Libin Bai Julian Eisenschlos Alex Korchemniy Tomy Tsai Mimi Jasarevic Weize Kong Phuong Dao Zeyu Zheng Frederick Liu Fan Yang Rui Zhu Tian Huey Teh Jason Sanmiya Evgeny Gladchenko Nejc Trdin Daniel Toyama Evan Rosen Sasan Tavakkol Linting Xue Chen Elkind Oliver Woodman John Carpenter George Papamakarios Rupert Kemp Sushant Kafle Tanya Grunina Rishika Sinha Alice Talbert Diane Wu Denese Owusu-Afriyie Cosmo Du Chloe Thornton Jordi Pont-Tuset Pradyumna Narayana Jing Li Saaber Fatehi John Wieting Omar Ajmeri Benigno Uria ## Contributors Yeongil Ko Laura Knight Amélie Héliou Ning Niu Shane Gu Chenxi Pang Yeqing Li Nir Levine Ariel Stolovich Rebeca Santamaria-Fernandez Sonam Goenka Wenny Yustalim Robin Strudel Ali Elqursh Charlie Deck Hyo Lee Zonglin Li Kyle Levin Raphael Hoffmann Dan Holtmann-Rice Olivier Bachem Sho Arora Christy Koh Soheil Hassas Yeganeh Siim Põder Mukarram Tariq Yanhua Sun Lucian Ionita Mojtaba Seyedhosseini Pouya Tafti Zhiyu Liu Anmol Gulati Jasmine Liu Xinyu Ye Bart Chrzaszcz Lily Wang Nikhil Sethi Tianrun Li Ben Brown Shreya Singh Wei Fan Aaron Parisi Joe Stanton Vinod Koverkathu Christopher A. Choquette-Choo Yunjie Li TJ Lu Abe Ittycheriah ## Contributors Prakash Shroff Mani Varadarajan Sanaz Bahargam Rob Willoughby David Gaddy Guillaume Desjardins Marco Cornero Brona Robenek Bhavishya Mittal Ben Albrecht Ashish Shenoy Fedor Moiseev Henrik Jacobsson Alireza Ghaffarkhah Morgane Rivière Alanna Walton Clément Crepy Alicia Parrish Zongwei Zhou Clement Farabet Carey Radebaugh Praveen Srinivasan Claudia van der Salm Andreas Fidjeland Salvatore Scellato Eri Latorre-Chimoto Hanna Klimczak-Plucińska David Bridson Dario de Cesare Tom Hudson Piermaria Mendolicchio Lexi Walker Alex Morris Matthew Mauger Alexey Guseynov Alison Reid Seth Odoom Lucia Loher Victor Cotruta Madhavi Yenugula Dominik Grewe Anastasia Petrushkina Tom Duerig Antonio Sanchez Steve Yadlowsky Amy Shen Amir Globerson Lynette Webb ## Contributors Sahil Dua Dong Li Surya Bhupatiraju Dan Hurt Haroon Qureshi Ananth Agarwal Tomer Shani Matan Eyal Anuj Khare Shreyas Rammohan Belle Lei Wang Chetan Tekur Mihir Sanjay Kale Jinliang Wei Ruoxin Sang Brennan Saeta Tyler Liechty Yi Sun Yao Zhao Stephan Lee Pandu Nayak Doug Fritz Manish Reddy Vuyyuru John Aslanides Nidhi Vyas Martin Wicke Xiao Ma Evgenii Eltyshev Nina Martin Hardie Cate James Manyika Keyvan Amiri Yelin Kim Xi Xiong Kai Kang Florian Luisier Nilesh Tripuraneni David Madras Mandy Guo Austin Waters Oliver Wang Joshua Ainslie Jason Baldridge Han Zhang Garima Pruthi Jakob Bauer Feng Yang Riham Mansour ## Contributors Jason Gelman Yang Xu George Polovets Ji Liu Honglong Cai Warren Chen XiangHai Sheng Emily Xue Sherjil Ozair Christof Angermueller Xiaowei Li Anoop Sinha Weiren Wang Julia Wiesinger Emmanouil Koukoumidis Yuan Tian Anand Iyer Madhu Gurumurthy Mark Goldenson Parashar Shah MK Blake Hongkun Yu Anthony Urbanowicz Jennimaria Palomaki Chrisantha Fernando Ken Durden Harsh Mehta Nikola Momchev Elahe Rahimtoroghi Maria Georgaki Amit Raul Sebastian Ruder Morgan Redshaw Jinhyuk Lee Denny Zhou Komal Jalan Dinghua Li Blake Hechtman Parker Schuh Milad Nasr Kieran Milan Vladimir Mikulik Juliana Franco Tim Green Nam Nguyen Joe Kelley Aroma Mahendru Andrea Hu ## Contributors Joshua Howland Ben Vargas Jeffrey Hui Kshitij Bansal Vikram Rao Rakesh Ghiya Emma Wang Ke Ye Jean Michel Sarr Melanie Moranski Preston Madeleine Elish Steve Li Aakash Kaku Jigar Gupta Ice Pasupat Da-Cheng Juan Milan Someswar Tejvi M. Xinyun Chen Aida Amini Alex Fabrikant Eric Chu Xuanyi Dong Amruta Muthal Senaka Buthpitiya Sarthak Jauhari Nan Hua Urvashi Khandelwal Ayal Hitron Jie Ren Larissa Rinaldi Shahar Drath Avigail Dabush Nan-Jiang Jiang Harshal Godhia Uli Sachs Anthony Chen Yicheng Fan Hagai Taitelbaum Hila Noga Zhuyun Dai James Wang Chen Liang Jenny Hamer Chun-Sung Ferng Chenel Elkind Aviel Atias Paulina Lee ## Contributors Vít Listík Mathias Carlen Jan van de Kerkhof Marcin Pikus Krunoslav Zaher Paul Müller Sasha Zykova Richard Stefanec Vitaly Gatsko Christoph Hirnschall Ashwin Sethi Xingyu Federico Xu Chetan Ahuja Beth Tsai Anca Stefanoiu Bo Feng Keshav Dhandhania Manish Katyal Akshay Gupta Atharva Parulekar Divya Pitta Jing Zhao Vivaan Bhatia Yashodha Bhavnani Omar Alhadlaq Xiaolin Li Peter Danenberg Dennis Tu Alex Pine Vera Filippova Abhipso Ghosh Ben Limonchik Bhargava Urala Chaitanya Krishna Lanka Derik Clive Yi Sun Edward Li Hao Wu Kevin Hongtongsak Ianna Li Kalind Thakkar Kuanysh Omarov Kushal Majmundar Michael Alverson Michael Kucharski Mohak Patel Mudit Jain Maksim Zabelin | Contributors Paolo Pelagatti Rohan Kohli Saurabh Kumar Joseph Kim Swetha Sankar Vineet Shah | Gemini App Program Leads | |-----------------------------------------------------------------------------------------------|--------------------------------------------------------------------| | | Amar Subramanya 7 | | | Sissie Hsiao | | | Gemini Program Leads | | | Demis Hassabis | | | Koray Kavukcuoglu | | Lakshmi Ramachandruni | | | Xiangkai Zeng | Overall Gemini App Technical Leads | | Ben Bariach | Adam Sadovsky 8 | | Laura Weidinger | Quoc Le | | Tu Vu | Trevor Strohman 9 | | Alek Andreev | Yonghui Wu 10 | | Antoine He | | | Kevin Hui | Overall Gemini Post-Training Lead | | Sheleem Kashem | Slav Petrov | | | Overall Gemini Technical Leads (equal con- tribution) Jeffrey Dean | | | Oriol Vinyals | The roles are defined as below: - Lead : Individual(s) responsible for the sub-team throughout the project. - Core Contributor : Individual that had significant impact throughout the project. - Contributor : Individual that had contributions to the project and was partially involved with the effort. - Program Lead : Responsible for the organizational aspects of the Gemini effort. - Overall Post-Training Lead : Responsible for the technical direction of post-training. - Overall Technical Lead : Responsible for the technical direction of the overall Gemini effort. Within each role, contributions are equal, and are listed in a randomized order. Ordering within each role does not indicate ordering of the contributions. Gemini is a cross-Google effort, with members from Google DeepMind (GDM), Google Research (GR), Bard/Assistant, Knowledge and Information (K&I), Core ML, Cloud, Labs, and more. We thank Aakanksha Chowdhery, Dustin Tran, Heng-Tze Cheng, Jack W. Rae, Kate Olszewska, Mariko Iinuma, Peter Humphreys, Shashi Narayan, and Steven Zheng for leading the preparation of this report. We also thank our reviewers and colleagues for their valuable discussions and feedback on the report - Alexandra Belias, Ana Ramalho, Anand Rao, Arielle Bier, Danielle Landress, Eleanor Tomlinson, Emily Hossellman, Gaby Pearl, Helen King, Hollie Dobson, Jaclyn Konzelmann, Jennifer 7 Lead, Gemini App Engineering 8 Lead, Gemini App Core Modeling, Eval, Data 9 Co-Lead, Gemini App Serving 10 Co-Lead, Gemini Text Beroshi, Joel Moss, Jon Small, Jonathan Fildes, Kathy Meier-Hellstern, Lisa Patel, Oli Gaymond, Rebecca Bland, Reena Jana, Tessa Lueth, and Tom Lue. Our work is made possible by the dedication and efforts of numerous teams at Google. We would like to acknowledge the support from Abhi Mohan, Adekunle Bello, Aishwarya Nagarajan, Alaa Saade, Alejandro Lince, Alexander Chen, Alexander Kolbasov, Alexander Schiffhauer, Ameya Shringi, Amin Vahdat, Anda Rabatić, Anthonie Gross, Antoine Yang, Anthony Green, Anton Ruddock, Art Khurshudov, Artemis Chen, Arthur Argenson, Avinatan Hassidim, Beiye Liu, Benjamin Schroeder, Bin Ni, Brett Daw, Bryan Chiang, Burak Gokturk, Carl Crous, Carrie Grimes Bostock, Charbel Kaed, Charlotte Banks, Che Diaz, Chris Larkin, Christy Lian, Claire Cui, Clare Bycroft, Corentin Tallec, Daniel Herndon, Dave Burke, David Battle, David Engel, Dipannita Shaw, Donghyun Koo, Doug Ritchie, Dragos Stefanescu, Elissa Wolf, Emre Sargin, Eric Herren, Estella King, Fatema Alkhanaizi, Felix Gimeno, Fernando Pereira, Florent Altché, Gabriel Carvajal, Gaurav Gandhi, George Powell, Goran Pavičić, Harry Richardson, Hassan Wassel, Hongji Li, Idan Szpektor, Igor Ivanisevic, Ivan Jambrešić, Ivan Jurin, Jade Fowler, James Assiene, Jay Yagnik, Jean-bastien Grill, Jeff Seibert, Jenna LaPlante, Jessica Austin, Jianxing Lu, Jim O'Keeffe, Jin Huang, Joe Heyward, Johannes Welbl, John Jumper, Jonathan Caton, Josh Woodward, Joshua Foster, Kathryn Tunyasuvunakool, Katrina Wong, Kavya Kopparapu, Kelvin Nguyen, Kira Yin, Konstantin Sharlaimov, Kun Li, Lee Hong, Lilly Taylor, Longfei Shen, Luc Mercier, Maciej Mikuła, Mania Abdi, Manuel Sanchez, Maria Ines Aranguren, Mario Carlos Cortes III, Matthew Tait, Matthias Lochbrunner, Mehdi Ghissassi, Micah Mosley, Michael Bendersky, Michael Figurnov, Michael Harris, Michael Mathieu, Michael O'Neill, Michael Vorburger, Mihir Paradkar, Nandita Dukkipati, Nathan Carter, Nathan Watson, Neil Rabinowitz, Nikhil Dandekar, Nishant Ranka, Olcan Sercinoglu, Olivier Lacombe, Ottavia Bertolli, Paul Caron, Pranesh Srinivasan, Praveen Kumar, Rahul Sukthankar, Raia Hadsell, Rajagopal Ananthanarayanan, Roberto Lupi, Rosie Zou, Sachin Menezes, Sadegh Jazayeri, Sam Cheung, Sameer Bidichandani, Sania Alex, Sanjiv Kumar, Sara Wiltberger, Sarah Fitzgerald, Saz Basu, Sebastian Nowozin, Shannon Hepburn, Shayne Cardwell,Srinivasan Venkatachary, Sugato Basu, Sundar Pichai, Sundeep Tirumalareddy, Susannah Young, Swetha Vijayaraghavan, Tania Bedrax-Weiss, Taylor Applebaum, Teiva Harsanyi, Terry Chen, Tim Blyth, Ting Liu, Tom Cobley, Tomas Izo, Trystan Upstill, Varun Singhai, Vedrana Klarić Trupčević, Victor Cai, Vladimir Pudovkin, Vu Dang, Wenbo Zhao, Wesley Crow, Wesley Szeng, Xiaodan Song, Yazhou Zu, Ye Tian, Yicong Wang, Yixing Wang, Yossi Matias, Yunlong Jiao, Zachary Jessup, Zhenchuan Pang, Žiga Avsec, Zimeng Yang, and Zoubin Ghahramani. We'd also like to recognize the AlphaCode team, the Borg Scheduling team, the Facilities team, the Gemini Demo Team, the Global Server Ops (GSO) team, the JAX team, the the Legal team, ML SRE team, the ML Supercomputer (MLSC) team, the PartIR team, the Platforms Infrastructure Engineering (PIE) team, and the XLA Compiler team. We thank everyone at Google not explicitly mentioned above, who have shared excitement, given feedback on early Gemini models or created interesting demo uses of Gemini, and worked with or supported the core Gemini team on many aspects of this project. ## 10. Appendix ## 10.1. Gemini Ultra Model Card | Model summary | Model summary | |--------------------|--------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------| | Model architecture | Gemini V1.0 is a new family of state-of-the-art language models, containing variants known as Nano, Pro and Ultra (ordered by parameter count) based on a decoder-only Transformer architecture (Vaswani et al., 2017a). Models are trained to support 32K context length, employing efficient attention mechanisms such as multi-query attention (Shazeer, 2019b). Gemini is trained jointly across image, audio, video and text data for the purpose of building a model with both strong generalist capabilities across modalities alongside cutting-edge understanding and reasoning performance in each respective domain. The post-trained models described in this model card are Gemini API and Gemini Apps model variants (Section 6) built on top of the Gemini Ultra pre-trained model. During the post-training process, additional architectural modifications are also made to support the training of multi-objective reward models for RLHF. | | Input(s) | Text (e.g. a question, a prompt, a document(s) to be summa- rized), images, video, audio files. | | Output(s) | Generated text in response to the input (e.g. an answer to the question, a summary of multiple documents, comparing documents/videos). | | Usage | Usage | | Application | Gemini is designed for accelerating research on language models, for use as a building block in features within Google products, and as a building block for select applications such as Gemini App and Search Generative Experience. Services and products built on top of Gemini Ultra are also being made available to external developers via Google Cloud Vertex API and Google Labs, with additional process and technical safeguards related to safety policies. | | Known Caveats | Gemini should not be made available as part of a general-purpose service or product, or used within a specific downstream appli- cation without a prior assessment and mitigation of the safety and fairness concerns specific to the downstream use. | | | Implementation Frameworks | |-----------------------|------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------| | Hardware & Software | Hardware: Training was conducted on TPUv4 and TPUv5e (Jouppi et al., 2020, 2023). Software: JAX (Bradbury et al., 2018), ML Pathways (Dean, 2021). JAX allows researchers to leverage the latest generation of hardware, including TPUs, for faster and more efficient training of large models. ML Pathways is infrastructure software to support Google's efforts to build artificially intelligent systems capable of generalizing across multiple tasks. This is specially suitable for foundation models, including large language models like the Gemini V1.0 models. Together, JAX and ML Pathways are used as described in Section 3. The 'single controller' programming model of JAX and ML Pathways allows a single Python process to orchestrate the entire training run, dramatically simplifying the development workflow. | | Compute Requirements | Not reported. | | Model Characteristics | Model Characteristics | | Model initialization | Initial pretraining used random initialization. Post-training was initialized from checkpoints obtained at the later stages of pre- training. These checkpoints were fine-tuned using supervised fine-tuning, and subsequently used to initialize reward model training and RLHF. | | Model Status | This is a static model trained on an offline dataset. | | Model Stats | Not reported. | | Data overview | Data overview | | Training Dataset | Gemini models are trained on a dataset that is both multimodal and multilingual. Our pre-training dataset uses data from web documents, books, and code, and includes image, audio, and video data. Refer to Section 4 (Pre-Training Dataset) for further de- tails. | | Evaluation Dataset | We compare pre- and post-trained Gemini Ultra models to a suite of external LLMs and our previous best model PaLM 2 across a series of text-based academic benchmarks covering reasoning, reading comprehension, STEM, and coding. We also evaluate Gemini models on four different mul- timodal capabilities: high-level object recognition using captioning or question-answering tasks such as VQAv2; fine- grained transcription using tasks such as TextVQA and DocVQA requiring the model to recognize low-level details; chart understanding requiring spatial understanding of input layout using ChartQA and InfographicVQA tasks; and multimodal reasoning using tasks such as Ai2D, MathVista and MMMU. | |--------------------------------|----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------| | Post-training Dataset | For post-training, we first collect a diverse set of prompts that are representative of real-world use cases. We then collect demonstration data of what the model's output should be for a given prompt for supervised fine-tuning. We further collect different possible responses to a given prompt, and collect feedback data over these to train reward models. Refer to Section 6.3 (Post-Training Methods and Data) for further details. | | Evaluation Results | Evaluation Results | | Benchmark Information | See Section 5 (Evaluation). | | Evaluation Results | See Section 5 (Evaluation) and Section 6.4 (Post-Training Hu- man Evaluation). | | Model Usage & Limitations | Model Usage & Limitations | | Sensitive Use | For an analysis of risks and sensitive uses associated with the Gemini models, see Section 7.1 (Impact Assessment). | | Known Limitations | Gemini models can exhibit limitations outlined in Section 7.1 (Impact Assessment). Gemini models should not be used for downstream applications without further analysis of potential harm in the proposed downstream application. | | Ethical Considerations & Risks | Areflection on the potential risks and impacts of the Gemini V1.0 models can be found in Section 7 (Responsible Deployment). For evaluation details for a range of risks, see Section 7.4 (Safety Evaluations). | ## 10.2. Chain-of-Thought Comparisons on MMLU benchmark We contrast several chain-of-thought approaches on MMLU and discuss their results in this section. We proposed a new approach where model produces k chain-of-thought samples, selects the majority vote if the model is confident above a threshold, and otherwise defers to the greedy sample choice. The thresholds are optimized for each model based on their validation split performance. The proposed approach is referred to as uncertainty-routed chain-of-thought . The intuition behind this approach is that chain-of-thought samples might degrade performance compared to the maximum-likelihood decision when the model is demonstrably inconsistent. We compare the gains from the proposed approach on both Gemini Ultra and GPT-4 in Figure 9. We find that Gemini Ultra benefits more from this approach compared to using only chain-of-thought samples. GPT-4's performance improves from 84.2% with greedy sampling to 87.3% with uncertainty-routed chain-of-thought approach with 32 samples, but it already achieves these gains from using 32 chain-of-thought samples. In contrast, Gemini Ultra improves its performance significantly from 84.0% with greedy sampling to 90.0% with uncertainty-routed chain-of-thought approach with 32 samples while it marginally improves to 85.0% with the use of 32 chain-of-thought samples only. Figure 9 | Chain-of-Thought with uncertainty routing on MMLU. <details> <summary>Image 16 Details</summary>  ### Visual Description \n ## Bar Chart: MMLU Accuracy Comparison - GPT-4 vs. Gemini Ultra ### Overview This bar chart compares the MMLU (Massive Multitask Language Understanding) accuracy, measured on a test split, between two language models: GPT-4 (gpt-4-0613) and Gemini Ultra. The comparison is made across three different evaluation methods: Score Eval, Chain-of-Thought@32, and Chain-of-Thought@32 (Uncertainty-Routed). ### Components/Axes * **X-axis:** Evaluation Method. Categories are: "Score Eval", "Chain-of-Thought@32", "Chain-of-Thought@32 (Uncertainty-Routed)". * **Y-axis:** MMLU accuracy (test split), ranging from 0 to 90. * **Legend:** * Grey: GPT-4 (gpt-4-0613) * Blue: Gemini Ultra ### Detailed Analysis The chart consists of six bars, grouped in sets of two for each evaluation method. Each set represents the accuracy of GPT-4 and Gemini Ultra. * **Score Eval:** * GPT-4: Accuracy is approximately 84.21. * Gemini Ultra: Accuracy is approximately 83.96. * **Chain-of-Thought@32:** * GPT-4: Accuracy is approximately 87.29. * Gemini Ultra: Accuracy is approximately 84.99. * **Chain-of-Thought@32 (Uncertainty-Routed):** * GPT-4: Accuracy is approximately 87.29. * Gemini Ultra: Accuracy is approximately 90.04. The bars are positioned side-by-side, allowing for direct visual comparison of the two models' performance for each evaluation method. ### Key Observations * Gemini Ultra consistently outperforms GPT-4 in the "Chain-of-Thought@32 (Uncertainty-Routed)" evaluation method, with a significant difference in accuracy (approximately 90.04 vs. 87.29). * In the "Score Eval" evaluation method, the performance of both models is very close, with GPT-4 slightly outperforming Gemini Ultra. * In the "Chain-of-Thought@32" evaluation method, GPT-4 outperforms Gemini Ultra. ### Interpretation The data suggests that Gemini Ultra benefits significantly from the "Uncertainty-Routed" approach within the Chain-of-Thought framework, achieving a notably higher accuracy compared to GPT-4. This indicates that incorporating uncertainty estimation into the reasoning process enhances Gemini Ultra's performance on the MMLU benchmark. The close performance in "Score Eval" suggests that both models have a similar baseline understanding of the tasks. The difference in performance in the Chain-of-Thought methods suggests that the models differ in their ability to leverage reasoning and potentially benefit from techniques like uncertainty routing. The MMLU benchmark tests a model's multi-task accuracy, and these results suggest that Gemini Ultra is more robust to the addition of uncertainty routing. </details> ## 10.3. Capabilities and Benchmarking Tasks We use more than 50 benchmarks as a holistic harness to evaluate the Gemini models across text, image, audio and video. We provide a detailed list of benchmarking tasks for six different capabilities in text understanding and generation: factuality, long context, math/science, reasoning, summarization, and multilinguality. We also enumerate the benchmarks used for image understanding, video understanding, and audio understanding tasks. - Factuality : We use 5 benchmarks: BoolQ (Clark et al., 2019), NaturalQuestions-Closed (Kwiatkowski et al., 2019a), NaturalQuestions-Retrieved (Kwiatkowski et al., 2019a), RealtimeQA (Kasai et al., 2022b), TydiQA-noContext and TydiQA-goldP (Clark et al., 2020). - Long Context : We use 6 benchmarks: NarrativeQA (Kočiský et al., 2018), Scrolls-Qasper, Scrolls-Quality (Shaham et al., 2022), XLsum (En), XLSum (non-English languages) (Hasan et al., 2021), and one other internal benchmark. - Math/Science : We use 8 benchmarks: GSM8k (with CoT) (Cobbe et al., 2021), Hendryck's MATHpass@1(Hendrycks et al., 2021b), MMLU (Hendrycks et al., 2021a), Math-StackExchange, Math-AMC 2022-2023 problems, and three other internal benchmarks. - Reasoning : We use 7 benchmarks: BigBench Hard (with CoT) (Srivastava et al., 2022; Suzgun et al., 2022), CLRS (Veličković et al., 2022), Proof Writer (Tafjord et al., 2020), Reasoning-Fermi problems (Kalyan et al., 2021), Lambada (Paperno et al., 2016), HellaSwag (Zellers et al., 2019), DROP (Dua et al., 2019). - Summarization : We use 5 benchmarks: XL Sum (English), XL Sum (non-English languages) (Hasan et al., 2021), WikiLingua (non-English languages), WikiLingua (English) (Ladhak et al., 2020), XSum (Narayan et al., 2018). - Multilinguality : We use 10 benchmarks: XLSum (Non-English languages) (Hasan et al., 2021), WMT22 (Kocmi et al., 2022), WMT23 (Tom et al., 2023), FRMT (Riley et al., 2023), WikiLingua (Non-English languages) (Ladhak et al., 2020), TydiQA (no context), TydiQA (GoldP) (Clark et al., 2020), MGSM (Shi et al., 2023), translated MMLU (Hendrycks et al., 2021a), NTREX (Federmann et al., 2022), FLORES-200 (Team et al., 2022). - Image and Video : We use 9 benchmarks for image understanding: MMMU (Yue et al., 2023), TextVQA (Singh et al., 2019), DocVQA (Mathew et al., 2021), ChartQA (Masry et al., 2022), InfographicVQA (Mathew et al., 2022), MathVista (Lu et al., 2023), AI2D (Kembhavi et al., 2016), VQAv2 (Goyal et al., 2017), XM3600 (Thapliyal et al., 2022) for multi-lingual image understanding, and 6 benchmarks for video understanding: VATEX (Wang et al., 2019) for captioning in two different languages, YouCook2 (Zhou et al., 2018), NextQA (Xiao et al., 2021), ActivityNet-QA (Yu et al., 2019), and Perception Test MCQA (Pătrăucean et al., 2023). - Audio : We use 5 benchmarks including automatic speech recognition (ASR) tasks such as FLEURS (Conneau et al., 2023), VoxPopuli (Wang et al., 2021), Multi-lingual Librispeech (Pratap et al., 2020), and automatic speech translation task such as CoVoST 2 (Wang et al., 2020). ## 10.4. Qualitative Examples This section shows sample qualitative examples from prompting the Gemini Ultra model. Some illustrative examples of multimodal reasoning for image understanding tasks over charts, natural images and memes are shown in Figures 10, 11, 13, 15, 16, and 17. Figure 12 shows an example of image generation capabilities of Gemini Ultra where the user generates an interleaved sequence of image and text to design a blog post. Beyond English, Figure 18 shows model's capability to understand images in a multilingual setting. Gemini models also show strong performance on multimodal image understanding and reasoning in mathematics, as shown in Figures 14, 20 and 21. Figure 22 is an example of complex multimodal reasoning demonstrating how the model composes complex image understanding, code generation, and instruction following capabilities for a given user task. In Figure 19, we see another example of the model being able to generate working code and follow complex user instructions. Finally, Figure 23 shows an example of Gemini Ultra's capability of understanding video by reasoning over temporally connected set of frames. ## 10.4.1. Chart understanding and reasoning over data <details> <summary>Image 17 Details</summary>  ### Visual Description \n ## Bar Charts: Share of Plastic Waste Management - 2019 ### Overview The image presents four horizontal bar charts illustrating the share of plastic waste that is recycled, landfilled, incinerated, and mismanaged in different regions: World, United States, Europe, and Asia (excl. China and India). The data is sourced from the OECD (2023) and visualized by Our World in Data. ### Components/Axes * **Title:** "Share of plastic waste that is recycled, landfilled, incinerated and mismanaged. 2019" * **Source:** "Data source: OECD (2023)" and "OurWorldinData.org/plastic-pollution | CC BY" * **Note:** "Regional aggregates were calculated by Our World in Data and are based on those specified by the OECD." * **Regions:** World, United States, Europe, Asia (excl. China and India) * **Categories:** Landfilled, Mismanaged, Incinerated, Recycled * **Axis:** The Y-axis represents the percentage of plastic waste (0% to 100%). The X-axis represents the categories of waste management. * **Colors:** * Landfilled: Light Blue * Mismanaged: Orange * Incinerated: Dark Green * Recycled: Light Green ### Detailed Analysis or Content Details **1. World:** * Landfilled: Approximately 49% (visually, the bar extends to just under the 50% mark). * Mismanaged: Approximately 22% (visually, the bar extends to just over the 20% mark). * Incinerated: Approximately 19% (visually, the bar extends to just under the 20% mark). * Recycled: Approximately 9% (visually, the bar extends to just over the 5% mark). **2. United States:** * Landfilled: Approximately 73% (visually, the bar extends to just over the 70% mark). * Mismanaged: Approximately 4% (visually, the bar extends to just over the 0% mark). * Incinerated: Approximately 19% (visually, the bar extends to just under the 20% mark). * Recycled: Approximately 4% (visually, the bar extends to just over the 0% mark). **3. Europe:** * Landfilled: Approximately 44% (visually, the bar extends to just under the 45% mark). * Mismanaged: Approximately 6% (visually, the bar extends to just over the 5% mark). * Incinerated: Approximately 38% (visually, the bar extends to just under the 40% mark). * Recycled: Approximately 12% (visually, the bar extends to just over the 10% mark). **4. Asia (excl. China and India):** * Landfilled: Approximately 39% (visually, the bar extends to just under the 40% mark). * Mismanaged: Approximately 34% (visually, the bar extends to just over the 30% mark). * Incinerated: Approximately 19% (visually, the bar extends to just under the 20% mark). * Recycled: Approximately 8% (visually, the bar extends to just over the 5% mark). ### Key Observations * The United States has the highest percentage of plastic waste landfilled (73%) and the lowest percentage recycled (4%). * The World average shows a significant portion of plastic waste is landfilled (49%) and mismanaged (22%). * Europe has a higher recycling rate (12%) compared to the World, United States, and Asia (excl. China and India). * Asia (excl. China and India) has a high percentage of mismanaged plastic waste (34%). ### Interpretation The data highlights significant disparities in plastic waste management practices across different regions. The United States relies heavily on landfilling, indicating a lack of investment in recycling and other waste treatment technologies. The high rate of mismanaged plastic waste in Asia (excl. China and India) suggests inadequate waste collection and disposal infrastructure. Europe demonstrates a more balanced approach, with a relatively higher recycling rate and a lower percentage of mismanaged waste. The overall trend suggests that a substantial portion of plastic waste globally is either landfilled or mismanaged, posing significant environmental risks. The data underscores the need for improved waste management infrastructure, increased recycling rates, and a reduction in plastic consumption to address the global plastic pollution crisis. The note regarding the OECD definition of regions is important. The data is based on how the OECD categorizes countries, which may influence the results. The fact that China and India are excluded from the "Asia" category is also noteworthy, as these countries are major contributors to plastic waste generation. A data point that stands out is the extremely low recycling rate in the United States (4%) compared to the other regions. This implies a systemic issue with recycling infrastructure or participation in the US, and has significant implications for environmental sustainability. </details> ## Model Response (rendered Markdown) The data point that stands out is the United States' landfilled plastic waste, which is 73%. This is significantly higher than the other countries and regions shown. This implies that the United States is not doing as much to recycle or incinerate its plastic waste as other countries and regions. | Country/Region | Landfilled (%) | Mismanaged (%) | Incinerated (%) | Recycled (%) | |------------------------------|------------------|------------------|-------------------|----------------| | World | 49 | 22 | 19 | 9 | | United States | 73 | 4 | 19 | 4 | | Europe | 44 | 6 | 38 | 12 | | Asia (excl. China and India) | 39 | 34 | 19 | 8 | Figure 10 | Solving a problem requiring multimodal chart understanding.The model has to read the text, understand the connections between different data points and reason over them to recommend an interesting point and follow the instructions to generate a markdown table (shown correctly rendered). Source: Our World In Data (Ritchie et al., 2023). ## 10.4.2. Multimodal question answering ## Prompt <details> <summary>Image 18 Details</summary>  ### Visual Description \n ## Photograph: Garden Plant Composition ### Overview The image is a close-up photograph of a garden bed containing a variety of plants with differing foliage colors and textures. The dominant plant features striking purple leaves with prominent veining. There are no charts, diagrams, or data tables present. The image is purely descriptive. ### Components/Axes There are no axes, scales, or legends present in this image. The components are solely the various plants and their arrangement within the garden bed. ### Detailed Analysis or Content Details The photograph showcases a dense arrangement of plants. * **Dominant Plant:** The central plant is characterized by its deep purple leaves, exhibiting a distinct ribbed or veined texture. The leaves are roughly triangular in shape with pointed tips. The plant appears to be relatively healthy and robust. * **Pink Flowers:** Several clusters of bright pink flowers are visible in the lower-left and left portions of the image. These flowers are small and densely packed. * **Green Foliage:** A variety of green foliage provides a backdrop to the purple plant. The green leaves vary in shape and size, suggesting different plant species. Some leaves are broad and rounded, while others are smaller and more delicate. * **Spotted Foliage:** In the bottom-right corner, there is a plant with leaves that have a speckled or mottled pattern, appearing to be a mix of green and pink. * **Light Pink Flowers:** Clusters of light pink flowers are visible on the left side of the image. There is no quantifiable data present in the image. The colors are approximate: * Purple: RGB(128, 0, 128) - approximate * Pink: RGB(255, 192, 203) - approximate * Green: RGB(0, 128, 0) - approximate ### Key Observations The composition emphasizes the contrast between the purple foliage and the surrounding green and pink plants. The arrangement appears natural and somewhat wild, suggesting a less formal garden style. The lighting is soft and diffused, highlighting the textures of the leaves and flowers. ### Interpretation The image likely aims to showcase the aesthetic appeal of diverse plant combinations in a garden setting. The striking purple color of the central plant serves as a focal point, drawing the viewer's eye. The variety of textures and colors creates visual interest and depth. The photograph could be used for horticultural purposes, garden design inspiration, or simply as a visually pleasing image. The image does not contain any data or information beyond the visual representation of the plants themselves. It is a descriptive image, not an analytical one. </details> Do you know what it this plant? How do I best take care of it? ## Model Response This is a Persian shield plant. It is a tropical plant that is native to Southeast Asia. It is a popular houseplant because it is easy to care for and has beautiful purple leaves. Persian shield plants prefer bright, indirect light. They can tolerate some direct sunlight, but too much direct sunlight can scorch their leaves. They also prefer warm temperatures and high humidity. Persian shield plants need to be watered regularly, but they should not be allowed to sit in water. The soil should be kept moist but not soggy. They also benefit from being misted regularly. Persian shield plants can be fertilized every two weeks during the growing season. A balanced fertilizer is a good choice. Persian shield plants are susceptible to pests, such as mealybugs and spider mites. These pests can be controlled with insecticidal soap or neem oil. Persian shield plants can be propagated by stem cuttings. The cuttings should be taken in the spring or summer and should be rooted in water or a moist potting mix. Figure 11 | Answering a multimodal information-seeking query. The model is able to recognize the specific plant shown in the image and provide information about it. The model shows robustness to typos as it is able to understand the user question despite them. Source: photo taken by an author from the Gemini team. ## 10.4.3. Interleaved image and text generation ## Prompt Please create a blog post about a trip to New York, where a dog and his owner had lots of fun. Include a few pictures of the dog posing happily at different landmarks. ## Model Response Hi, I'm Cooper. I'm a 2 year old Goldendoodle who loves to travel. Last month, my family and I went to New York City. It was my first time in the Big Apple, and I had a blast! We stayed in a hotel near Times Square. It was so much fun to see all the bright lights and people. We went to see a Broadway show, and I even got to meet some of the actors. One of my favorite parts of New York City was Central Park. It was so big and beautiful. I loved running around and playing fetch. We also went to the Statue of Liberty and the Empire State Building. I was so excited to see all the famous landmarks. I had so much fun in New York City. It's definitely one of my favorite places to travel. I can't wait to go back someday! Figure 12 | Generating interleaved text and images. The model is able to follow the instructions of generating a blog post with images closely related to the text and with dog images showing high levels of consistency across all images. <details> <summary>Image 19 Details</summary>  ### Visual Description \n ## Photographs: Dog in New York City - Triptych ### Overview The image is a triptych (a set of three images side-by-side) depicting a medium-sized, golden-brown, curly-haired dog (likely a Goldendoodle or similar breed) in various locations within New York City. The images appear to be professionally taken, with a shallow depth of field focusing on the dog. There is no quantitative data present in the image. ### Components/Axes There are no axes, legends, or scales present. The image consists solely of visual information. The three panels are arranged horizontally. ### Detailed Analysis or Content Details * **Panel 1 (Left):** The dog is positioned in the foreground, looking directly at the camera with its mouth open and tongue visible. It is standing on a wet street, and a yellow taxi cab is blurred in the background. Tall buildings are visible on either side of the street, suggesting a typical urban environment. * **Panel 2 (Center):** The dog is sitting on a pathway covered in fallen autumn leaves. The background features trees with golden foliage and blurred figures of people walking in the distance. The scene evokes a park or tree-lined street. * **Panel 3 (Right):** The dog is sitting with its back to the camera, looking out towards a body of water (likely a river or harbor). The skyline of New York City is visible in the background. The dog is positioned on a concrete ledge. ### Key Observations The consistent subject (the dog) across all three panels suggests a narrative or theme. The changing backgrounds showcase different aspects of New York City – urban streets, parks, and waterfront views. The dog appears calm and well-behaved in all settings. ### Interpretation The triptych likely aims to portray a dog experiencing the city life of New York. The images evoke a sense of adventure, exploration, and the integration of a pet into an urban environment. The use of different settings highlights the diversity of experiences available within the city. The images are likely intended to be aesthetically pleasing and emotionally engaging, rather than to convey specific data or information. The composition and lighting suggest a focus on capturing a positive and heartwarming narrative. There is no factual data to analyze, only visual storytelling. </details> ## 10.4.4. Image understanding and reasoning ## Prompt <details> <summary>Image 20 Details</summary>  ### Visual Description \n ## Diagram: Shape Sequence ### Overview The image presents a sequence of geometric shapes: a triangle, a square, a pentagon, and a question mark. The shapes are arranged horizontally in a row. The question mark suggests a continuation of the sequence, implying a pattern recognition task. ### Components/Axes There are no axes or formal components beyond the shapes themselves. The sequence is presented linearly. The shapes are: * Triangle (3 sides) * Square (4 sides) * Pentagon (5 sides) * Question Mark (?) ### Detailed Analysis or Content Details The sequence consists of polygons with an increasing number of sides. * Shape 1: Triangle - 3 sides * Shape 2: Square - 4 sides * Shape 3: Pentagon - 5 sides * Shape 4: Question Mark - Unknown number of sides. ### Key Observations The number of sides in each polygon increases by one. The question mark indicates the next shape in the sequence should be a polygon with 6 sides. ### Interpretation The diagram demonstrates a simple mathematical sequence based on the number of sides in a polygon. The question mark serves as a prompt to identify the next logical element in the sequence, which would be a hexagon (6 sides). This is a basic pattern recognition exercise. The diagram is not presenting data in a quantitative sense, but rather a qualitative progression of geometric forms. The image is designed to test logical reasoning and the ability to identify patterns. </details> Look at this sequence of three shapes. What shape should come as the fourth shape? Explain your reasoning with detailed descriptions of the first shapes. ## Model Response The fourth shape should be a hexagon. The first shape is a triangle, the second shape is a square, and the third shape is a pentagon. The number of sides in each shape is increasing by one Figure 13 | Solving a multimodal reasoning problem.The model is able to recognize shapes in the image, understand their properties and reason about the relationship between them to predict the next object. Source: photo taken by an author from the Gemini team. ## 10.4.5. Geometrical reasoning <details> <summary>Image 21 Details</summary>  ### Visual Description \n ## Diagram: Parallelogram Area Problem ### Overview The image presents a geometry problem involving a parallelogram. It includes a diagram of a parallelogram with labeled sides and height, along with a textual problem statement and a model response providing the solution. ### Components/Axes * **Diagram:** A parallelogram is depicted. * Base: Labeled as "x + 15". * Height: Labeled as "x". * Right Angle Symbol: Indicates the height is perpendicular to the base. * **Problem Statement:** "Find the height of the parallelogram given its area with 100 square units." * **Model Response:** A textual solution to the problem. ### Detailed Analysis or Content Details The model response provides the following steps: 1. **Area Formula:** States that the area of a parallelogram is equal to the product of the base and the height. 2. **Equation Setup:** Sets up the equation: 100 = (x + 15)x 3. **Quadratic Equation:** Expands the equation to: x² + 15x - 100 = 0 4. **Roots:** Finds the roots of the quadratic equation: x₁ = 5 and x₂ = -20 5. **Root Selection:** Discards the negative root (-20) as unacceptable for a height. 6. **Final Answer:** States the final answer is 5. ### Key Observations The problem is a straightforward application of the area formula for a parallelogram, leading to a quadratic equation. The solution correctly identifies and discards the negative root, recognizing that height cannot be negative. ### Interpretation The problem demonstrates the relationship between the area, base, and height of a parallelogram. The solution process illustrates how to translate a geometric problem into an algebraic equation and solve it. The discarding of the negative root highlights the importance of considering the physical constraints of the problem (height must be positive). The final answer of 5 indicates that the height of the parallelogram is 5 units, and the base is 20 units (5 + 15). </details> Figure 14 | Solving a geometrical reasoning task. The model shows good understanding of the task and is able to provide meaningful reasoning steps despite slightly unclear instructions. Source: Lu et al. (2021). 10.4.6. Information seeking about objects <details> <summary>Image 22 Details</summary>  ### Visual Description \n ## Image: Lunar Golf Connection ### Overview The image presents two distinct visuals: a close-up photograph of the Moon, and a photograph of a hand holding a golf ball indoors. The accompanying text prompts a connection between the two, hinting at a historical event. ### Components/Axes There are no axes or components in the traditional sense of a chart or diagram. The image consists of two photographs and a text prompt. ### Detailed Analysis or Content Details The left image is a detailed photograph of the Moon's surface, showing craters and varying shades of gray. The right image depicts a hand, presumably human, holding a white golf ball against a blurred indoor background with visible ceiling lights and what appears to be a screen or display in the background. The text below the images reads: "Find a connection between these. Hint: think about historical events." The text below that reads: "The Moon is the only celestial body in which humans have played golf. In 1971, the Apollo 14 crew hit two golf balls on the lunar surface." ### Key Observations The key observation is the connection between the Moon and a golf ball, prompted by the text. The hint directs the viewer to consider historical events. ### Interpretation The image and text reveal a historical fact: astronauts from the Apollo 14 mission played golf on the Moon in 1971. This makes the Moon the only celestial body where golf has been played by humans. The image serves as a visual representation of this unique event, juxtaposing the lunar surface with a golf ball to highlight the connection. The prompt encourages the viewer to recall or discover this historical anecdote. The image is not presenting data, but rather illustrating a fact. </details> Figure 15 | Solving a puzzle using multimodal inputs. The model recognizes the objects in the images and identifies a commonality that connects the two objects. Source: photo taken by an author from the Gemini team. ## 10.4.7. Multimodal reasoning based on visual cues ## Prompt <details> <summary>Image 23 Details</summary>  ### Visual Description \n ## Photograph: Nighttime Cityscape ### Overview The image is a nighttime photograph of a city street, likely in New York City, featuring prominent skyscrapers, pedestrian activity, and street infrastructure. The Empire State Building is visible in the distance, illuminated. The scene appears to be a busy urban environment. ### Components/Axes There are no axes or legends present in this image. The key components are: * **Buildings:** Various skyscrapers of differing heights and architectural styles. * **Street:** A wide street with pedestrian crosswalks. * **Vehicles:** Cars and potentially buses are visible on the street. * **Pedestrians:** People are walking along the sidewalks and crossing the street. * **Streetlights:** Numerous streetlights illuminate the scene. * **Construction Barriers:** Orange construction barriers are present on the right side of the image. * **Traffic Signals:** A traffic signal is visible on the left side of the image. * **Signage:** A street sign reading "34th St" is visible. ### Detailed Analysis or Content Details The image depicts a street scene at night. The Empire State Building is the most prominent landmark, appearing as a brightly lit spire in the center-background. The buildings on either side of the street vary in height and design. The street is wet, suggesting recent rain. Pedestrians are present, some waiting to cross the street, others walking along the sidewalks. Vehicles are visible, including a dark-colored SUV in the foreground. The construction barriers indicate ongoing work in the area. The traffic signal on the left is red. The street sign clearly indicates "34th St". ### Key Observations * The Empire State Building serves as a focal point, drawing the eye towards the center of the image. * The wet street surface reflects the city lights, creating a shimmering effect. * The presence of construction barriers suggests ongoing urban development. * The red traffic signal indicates a stop for vehicles. * The image captures a typical bustling nighttime scene in a major city. ### Interpretation The photograph portrays a vibrant urban environment, likely in Midtown Manhattan, New York City. The presence of the Empire State Building immediately establishes the location. The combination of illuminated skyscrapers, pedestrian activity, and street infrastructure conveys a sense of energy and dynamism. The construction barriers suggest a city constantly evolving and adapting. The wet street and nighttime setting contribute to a moody and atmospheric quality. The image doesn't present quantifiable data, but rather a snapshot of urban life, offering a visual representation of the city's scale, activity, and ongoing development. The image is a documentation of a moment in time, capturing the essence of a bustling city at night. </details> Do you know the precise location where this image was taken? ## Model Response This image was taken in New York City. The Empire State Building is visible in the background. The street in the foreground is 8th Avenue. The cross street is West 34th Street. Figure 16 | Identifying the objects in the image (the Empire State Building) and recognizing what those are even with small levels of visual distortion in the image. Based on the image, the model is also able to correctly identify the precise location of the person taking the photo. Source: photo taken by an author from the Gemini team. ## 10.4.8. Multimodal humor understanding ## Prompt <details> <summary>Image 24 Details</summary>  ### Visual Description \n ## Meme: Gaming Performance Illustration ### Overview The image is a two-panel meme illustrating the perceived disconnect between high frame rates (FPS) in a game and the refresh rate of a monitor. It uses a common meme format of a person reaching for a large sphere representing a desirable outcome, with the second panel showing a diminished outcome due to a limiting factor. The image does not contain factual data, but rather a conceptual illustration. ### Components/Axes The image consists of two panels, each containing a simplified human figure and a large sphere. Each sphere has text labeling the game's performance. * **Panel 1:** * Figure: A white, simplistic human figure with an excited expression. * Sphere: A yellow sphere labeled "Game at 300 FPS" in black text. * **Panel 2:** * Figure 1: A pink, simplistic human figure with an excited expression. This figure is labeled "75 hz Monitor" in black text. * Figure 2: A white, simplistic human figure with an excited expression. * Sphere: A smaller, yellow sphere labeled "Game at FPS" in black text. The FPS value is incomplete. ### Detailed Analysis or Content Details The image does not contain numerical data beyond the stated FPS values. The key elements are the labels and the visual representation of the size difference between the spheres in each panel. * **Panel 1:** The figure is reaching for a large sphere representing a game running at 300 frames per second (FPS). * **Panel 2:** The figure is reaching for a smaller sphere representing a game running at an unspecified FPS. The figure is labeled as a "75 hz Monitor", indicating that the monitor's refresh rate is 75 Hertz. ### Key Observations The primary observation is the contrast in sphere size between the two panels. The larger sphere in the first panel visually represents the potential for a high-performance gaming experience, while the smaller sphere in the second panel suggests a limited experience due to the monitor's refresh rate. The incomplete FPS value in the second panel implies that the monitor caps the displayed FPS. ### Interpretation The meme illustrates the concept that a high frame rate generated by a gaming PC is only fully realized if the monitor has a correspondingly high refresh rate. A monitor with a lower refresh rate (like the 75 Hz monitor in the image) cannot display all the frames generated by the game, effectively limiting the perceived smoothness and responsiveness. The meme suggests that investing in a high-end gaming PC is less beneficial if paired with a monitor that cannot take full advantage of its capabilities. The incomplete FPS value in the second panel is a humorous way of indicating that the actual FPS displayed will be capped at 75, regardless of the game's potential to render more frames. The meme is a commentary on the importance of matching components for optimal gaming performance. It highlights the bottleneck effect where the weakest link in the system limits the overall performance. </details> Can you explain this meme? ## Model Response The meme is about a person who is playing a game at 300 FPS, but their monitor is only 75Hz. The person is trying to catch the game, but it is too fast for their monitor to display. The meme is funny because it is relatable to anyone who has ever played a game on a monitor that is not fast enough. Figure 17 | Explanation of humor in a meme. The model is showing the ability to not only describe what is happening in the image but also what it means even though the cultural context is not mentioned explicitly in the image or the prompt. Source: Hwang and Shwartz (2023). ## 10.5. Commonsense reasoning in a multilingual setting <details> <summary>Image 25 Details</summary>  ### Visual Description \n ## Diagram: The Basic Chinese Family Tree ### Overview The image presents a diagram illustrating a basic Chinese family tree, depicting familial relationships and their corresponding Chinese terms. The diagram uses circular nodes to represent individuals, connected by lines indicating relationships. A red box highlights the "我 (wǒ)" node, representing "I" or "me". Below the diagram is a question prompting the user to determine the correct terms for grandparents on each side of the family. ### Components/Axes The diagram consists of the following components: * **Nodes:** Circular representations of family members. * **Lines:** Connecting lines indicating familial relationships (parent-child, sibling). * **Labels:** Chinese characters and Pinyin transliterations labeling each family member. * **Title:** "The Basic Chinese Family Tree" at the top of the diagram. * **Highlighted Node:** A red box surrounds the node labeled "我 (wǒ)". The labels are arranged in a hierarchical structure, with grandparents at the top, parents in the middle, and children at the bottom. ### Content Details Here's a transcription of the labels, including Chinese characters and Pinyin: * **爷爷 (yéye):** Paternal Grandfather * **奶奶 (nǎinai):** Paternal Grandmother * **外婆 (wàipó):** Maternal Grandmother * **外公 (wàigōng):** Maternal Grandfather * **爸爸 (bàba):** Father * **妈妈 (māmā):** Mother * **弟弟 (dìdi):** Younger Brother * **妹妹 (mèimei):** Younger Sister * **我 (wǒ):** I/Me * **哥哥 (gēge):** Older Brother * **姐姐 (jiějie):** Older Sister The diagram shows the following relationships: * 爷爷 (yéye) and 奶奶 (nǎinai) are the parents of 爸爸 (bàba). * 外公 (wàigōng) and 外婆 (wàipó) are the parents of 妈妈 (māmā). * 爸爸 (bàba) and 妈妈 (māmā) are the parents of 弟弟 (dìdi), 妹妹 (mèimei), 我 (wǒ), 哥哥 (gēge), and 姐姐 (jiějie). ### Key Observations The diagram clearly distinguishes between paternal and maternal sides of the family using different sets of terms. The highlighted node "我 (wǒ)" serves as the reference point for understanding the relationships. The arrangement of nodes visually represents the generational structure of the family. ### Interpretation The diagram serves as a visual aid for learning and understanding Chinese kinship terms. It demonstrates how the Chinese language uses specific terms to differentiate between grandparents on the father's and mother's sides. The diagram's structure emphasizes the importance of family relationships in Chinese culture. The question posed below the diagram tests the user's comprehension of the presented information. The diagram is a simplified representation of a family tree, focusing on the core relationships and terminology. It does not include extended family members or complex relationships. </details> Figure 18 | Common-sense reasoning in images. The model is able to understand the relationships represented in the graphs and reason about them in a multilingual setting. Source: image created by an author from the Gemini team. ## 10.5.1. Reasoning and code generation ## Prompt Create a web app called "Opossum Search": 1. Every time you make a search query, it should redirect you to a google search with the same query, but the word opossum before it. 2. It should be visually similar to Google search, 3. Instead of the google logo, it should have a picture of an opossum from the internet. 4. It should be a single html file, no separate js or css files. 5. It should say "Powered by google search" in the footer. ## Model Response (rendered code) ## Here is the code for Opossum Search: ``` 5. It should say "Powered by google search" in the footer. ``` Figure 19 | Writing code for a website based on user request. The model follows the instructions and requirements defined by the user and converts them to functioning code. <details> <summary>Image 26 Details</summary>  ### Visual Description \n ## Code Snippet: HTML and JavaScript ### Overview The image presents a snippet of HTML and JavaScript code. The code defines a simple web application that redirects the user to a Google search with the term "opossum" prepended to their search query. The code includes a footer indicating it is powered by Google Search. ### Components/Axes The code is structured as follows: * **HTML:** Contains the `<footer>` and `<body>` tags. * **JavaScript:** Contained within `<script>` tags, handles the search functionality. * **Footer:** Displays "Powered by Google Search". * **JavaScript Variables:** `searchInput`, `searchButton`. * **Event Listener:** Attached to the `searchButton` for the 'click' event. * **Search Query:** Retrieved from the `searchInput` value. * **Redirection:** Uses `window.location.href` to redirect to a Google search URL. ### Detailed Analysis or Content Details The code can be transcribed as follows: ```html 62 <footer> 63 Powered by Google Search 64 </footer> 65 <script> 66 const searchInput = document.querySelector('.search-input'); 67 const searchButton = document.querySelector('.search-button'); 68 69 searchButton.addEventListener('click', () => { 70 const query = searchInput.value; 71 if (query) { 72 window.location.href = 'https://www.google.com/search?q=opossum+' + query; 73 } 74 }); 75 </script> 76 </body> ``` The accompanying text states: "This code creates a simple web app that looks similar to Google Search, but with an opossum logo. When you enter a search query and click the 'Search' button, it will redirect you to a Google search with the word 'opossum' added to the beginning of your query. The app is powered by Google Search, as indicated in the footer." ### Key Observations * The code uses `document.querySelector` to select HTML elements based on their CSS classes (`.search-input`, `.search-button`). * The JavaScript code is concise and uses an arrow function for the event listener. * The redirection URL is constructed by concatenating the base Google search URL with the query string "q=opossum+" followed by the user's input. * The code includes a basic check to ensure that the search query is not empty before redirecting. ### Interpretation The code demonstrates a simple web application that leverages Google Search functionality while adding a playful twist (the "opossum" prefix). The code is straightforward and effectively implements the desired behavior. The use of JavaScript event listeners and DOM manipulation is standard practice for creating interactive web applications. The footer attribution to Google Search suggests that the application is built upon or utilizes Google's search infrastructure. The code's simplicity suggests it's likely a proof-of-concept or a small-scale project rather than a complex, production-ready application. The addition of "opossum" to the search query is likely intended as a humorous or branding element. </details> <details> <summary>Image 27 Details</summary>  ### Visual Description \n ## Screenshot: Opossum Search Website ### Overview The image is a screenshot of a simple website interface labeled "Opossum Search". It features a search bar powered by Google Search, with a header containing the website title and an image of an opossum. The overall design is minimalist, with a dark color scheme. ### Components/Axes The screenshot contains the following elements: * **Header:** Contains the text "Opossum Search" in a large font, and an image of an opossum. * **Search Bar:** A text input field labeled "Search the web" with a "Search" button to its right. * **Footer:** Contains the text "Powered by Google Search". * **Opossum Image:** A photograph of an opossum. ### Detailed Analysis or Content Details The text elements are as follows: * **Title:** "Opossum Search" * **Search Bar Placeholder:** "Search the web" * **Search Button:** "Search" * **Powered By:** "Powered by Google Search" The opossum image appears to show a gray opossum on a wooden surface. The image is centered within the header. ### Key Observations The website is a search interface specifically themed around opossums. The use of "Opossum Search" as the title and the inclusion of an opossum image suggest a focus on information related to opossums. The integration with Google Search indicates that the website utilizes Google's search engine to provide search results. ### Interpretation The website appears to be a novelty search engine, or a themed portal to Google Search. It's likely designed to attract users interested in opossums, providing a branded search experience. The simplicity of the design suggests a focus on functionality over aesthetics. The "Powered by Google Search" attribution indicates that the website doesn't maintain its own search index but relies on Google's infrastructure. The image of the opossum serves as a visual cue to reinforce the website's theme. There is no data or trends to analyze, it is a simple UI. </details> ## 10.5.2. Mathematics: Calculus ## Prompt Find the derivative of sinh 𝑥 + cosh 𝑦 = 𝑥 + 𝑦 . ## Model Response (rendered latex) We have that 𝑑 𝑑𝑥 sinh 𝑥 = cosh 𝑥 and 𝑑 𝑑𝑥 cosh 𝑦 = sinh 𝑦 𝑑𝑦 𝑑𝑥 . Therefore, $$\cosh x + \sinh y { \frac { d y } { d x } } = 1 + { \frac { d y } { d x } } .$$ $${ \frac { d y } { d x } } = { \frac { 1 - \cosh x } { \sinh y - 1 } } .$$ Final Answer: The final answer is 1 - cosh 𝑥 sinh 𝑦 - 1 . Figure 20 | Solving a calculus problem. The model is able to get a solution to a calculus problem with step-by-step explanation and correctly defined LaTeX equations. Source: question is provided by Macmillan Learning. Solving for 𝑑𝑦 𝑑𝑥 , we find that ## 10.6. Multi-step reasoning and mathematics ## Prompt In the year 2018 factory A produced goods worth 500 million forints, while factory B produced goods worth 400 million forints. According to their long-term plans, factory A will increase production by 5% each year, while factory B will increase production by 6% per year. The above plans were published on an economy website. A debate started in the comment section of the article: one user claimed that the difference between the volumes of production between the two factories would get smaller and smaller in the next couple of years. Calculate the missing values of the table below and show that the above claim is not true. ``` Header: | 2018 | 2019 | 2020 | 2021 | Production volume of factory A (mln) | 500 | | | | | Production volume of factory B (mln) | 400 | | | | | ``` ## Model Response (rendered Markdown) | Year | Production volume of factory A (mln) | Production volume of factory B (mln) | Difference (mln) | |--------|----------------------------------------|----------------------------------------|--------------------| | 2018 | 500 | 400 | 100 | | 2019 | 525 | 424 | 101 | | 2020 | 551.25 | 449.44 | 101.81 | | 2021 | 578.81 | 476.41 | 102.4 | The production volume of factory A in 2019 is 500 * 1.05 = 525 million forints. The production volume of factory B in 2019 is 400 * 1.06 = 424 million forints. The difference in production volume in 2019 is 525 - 424 = 101 million forints. The production volume of factory A in 2020 is 525 * 1.05 = 551.25 million forints. The production volume of factory B in 2020 is 424 * 1.06 = 449.44 million forints. The difference in production volume in 2020 is 551.25 - 449.44 = 101.81 million forints. The production volume of factory A in 2021 is 551.25 * 1.05 = 578.81 million forints. The production volume of factory B in 2021 is 449.44 * 1.06 = 476.41 million forints. The difference in production volume in 2021 is 578.81 - 476.41 = 102.4 million forints. As you can see, the difference in production volume between the two factories is actually increasing, not decreasing. Therefore, the claim that the difference between the volumes of production between the two factories would get smaller and smaller in the next couple of years is not true. Figure 21 | Solving a multi-step math problem. The model is able to understand the task and generate a markdown table with correctly calculated values. It also explicitly follows the instructions to show where the numbers come from and answer the question given in the task. Source: Oktatási Hivatal (2023, p. 20) ## 10.6.1. Complex image understanding, code generation, and instruction following ## Prompt <details> <summary>Image 28 Details</summary>  ### Visual Description \n ## Charts/Graphs: Mathematical Function Plots ### Overview The image contains four separate plots of mathematical functions. The plots are arranged in a 2x2 grid. Each plot displays a function's behavior over a domain of approximately 0 to 10 on the x-axis, with varying ranges on the y-axis. The plots appear to be generated using a simple line graph style. ### Components/Axes All four plots share similar axis characteristics: * **X-axis:** Ranges from approximately 0 to 10, with tick marks at integer values. No explicit label is present, but it represents the independent variable. * **Y-axis:** Each plot has a different scale and range, as described below. No explicit label is present, but it represents the dependent variable. ### Detailed Analysis or Content Details **Plot 1 (Top-Left):** * **Trend:** The line oscillates sinusoidally, completing approximately 2.5 cycles over the domain. * **Y-axis Range:** Approximately -1.0 to 1.0. * **Data Points (approximate):** * x = 0, y = 1.0 * x = 2, y = -0.75 * x = 4, y = 1.0 * x = 6, y = -0.75 * x = 8, y = 1.0 * x = 10, y = 0.25 **Plot 2 (Top-Right):** * **Trend:** The line exhibits a series of vertical asymptotes and curves, suggesting a rational function with poles. There are three distinct vertical sections. * **Y-axis Range:** Approximately -40 to 20. * **Data Points (approximate):** * x = 1, y = 10 * x = 2, y = -10 * x = 3, y = 10 * x = 4, y = -10 * x = 5, y = 10 * x = 6, y = -10 * x = 7, y = 10 * x = 8, y = -10 * x = 9, y = 10 **Plot 3 (Bottom-Left):** * **Trend:** The line exhibits exponential growth, starting near zero and rapidly increasing. * **Y-axis Range:** Approximately 0 to 20000. * **Data Points (approximate):** * x = 0, y = 0 * x = 2, y = 2000 * x = 4, y = 6000 * x = 6, y = 12000 * x = 8, y = 18000 * x = 10, y = 20000 **Plot 4 (Bottom-Right):** * **Type:** 3D surface plot. * **X-axis Range:** Approximately 0 to 1. * **Y-axis Range:** Approximately 0 to 1. * **Z-axis Range:** Approximately 0 to 2. * **Trend:** The surface rises from a minimum value near z=0 to a maximum value near z=2, forming a curved shape. The surface is relatively flat near x=0 and y=0, and becomes steeper as x and y increase. * **Color Gradient:** The color gradient indicates the z-value, with darker shades representing lower values and lighter shades representing higher values. ### Key Observations * The plots demonstrate a variety of mathematical functions, including trigonometric, rational, exponential, and a 3D surface. * The exponential function (Plot 3) exhibits the most rapid growth. * The rational function (Plot 2) has repeating vertical asymptotes. * The 3D surface plot (Plot 4) shows a non-linear relationship between x, y, and z. ### Interpretation The image serves as a visual representation of different mathematical functions and their properties. The plots demonstrate how functions can behave in different ways, exhibiting oscillations, asymptotes, exponential growth, and complex surfaces. The choice of functions suggests an exploration of fundamental mathematical concepts. The lack of labels makes it difficult to determine the specific functions being plotted, but the visual characteristics provide clues about their nature. The 3D plot suggests a function of two variables, while the other plots represent functions of a single variable. The image could be used for educational purposes to illustrate the diversity of mathematical functions and their graphical representations. </details> I want you to take the function depicted in the top left subplot, multiply it by 1000, and then add it to the function depicted in the bottom left subplot. Generate matplotlib code for the single resulting plot. ## Model Response (rendered code) ``` import numpy as np import matplotlib.pyplot as plt x = np.linspace(0, 10, 1000) y1 = np.sin(x) y2 = np.exp(x) plt.plot(x, 1000*y1 + y2) plt.show() ``` ## Rendered Graph Figure 22 | Multimodal reasoning capabilities applied to code generation. Gemini Ultra needs to perform inverse graphics task to infer the code that would have generated the plots, perform additional mathematical transformations, and generate relevant code. <details> <summary>Image 29 Details</summary>  ### Visual Description \n ## Chart: Single Line Graph ### Overview The image displays a line graph with a single data series plotted against an unspecified variable on the x-axis and a numerical value on the y-axis. The graph shows an initial decrease, followed by a minimum, and then a rapid increase. ### Components/Axes * **X-axis:** Ranges from 0 to 10, with tick marks at integer values. No label is present. * **Y-axis:** Ranges from 0 to 20000, with tick marks at 5000-unit intervals. No label is present. * **Data Series:** A single blue line representing the plotted data. * **Legend:** No legend is present. ### Detailed Analysis The blue line begins at approximately (0, 500). It initially decreases to a minimum value around (3, -500). From this minimum, the line increases, becoming steeper as the x-value increases. At x = 8, the y-value is approximately 5000. The line continues to increase rapidly, reaching approximately (10, 20000). Here's a breakdown of approximate data points: * (0, 500) * (1, 1500) * (2, 2000) * (3, -500) * (4, 0) * (5, 500) * (6, 1500) * (7, 3000) * (8, 5000) * (9, 10000) * (10, 20000) The line exhibits a clear trend: a decrease to a minimum, followed by exponential growth. ### Key Observations * The graph shows a significant change in the rate of increase after x = 6. * The minimum value of the function is negative, indicating a portion of the curve lies below the x-axis. * The rapid increase suggests a potential exponential or polynomial relationship. ### Interpretation The graph likely represents a function that initially decreases before experiencing a period of growth. The shape of the curve suggests a relationship where the rate of change is initially negative, reaches a minimum, and then becomes increasingly positive. This could model various phenomena, such as population growth after an initial decline, the learning curve of a skill, or the acceleration of a process. Without axis labels, the specific meaning of the graph remains ambiguous, but the mathematical relationship is clearly non-linear. The absence of a legend is not problematic as there is only one data series. The graph demonstrates a function with a local minimum and a rapid increase towards larger values. </details> Source: figure generated by an author from the Gemini team. ## 10.6.2. Video understanding and reasoning ## Prompt (video) <details> <summary>Image 30 Details</summary>  ### Visual Description \n ## Photograph: Soccer Kick Sequence ### Overview The image is a composite photograph showing four sequential frames of a person kicking a soccer ball on a grass field. The frames capture the motion of the kick, from the initial stance to the follow-through. A soccer goal is visible in the background of each frame. ### Components/Axes There are no axes or legends present in this image. The components are: * **Person:** A single individual wearing a long-sleeved shirt and shorts. * **Soccer Ball:** A white soccer ball. * **Soccer Goal:** A standard soccer goal with netting. * **Grass Field:** A green grass field serving as the playing surface. * **Trees:** Trees are visible in the background, suggesting a park or recreational area. * **Sky:** A partly cloudy sky. ### Detailed Analysis or Content Details The image is divided into four quadrants, each containing one frame of the sequence: * **Top-Left:** The person is standing facing the goal, with the ball positioned slightly in front of them. Their kicking leg is bent, preparing for the kick. * **Top-Right:** The person is in the midst of kicking the ball. Their leg is extended forward, making contact with the ball. The body is leaning forward. * **Bottom-Left:** The person is following through with the kick. Their leg is continuing its motion forward and upward. The ball is now airborne. * **Bottom-Right:** The person is completing the follow-through, with their leg extended and body rotating slightly. They are looking towards where the ball was kicked. There is no quantifiable data present in the image. The image is purely visual and demonstrates a physical action. ### Key Observations The sequence clearly demonstrates the mechanics of a soccer kick. The frames show the progression from preparation to impact to follow-through. The lighting and background are consistent across all four frames, indicating they were taken in quick succession. ### Interpretation The image illustrates a fundamental skill in soccer – kicking the ball. The sequence could be used for instructional purposes, demonstrating proper technique. The setting suggests a casual practice session rather than a competitive game. The image focuses on the biomechanics of the kick, highlighting the coordination and movement involved. The absence of other players or spectators suggests a solitary practice session. The image does not provide any information about the skill level of the person, the distance to the goal, or the intended target. It is a purely descriptive representation of a physical action. </details> How could this person improve their technique? ## Model Response This soccer player needs to work on their ball-striking mechanics. Their foot is not making good contact with the ball, which is causing the ball to fly up and away from the goal. They also need to work on their body positioning. They should be leaning into the shot with their non-kicking leg in front of their kicking leg, and they should be following through with their kicking leg. Figure 23 | Video understanding and reasoning over the situation presented in the video. Here, we provide a video as input to the model together with a text prompt (images are provided here only for visualization purposes). The model is able to analyze what happened in the video and provide recommendations on how the actions in the video could have been better. Video source: "Football/Soccer Penalty Miss" https://www.youtube.com/watch?v=VmWxjmJ3mvs