2406.04271

# Buffer of Thoughts: Thought-Augmented Reasoning with Large Language Models > Equal Contribution. 🖂  yangling0818@163.com Abstract We introduce Buffer of Thoughts (BoT), a novel and versatile thought-augmented reasoning approach for enhancing accuracy, efficiency and robustness of large language models (LLMs). Specifically, we propose meta-buffer to store a series of informative high-level thoughts, namely thought-template, distilled from the problem-solving processes across various tasks. Then for each problem, we retrieve a relevant thought-template and adaptively instantiate it with specific reasoning structures to conduct efficient reasoning. To guarantee the scalability and stability, we further propose buffer-manager to dynamically update the meta-buffer, thus enhancing the capacity of meta-buffer as more tasks are solved. We conduct extensive experiments on 10 challenging reasoning-intensive tasks, and achieve significant performance improvements over previous SOTA methods: 11% on Game of 24, 20% on Geometric Shapes and 51% on Checkmate-in-One. Further analysis demonstrate the superior generalization ability and model robustness of our BoT, while requiring only 12% of the cost of multi-query prompting methods (e.g., tree/graph of thoughts) on average. Notably, we find that our Llama3-8B + BoT has the potential to surpass Llama3-70B model. Our project is available at https://github.com/YangLing0818/buffer-of-thought-llm 1 Introduction A series of Large Language Models (LLMs) [1, 2, 3, 4, 5] like GPT-4 [3], PaLM [2] and LLaMA [6, 7] have showcased the impressive performance in various reasoning tasks. In addition to scaling up the model size to improve the reasoning performance, there are more effective prompting methods that further enhance the functionality and performance of LLMs. We divide these methods into two categories: (i) single-query reasoning: these methods [8, 9, 10] usually focus on prompt engineering and their reasoning process can be finished within a single query, such as CoT [8] that appends the input query with ’Let’s think step by step’ to produce rationales for increasing reasoning accuracy, and Few-shot Prompting [11, 12, 9, 13] which provides task-relevant exemplars to assist the answer generation; (ii) multi-query reasoning: these methods [14, 15] focus on leveraging multiple LLM queries to elicit different plausible reasoning paths, thus decomposing a complex problem into a series of simpler sub-problems, such as Least-to-Most [16], ToT [14] and GoT [17]. However, both kinds of methods face some limitations: (1) single-query reasoning usually requires prior assumption or relevant exemplars of reasoning process, which makes it impractical to manually design them task by task, thus lacking universality and generalization; (2) Due to the recursive expansion of reasoning paths, multi-query reasoning is usually computationally-intensive when finding a unique intrinsic structure underlying the reasoning process for each specific task; (3) Both single-query and multi-query reasoning processes are limited by their designed exemplars and reasoning structures, and they neglect to derive general and high-level guidelines or thoughts from previously-completed tasks, which are informative for improving efficiency and accuracy when solving similar problems. <details> <summary>x1.png Details</summary>  ### Visual Description \n ## Diagram: LLM Reasoning Approaches ### Overview This diagram illustrates three different approaches to reasoning with Large Language Models (LLMs), comparing their processes and outcomes in terms of accuracy and efficiency. The approaches are Single-Query, Multi-Query, and a Buffer of Thoughts (BoT) method. Each approach is depicted as a flow diagram, showing the progression from input query to output, with intermediate steps of LLM processing and reasoning. The diagram also visually indicates the relative accuracy and efficiency of each approach using downward or upward arrows and corresponding emoticons. ### Components/Axes The diagram consists of three horizontal sections, each representing a different reasoning approach. Each section includes the following components: * **Input query:** Represented by a cloud-shaped icon. * **LLM:** A rectangular box labeled "LLM". * **Intermediate Steps:** Boxes representing processing stages (e.g., Reasoning, Thought Expansion, Problem Distiller, Instantiated Reasoning). * **Output:** A rounded rectangular box labeled "Output". * **Arrows:** Indicate the flow of information between components. * **Text Annotations:** Explanatory text describing the process within each approach. * **Accuracy/Efficiency Indicators:** Downward or upward arrows with corresponding emoticons to indicate relative performance. * **"N-hop Iteration"**: Text indicating iterative process in the Multi-Query approach. * **"Buffer of Thoughts (BoT)"**: Label for the third approach. * **"Meta Buffer"**: A rectangular box within the BoT approach. ### Detailed Analysis or Content Details **1. Single-Query Approach (Top Section):** * Input query flows to the LLM. * LLM performs "Reasoning" (Manual Prompt for Specific Task, e.g., CoT, Few-shot Prompting). * Output is generated. * Accuracy is indicated by a downward arrow and a sad face emoticon. **2. Multi-Query Approach (Middle Section):** * Input query flows to the LLM. * LLM utilizes a "Multi-Query" approach with a "Pre-defined Query Structure" (e.g., ToT, GoT). * This leads to "Thought Expansion". * "Thought Expansion" feeds into "Reasoning". * "Reasoning" generates the Output. * The process is labeled as "N-hop Iteration" with a curved arrow indicating a loop. * Efficiency is indicated by a downward arrow and a sad face emoticon. **3. Buffer of Thoughts (BoT) Approach (Bottom Section):** * Input query flows to the LLM. * LLM uses a "Problem Distiller". * "Problem Distiller" generates "High-level Thoughts". * "High-level Thoughts" are retrieved from a "Buffer of Thoughts" (BoT) and are represented by a person icon with a lightning bolt. * "High-level Thoughts" are used in "Instantiated Reasoning". * "Instantiated Reasoning" undergoes "Thought Distillation and Update" using a "Meta Buffer". * Output is generated. * Accuracy and Efficiency are indicated by an upward arrow and a smiling face emoticon. ### Key Observations * The diagram visually represents a progression in complexity and performance. The Single-Query approach is the simplest but least effective, while the BoT approach is the most complex but most effective. * The use of arrows and emoticons provides a clear and intuitive understanding of the relative performance of each approach. * The BoT approach incorporates a feedback loop ("Thought Distillation and Update") suggesting a continuous improvement process. * The "N-hop Iteration" in the Multi-Query approach indicates a repetitive process, potentially contributing to its lower efficiency. ### Interpretation The diagram demonstrates a comparative analysis of different reasoning strategies for LLMs. It suggests that more sophisticated approaches, like the Buffer of Thoughts (BoT) method, can significantly improve both accuracy and efficiency compared to simpler methods like Single-Query. The BoT approach, by leveraging a "Buffer of Thoughts" and a "Meta Buffer" for distillation and update, appears to mimic a more human-like reasoning process, allowing the LLM to learn and refine its responses over time. The diagram highlights the trade-offs between simplicity and performance, suggesting that investing in more complex reasoning architectures can yield substantial benefits in terms of output quality and resource utilization. The use of visual cues (arrows, emoticons) effectively communicates the relative strengths and weaknesses of each approach, making the diagram accessible to a broad audience. The diagram is not presenting data, but rather a conceptual model of different approaches. It is a qualitative comparison, not a quantitative one. </details> Figure 1: Comparison between single-query [8, 11], multi-query [14, 17], and (c) our BoT methods. To address these limitations, we propose Buffer of Thoughts (BoT), a novel and versatile thought-augmented reasoning framework aimed at enhancing reasoning accuracy, efficiency and robustness of LLMs across various tasks. Specifically, we design meta-buffer, a lightweight library housing a series of universal high-level thoughts (thought-template), which are distilled from different problem-solving processes and can be shared across tasks. Then, for each problem, we retrieve a relevant thought-template and instantiate it with specific reasoning structure for efficient thought-augmented reasoning. In order to guarantee the scalability and stability of our BoT, we further propose buffer-manager to dynamically update the meta-buffer, which effectively enhances the capacity of meta-buffer as more tasks are solved. Our method has three critical advantages: (i) Accuracy Improvement: With the shared thought-templates, we can adaptively instantiate high-level thoughts for addressing different tasks, eliminating the need to build reasoning structures from scratch, thereby improving reasoning accuracy. (ii) Reasoning Efficiency: Our thought-augmented reasoning could directly leverage informative historical reasoning structures to conduct reasoning without complex multi-query processes, thus improving reasoning efficiency. (iii) Model Robustness: The procedure from thought retrieval to thought instantiation is just like the human thought process, enabling LLMs to address similar problems in a consistent way, thus significantly enhancing the model robustness of our method. Our empirical studies demonstrate that Buffer of Thoughts significantly improves precision, efficiency, and robustness over a diverse array of tasks. Here, we summarize our contributions as follows: 1. We propose a novel thought-augmented reasoning framework Buffer of Thoughts (BoT) for improving the accuracy, efficiency and robustness of LLM-based reasoning. 1. We propose meta-buffer for store informative high-level thoughts distilled from different problems, and adaptively instantiate each thought template to address each specific task. 1. We design buffer-manager to distill thought-templates from various solutions, and is continually improves the capacity of meta-buffer as more tasks are solved. 1. We conduct extensive experiments on 10 challenging reasoning-intensive tasks. Our BoT achieves significant performance improvements over previous SOTA methods: 11% on Game of 24, 20% on Geometric Shapes and 51% on Checkmate-in-One, while requiring only 12% of the cost of multi-query prompting methods on average. 2 Related Work and Discussions Retrieval-Augmented Language Models The retrieval-augmented (Large) Language Model is introduced as a solution to mitigate the phenomenon of hallucination and enhance the output quality of language models [18, 19, 20, 21, 22]. When presented with an input question, the retrieval-augmented LLM first queries an external database with billion-level tokens [23] for retrieving a subset of the text corpus to help generating the final answer. Notably, the retrieval-augmented LLM achieves superior question-answering performance using fewer parameters compared to conventional LLMs [19], and it has found application across various downstream tasks [24, 25, 26], including multi-modal generation [24, 22, 23, 25] and biomedical applications [26, 27]. In this paper, we construct a novel category of retrieval database, termed meta-buffer, which contains a series of high-level thoughts rather than specific instances, aiming to universally address various tasks for LLM-based reasoning. Prompt-based Reasoning with Large Language Models Prompting techniques have significantly enahnced the arithmetic and commonsense reasoning capabilities of LLMs. Chain-of-Thought (CoT) prompting [8] and its variants [28, 29, 30], such as Least-to-Most [16], Decomposed Prompting [31], and Auto-CoT [13] —prompt LLMs to break down complex questions into simpler subtasks and systematically solve them before summarizing a final answer. Numerous studies [32, 33, 34, 35, 36, 37] have demonstrated the effectiveness of these prompting methods across a wide range of tasks and benchmarks. Innovations like Tree-of-Thought [14] and Graph-of-Thought [17], have further advanced this field by exploring dynamic, non-linear reasoning pathways to expand heuristic capabilities of LLMs [38, 39]. However, they suffer from increased resource demands and greater time complexity, depend on manual prompt crafting, and are often tailored to specific task types. Recent meta prompting methods [15, 40] utilize a same task-agnostic form of prompting for various tasks and recursively guide a single LLM to adaptively addressing different input queries. Nevertheless, such a long meta prompt may require a considerable context window, and these methods fail to leverage historical informative guidelines or thoughts for potential similar tasks. Analogical Reasoning Analogical reasoning is a useful technique for natural language reasoning [41, 42, 43, 44, 45]. Recent works demonstrate that LLMs can perform analogical reasoning just like humans [46, 47, 12, 48, 49]. For example, Analogical Prompting [12] and Thought Propagation [48] prompt LLMs to self-generate a set of analogous problems, and then utilize the results of analogous problems to produce a solution for input problem. However, the specific solutions for self-explored problems may introduce additional noise and cause error accumulation. Recent Thought-Retriever [49] uses the intermediate thoughts generated when solving past user to address analogous queries, but it only focuses on textual comprehension/generation instead of general reasoning problems. Thus, a more high-level and general analogical approach for LLM complex reasoning is still lacking. 3 Buffer of Thoughts Overview of Buffer of Thoughts In this section, we introduce our Buffer of Thoughts in detail and we also illustrate our core thought-augmented reasoning process in Figure 2. Given a specific task, we utilize our problem-distiller (Section 3.1) to extract critical task-specific information along with relevant constraints. Based on the distilled information, we search in meta-buffer (Section 3.2) that contains a series of high-level thoughts (thought-template) and retrieve a most relevant thought-template for the task. Subsequently, we instantiate the retrieved thought-template with more task-specific reasoning structures and conduct reasoning process. Finally, we employs a buffer-manager (Section 3.3) for summarizing the whole problem-solving process and distilling high-level thoughts for imcreasing the capacity of meta-buffer. <details> <summary>x2.png Details</summary>  ### Visual Description \n ## Diagram: Problem Solving Chain-of-Thought ### Overview This image presents a diagram illustrating a chain-of-thought problem-solving process, likely generated by a large language model. It details the steps taken to solve a word problem concerning pricing and sales volume. The diagram is structured with distinct sections: "Input Problem", "Chain-of-Thought", "Buffer of Thoughts", "Problem Distillation & Thought Retrieval", and "Instantiated Reasoning". There are also sections for "Plan-and-Solve" and a visual "Meta Buffer" element. ### Components/Axes The diagram is not a traditional chart with axes. Instead, it's a flow diagram with text blocks and visual elements. Key components include: * **Input Problem:** A text box containing the original word problem. * **Chain-of-Thought:** A numbered list outlining the steps taken to solve the problem. * **Buffer of Thoughts:** Two columns labeled "Thought Template T₁" and "Thought Template T₂", containing code-like snippets and explanations. * **Problem Distillation & Thought Retrieval:** A section with a visual "X" symbol and text describing the process of distilling the problem. * **Instantiated Reasoning:** A section presenting the solution to the problem, including variable definitions and the final answer. * **Plan-and-Solve:** A section outlining the plan to solve the problem. * **Meta Buffer:** A visual element resembling a radar screen with dots, likely representing the model's internal state. ### Detailed Analysis or Content Details **Input Problem:** "A certain shopping mall sells a batch of branded shirts, with an average daily sales of 20 pieces and a profit of 40 yuan per piece. In order to expand sales, increase profits, and reduce inventory as soon as possible, the mall has decided to take appropriate price reduction measures. After investigation, it was found that for every 1 yuan decrease in the price of this shirt, an average of 2 more shirts are sold per day. If the mall wants to make an average profit of 1200 yuan per day, how much price should each shirt be reduced?" **Chain-of-Thought:** 1. Calculate the current daily profit: Current daily profit = 20 \* 40 = 800 yuan 2. Calculate additional daily profit from selling more shirts: Additional daily profit = 2 \* 40 = 80 yuan 3. Calculate Price reduction: Price reduction = 320 / 2 = 160 yuan Therefore, each shirt should be reduced by 160 yuan in order for the mall to make an average profit of 1200 yuan per day. **Buffer of Thoughts:** * **Thought Template T₁:** * `def process_element(element):` * `# Define how to process each individual element,` * `# e.g., applying a filter or transformation.` * `def combine_elements(element1, element2):` * `# Define how to combine elements, e.g.,` * `# summing numbers, concatenating strings, etc.` * `def check_condition(accumulated_result):` * `# Define the condition that the accumulated` * `# result must meet.` * `def solve_problem(input_list):` * **Thought Template T₂:** * To solve any quadratic equation of the form ax² + bx + c = 0, we can follow a general approach based on the method described. Here is the structured template for solving such equations: * Step 1: Calculate the Discriminant * Compute the discriminant D using the formula D = b² - 4ac. * Step 2: Determine the Nature of the Roots * **Meta Buffer:** Contains several dots, seemingly randomly placed within a circular area. **Problem Distillation & Thought Retrieval:** Contains the "X" symbol and text describing the process of distilling the problem. **Instantiated Reasoning:** * Solution: * `pb` the original profit per shirt (40 yuan). * `x` be the amount by which the price of each shirt is reduced. * `qb` the original quantity of shirts sold daily (20 pieces). * `re` the rate at which the quantity of shirts sold increases as the mall wants to maintain an average daily profit of 1200 yuan. * Equation: (40 - x)(20 + 2x) = 1200 * Expanding: 800 + 80x - 20x - 2x² = 1200 * Simplifying: -2x² + 60x - 400 = 0 * Dividing by -2: x² - 30x + 200 = 0 * Solving the quadratic equation: x = 10 or x = 20 * Therefore, each shirt should be reduced by 10 yuan or 20 yuan. **Plan-and-Solve:** "To devise a plan to solve the problem, we'll break down the process into steps:" 1. Define variables to represent the unknown quantities. 2. Formulate an equation that represents the relationship between the variables and the given information. 3. Solve the equation to find the value of the unknown quantity. 4. Check the solution to ensure it makes sense in the context of the problem. ### Key Observations The diagram showcases a step-by-step reasoning process. The "Buffer of Thoughts" sections suggest the underlying code or templates used by the model to generate the solution. The "Instantiated Reasoning" section provides a clear mathematical solution to the problem. The diagram highlights the model's ability to translate a word problem into a mathematical equation and solve it. There are two possible solutions (x=10 or x=20) presented. ### Interpretation This diagram demonstrates a sophisticated approach to problem-solving, characteristic of advanced language models. The chain-of-thought methodology allows the model to break down a complex problem into smaller, manageable steps. The "Buffer of Thoughts" provides insight into the model's internal reasoning process, revealing the use of templates and code-like structures. The diagram suggests that the model doesn't simply provide an answer but rather explains its reasoning, making the solution more transparent and understandable. The presence of two solutions indicates the model is capable of identifying multiple valid answers, although further context might be needed to determine the most appropriate one. The "Meta Buffer" is a less clear element, potentially representing the model's confidence or internal state during the reasoning process. Overall, the diagram illustrates a powerful combination of natural language processing and mathematical reasoning. </details> Figure 2: Illustration of different reasoning process. Buffer of Thoughts enables large language models to tackle complex reasoning tasks through our thought-augmented reasoning process. Thought template is marked in orange and instantiated thought is marked in blue. 3.1 Problem Distiller Most of complex tasks contain implicit constraints, complex object relationships, and intricate variables and parameters within their contexts. Consequently, during the reasoning stage, LLMs need to overcome three main challenges: extracting vital information, recognizing potential constraints, and performing accurate reasoning. These challenges would impose a significant burden on a single LLM. Therefore, we separate the extraction and comprehension stages of task information from the final reasoning stage, through prepending a problem distiller to the reasoning process. More concretely, we design a meta prompt $\mathcal{\phi}$ to first distill and formalize the task information. The distilled task information could be denoted as: $$ x_{d}=LLM(\mathcal{\phi}(x)), \tag{1} $$ where $x$ is the task statement. Due to the page limit, we put the detailed meta prompt for problem-distiller in Section A.2. Problem Condensation and Translation We use the problem distiller to extract key elements from input tasks, focusing on: (1). Essential parameters and variables for problem-solving; (2). The objectives of the input tasks and their corresponding constraints. We then re-organize this distilled information into a clear, comprehensible format for the subsequent reasoning stage. We then translate the specific problems into high-level concepts and structures. This translation procedure decomposes complex real-world problems, like intricate mathematical application scenarios, into simpler, multi-step calculations, making it easier for later retrieval of high-level thought. 3.2 Thought-Augmented Reasoning with Meta Buffer Motivation Human often summarize and induce higher-level guidelines when solving problems and then apply them to relevant problems. Motivated by this, we propose meta-buffer, a lightweight library that contains a series of high-level thoughts (thought-template) for addressing various types of problems. Unlike traditional methods [11, 46, 12, 36, 9] that require specific instructions or exemplars, our high-level thought-templates can be adaptively instantiated when solving different problems, thereby enhancing LLMs with superior precision and flexibility. Thought Template As a kind of high-level guideline, our thought-template is stored in meta-buffer , and is obtained from various problem-solving processes by our buffer-manager. The details about acquiring thought-templates would be introduced in Section 3.3. Since our BoT aims to provide a general reasoning approach for various tasks, we correspondingly classify the thought-templates into six categories: Text Comprehension, Creative Language Generation, Common Sense Reasoning, Mathematical Reasoning, Code Programming and Application Scheduling. We provide some example thought-templates in Section A.1. Such classification of thought-templates can facilitate the template retrieval for finding most suitable solutions to different problems. Here we denote thought template, template description and its corresponding category as $(T_{i},D_{T_{i}},C_{k})$ , where $i$ denotes the index of meta-template, $k∈\mathbb{Z^{+}}$ and $1≤ k≤ 6$ , which means $C_{k}$ is in one of the six categories, and $D_{T_{i}}$ is the description of thought template. Template Retrieval For each task, our BoT retrieves a thought-template $T_{i}$ that is highly similar to the distilled problem $x_{d}$ by calculating the embedding similarity between the description $D_{T_{i}}$ and $x_{d}$ . The retrieval process can be formulated as: $$ j=\text{argmax}_{i}(\text{Sim}(f(x_{d}),\{f(D_{T_{i}})\}_{i=1}^{N})),\quad% \text{where}\quad\text{Sim}(f({x}_{d}),\{f(D_{T_{i}})\}_{i=0}^{n})>=\delta, \tag{2} $$ $N$ is the size of the meta-buffer, $f(·)$ is a normal text embedding model, and $T_{j}$ denotes the retrieved thought template. We set a threshold $\delta$ (0.5 $\sim$ 0.7 is recommended) to determine whether the current task is new. Therefore, if $\text{Sim}(f({x}_{d}),\{f(D_{T_{i}})\}_{i=0}^{n})<\delta$ , we identify the task $x$ as a new task. Instantiated Reasoning For each specific task, we discuss two situations for the instantiated reasoning, depending on whether the current task is new: The first situation is that we successfully retrieve a thought-template $T_{j}$ for the task. In this case, as presented in Figure 2, our thought-augmented reasoning will be adaptively instantiated to suitable reasoning structures with our designed instantiation prompt (in Section A.3). For example, in a Checkmate-in-One problem, we instantiate the template of updating chess board state to solve the problem step by step. Thus we conduct the instantiated reasoning for task $x$ using the distilled information $x_{d}$ and the retrieved template $T_{j}$ , and produce its solution $S_{x}$ as: $$ S_{x}=LLM_{\text{instantiation}}(x_{d},T_{j}), \tag{3} $$ where $LLM_{\text{instantiation}}$ denotes the instantiated reasoner with a LLM. In the second situation, the task is identified as a new task. To enable proper instantiated reasoning, we prepare three general coarse-grained thought-templates for utilization. Based on the distilled task information $x_{d}$ , our BoT would automatically assign a suitable thought-template to the reasoning process. The detailed pre-defined thought-templates are included in Section A.3). 3.3 Buffer Manager We propose buffer-manager to summarize the high-level guidelines and thoughts that are gained from each problem-solving process. It can generalize each specific solution to more problems, storing the critical distilled knowledge in the form of thought-templates within the meta buffer. In contrast to methods that temporarily generate exemplars or instructions for each problem, our buffer-manager can ensure permanent advancements in accuracy, efficiency, and robustness for LLM-based reasoning. Template Distillation To extract a general though-template, we propose a three-step approach: (1) Core task summarization: identifying and describing basic types and core challenges of problems; (2) Solution steps description: summarize the general steps for solving a problem; (3) General answering template: based on the above analysis, propose a solution template or approach that can be widely applied to similar problems. Additionally, to boost the generalization ability and stability of template distillation, we carefully design two types of in-context examples of how to generate thought-template— in-task and cross-task examples. Cross-task means we choose the template distilled from one task to tackle the problem of other tasks, such as addressing a mathematical problem with a code-related thought-template. The new template distilled from input task $x$ can be denoted as: $$ T_{new}=LLM_{\text{distill}}(x_{d},S_{x}), \tag{4} $$ where $LLM_{\text{distill}}$ is the LLM-based template distiller initialized with the following prompt: Prompt for Template Distillation: User: [Problem Description] + [Solution Steps or Code] To extract and summarize the high-level paradigms and general approaches for solving such problems, please follow these steps in your response: 1. Core task summarization: Identify and describe the basic type and core challenges of the problem, such as classifying it as a mathematical problem (e.g., solving a quadratic equation), a data structure problem (e.g., array sorting), an algorithm problem (e.g., search algorithms), etc. And analyze the most efficient way to solve the problem. 2. Solution Steps Description: Outline the general solution steps, including how to define the problem, determine variables, list key equations or constraints, choose appropriate solving strategies and methods, and how to verify the correctness of the results. 3. General Answer Template: Based on the above analysis, propose a template or approach that can be widely applied to this type of problem, including possible variables, functions, class definitions, etc. If it is a programming problem, provide a set of base classes and interfaces that can be used to construct solutions to specific problems. Please ensure that your response is highly concise and structured, so that specific solutions can be transformed into generalizable methods. [Optional] Here are some exemplars of the thought-template: (Choose cross-task or in-task exemplars based on the analysis of the Core task summarization.) Dynamic Update of Meta-Buffer After template distillation, we need to consider whether the distilled template should be updated into the meta-buffer. If we initialize an empty meta-buffer or encounter a problem without a proper thought-template, the distilled thought-templates will be directly stored in the meta-buffer. If we solve problem with a retrieved thought-template, new insights may arise during the instantiation of a certain thought-template. Therefore, to avoid the redundancy of the meta-buffer while maintaining newly-generated informative thoughts, we will calculate the similarity between the embedding vectors of $D_{T_{new}}$ and $\{D_{T_{i}}\}_{i=0}^{n}$ and update the meta-buffer with the following rule: $$ \text{Max}(\text{Sim}(f(D_{T_{new}}),\{f(D_{T_{i}})\}_{i=0}^{n}))<\delta. \tag{5} $$ Otherwise, it means the meta-buffer has already possessed the necessary knowledge to solve this task and does not need to perform the update. Our dynamic update strategy effectively reduces the computational burden of template retrieval while ensuring the lightweight property of our meta-buffer. We further conduct ablation study to analyze it in Section 6. 4 Experiments Datasets and Tasks To evaluate the efficacy of our proposed Buffer of Thoughts and compare with previous methods, we consider a diverse set of tasks and datasets that require varying degrees of mathematical and algorithmic reasoning, domain-specific knowledge, and literary creativity: (a). The Game of 24 from ToT [14], where the objective is to form an arithmetic expression that equals 24 using each of four given numbers exactly once; (b). Three BIG-Bench Hard (BBH) [35] tasks: Geometric Shapes, Multi-Step Arithmetic Two, and Word Sorting; (c). Three reasoning tasks directly obtained from the BIG-Bench suite [50]: Checkmate-in-One, Penguins —where the task is to answer questions about penguins’ attributes based on a given table and additional natural language information, and DateUnderstanding —a task that involves inferring dates from natural language descriptions, performing arithmetic operations on dates, and utilizing global knowledge such as the number of days in February; (d). Python Programming Puzzles (P3) [51, 52], a collection of challenging programming puzzles written in Python with varying difficulty levels; (e). Multilingual Grade School Math (MGSM) [33], a multilingual version of the GSM8K dataset [53] featuring translations of a subset of examples into ten typologically diverse languages, including Bengali, Japanese, and Swahili; (f). Shakespearean Sonnet Writing from meta-prompting [15], a novel task where the goal is to write a sonnet following the strict rhyme scheme "ABAB CDCD EFEF GG" and incorporating three provided words verbatim. Implementation and Baselines For the fair comparisons with previous methods, we use GPT-4 as the base model of our BoT, including the main experiment and the ablation study (in Section 6). We also use Llama3-8B and Llama3-70B in our analysis part on NVIDIA A100-PCIE-40GB GPU. We compare our Buffer of Thoughts with the following prompting methods: 1. Standard Prompting: This is our most basic baseline, where an LLM is asked to generate a response directly from the input query, without any specific guiding input-output examples or additional instructions beyond the task description included in the query. 2. Single-query Method: This includes Zero-shot CoT [8] and PAL [10], which use the LLM to analyze natural language problems and generate intermediate reasoning steps. We also include Expert Prompting [9], which creates an expert identity tailored to the specific context of the input query, and then integrates this expert profile into the input to generate a well-informed response. 3. Multi-query Method: This includes ToT [14] and GoT [17], which enable LLMs to make deliberate decisions by considering multiple reasoning paths and self-evaluating choices to determine the next course of action. These methods also allow for looking ahead or backtracking when necessary to make global decisions. Additionally, we include Meta Prompting [15], which employs an effective scaffolding technique designed to enhance the functionality of LLMs. Table 1: Comparing BoT with previous methods across various tasks. We denote the best score in blue, and the second-best score in green. Our BoT significantly outperforms other methods on all tasks, especially on general reasoning problems. | Game of 24 MGSM (avg) Multi-Step Arithmetic | 3.0 84.4 84.0 | 11.0 85.5 83.2 | 3.0 85.0 83.2 | 64.0 72.0 87.4 | 74.0 86.4 88.2 | 73.2 87.0 89.2 | 67.0 84.8 90.0 | 82.4 89.2 99.8 | | --- | --- | --- | --- | --- | --- | --- | --- | --- | | WordSorting | 80.4 | 83.6 | 85.2 | 93.2 | 96.4 | 98.4 | 99.6 | 100.0 | | Python Puzzles | 31.1 | 36.3 | 33.8 | 47.3 | 43.5 | 41.9 | 45.8 | 52.4 | | Geometric Shapes | 52.6 | 69.2 | 55.2 | 51.2 | 56.8 | 54.2 | 78.2 | 93.6 | | Checkmate-in-One | 36.4 | 32.8 | 39. 6 | 10.8 | 49.2 | 51.4 | 57.2 | 86.4 | | Date Understanding | 68.4 | 69.6 | 68.4 | 76.2 | 78.6 | 77.4 | 79.2 | 88.2 | | Penguins | 71.1 | 73.6 | 75.8 | 93.3 | 84.2 | 85.4 | 88.6 | 94.7 | | Sonnet Writing | 62.0 | 71.2 | 74.0 | 36.2 | 68.4 | 62.8 | 79.6 | 80.0 | 4.1 BoT Achieves Better Accuracy, Efficiency and Robustness Reasoning Accuracy As shown in LABEL:tab-accuracy, our BoT consistently outperforms all previous prompting methods across multiple kinds of challenging benchmarks, particularly demonstrated in complicated reasoning tasks such as Game of 24 and Checkmate-in-One. Taking GPT-4 as a baseline, our method achieves an astonishing 79.4% accuracy improvement in Game of 24, and compared to ToT, which has a good performance on this task, we also achieve an 8.4% accuracy improvement. What’s more, compared to recent Meta-prompting method [15], we see significant accuracy improvements: 23% on Game of 24, 20% on Geometric Shapes and 51% on Checkmate-in-One. Existing methods need complex, iterative, and heuristic search strategies to address these problems on a case-by-case basis. Conversely, our BoT leverages the historical insights and informative guidelines from thought-templates, and further adaptively instantiate a more optimal reasoning structure for addressing these complex problems. <details> <summary>x3.png Details</summary>  ### Visual Description \n ## Bar Chart: Comparison of the inference time ### Overview This bar chart compares the inference time (in logarithmic seconds) of five different methods: Expert, PAL, ToT, Meta-prompting, and Ours, across three different tasks: Game of 24, MGSM, and Checkmate-in-One. The inference time is represented on the y-axis, while the tasks are displayed on the x-axis. Each task has five bars, one for each method, showing the time taken for inference. ### Components/Axes * **Title:** "Comparison of the inference time" (centered at the top) * **X-axis Label:** Tasks (Game of 24, MGSM, Checkmate-in-One) * **Y-axis Label:** "Logarithmic time (s)" (ranging from 0 to 10) * **Legend:** Located at the top-right corner, identifying the colors for each method: * Expert (Blue) * PAL (Orange) * ToT (Gray) * Meta-prompting (Yellow) * Ours (Light Blue) ### Detailed Analysis The chart consists of three groups of five bars, one group for each task. **Game of 24:** * Expert: Approximately 4.64 seconds. * PAL: Approximately 5.5 seconds. * ToT: Approximately 8.73 seconds. * Meta-prompting: Approximately 8.47 seconds. * Ours: Not present. **MGSM:** * Expert: Approximately 5.17 seconds. * PAL: Approximately 4.16 seconds. * ToT: Approximately 8.34 seconds. * Meta-prompting: Approximately 8.04 seconds. * Ours: Approximately 5 seconds. **Checkmate-in-One:** * Expert: Approximately 4.81 seconds. * PAL: Approximately 5.21 seconds. * ToT: Approximately 9.03 seconds. * Meta-prompting: Approximately 8.43 seconds. * Ours: Approximately 6.39 seconds. ### Key Observations * **ToT consistently exhibits the longest inference times** across all three tasks, significantly exceeding the other methods. * **PAL generally has the shortest inference time** for MGSM and is competitive for Game of 24 and Checkmate-in-One. * **Expert consistently performs well**, with relatively low inference times across all tasks. * **Meta-prompting and Ours have similar performance**, generally falling between Expert and ToT. * The differences in inference times are more pronounced for the Checkmate-in-One task. ### Interpretation The data suggests that the "ToT" method is the most computationally expensive, requiring significantly more time for inference compared to the other methods across all tested tasks. The "PAL" method appears to be the most efficient, particularly for the MGSM task. The "Expert" method provides a good balance between performance and efficiency. The "Ours" method shows competitive performance, generally comparable to "Meta-prompting". The larger differences observed in the "Checkmate-in-One" task might indicate that this task is more sensitive to the choice of inference method, or that the limitations of the "ToT" method are more pronounced for more complex tasks. The logarithmic scale on the y-axis emphasizes the relative differences in inference times, making it easier to compare the performance of the different methods. The chart provides valuable insights into the computational cost of different inference methods for these specific tasks, which can inform the selection of the most appropriate method based on performance requirements and available resources. </details> Figure 3: Comparison of logarithmic inference time between our Buffer of Thoughts and GPT4 [3], GPT4+CoT [8], Expert-prompting [9], PAL [10], ToT [14] across different benchmarks. <details> <summary>x4.png Details</summary>  ### Visual Description \n ## Bar Chart: Success Rate Comparison ### Overview This bar chart compares the success rates of four different models – GPT4, Expert, PAL, and “Ours” – across three distinct game-solving tasks: Game of 24, MGSM, and Checkmate-in-One. An overall average success rate is also presented. The y-axis represents the average accuracy in percentage, ranging from 0 to 100. ### Components/Axes * **Title:** "Success rate" (positioned at the top-center) * **X-axis:** Game/Task Name (labeled: "Game of 24", "MGSM", "Checkmate-in-One", "Average") * **Y-axis:** Average accuracy (%) (labeled, ranging from 0 to 100, with increments of 10) * **Legend:** Located at the top-center, identifying the models by color: * GPT4 (Blue) * Expert (Orange) * PAL (Gray) * Ours (Yellow) ### Detailed Analysis The chart consists of four groups of bars, one for each task/average, with each group containing four bars representing the success rate of each model. **Game of 24:** * GPT4: Approximately 98% (visually, almost reaching 100%) * Expert: Approximately 85% * PAL: Approximately 76% * Ours: Approximately 71% * The trend is a decreasing success rate from GPT4 to "Ours". **MGSM:** * GPT4: Approximately 96.8% (visually, very close to 100%) * Expert: Approximately 87% * PAL: Approximately 84% * Ours: Approximately 84% * GPT4 has the highest success rate, followed by Expert, and PAL and Ours are tied. **Checkmate-in-One:** * GPT4: Approximately 93.4% * Expert: Approximately 78.4% * PAL: Approximately 53.4% * Ours: Approximately 48.2% * The success rate decreases significantly from GPT4 to "Ours". **Average:** * GPT4: Approximately 95.15% * Expert: Approximately 84.57% * PAL: Approximately 70.12% * Ours: Approximately 67.13% * GPT4 has the highest average success rate, followed by Expert, PAL, and "Ours". ### Key Observations * GPT4 consistently outperforms all other models across all tasks and in the overall average. * The "Ours" model generally has the lowest success rate, except for MGSM where it ties with PAL. * The largest performance gap between models is observed in the "Checkmate-in-One" task. * The success rates for all models are relatively high, generally above 60%, indicating a good level of performance overall. ### Interpretation The data suggests that GPT4 is the most effective model for solving these game-solving tasks, demonstrating a significantly higher success rate compared to the other models. The "Expert" model performs reasonably well, consistently ranking second. PAL and "Ours" exhibit lower success rates, with "Ours" generally being the least effective. The substantial difference in performance on the "Checkmate-in-One" task could indicate that this task is particularly challenging and highlights the strengths of GPT4 in handling complex strategic problems. The relatively high success rates across all models suggest that the tasks are not overly difficult, but GPT4's consistent superiority indicates a significant advantage in problem-solving capabilities. The fact that "Ours" ties with PAL on MGSM suggests that the model may have specific strengths in that particular task. The chart provides a clear comparison of the performance of different models, allowing for a quantitative assessment of their effectiveness in game-solving. This information could be valuable for researchers and developers seeking to improve the performance of AI models in similar domains. </details> Figure 4: Comparison of reasoning robustness between our Buffer of Thoughts and GPT4 [3], GPT4+CoT [8], Expert-prompting [9], PAL [10], ToT [14] across different benchmarks. Reasoning Efficiency In addition to significant improvements in accuracy, as a multi-query method, our BoT can achieve comparable reasoning time to single-query method across various tasks, while being considerably less than conventional multi-query method like ToT [14] as shown in Figure 3. For example, in Game of 24, both single-query and multi-query methods necessitate iterative and heuristic searches to identify feasible solutions. This process is particularly time-consuming and inefficient, especially for the multi-query method, which involves conducting multi-query search and backtrace phases. In contrast, our BoT directly retrieves a thought-template in code format, thus a program is instantiated to traverse combinations of numbers and symbols, thereby eliminating the need to build the reasoning structure from scratch. This allows for solving the problem with just one query after invoking the problem-distiller, significantly reducing the time required for complex reasoning. Notably, our BoT requires only 12% of the cost of multi-query methods (e.g., tree of thoughts and meta-prompting) on average. Reasoning Robustness To better evaluate our BoT, we devise a new evaluation metric: success rate, which is used to assess the reasoning robustness. We randomly sample 1000 examples from various benchmarks as a test subset and evaluate different methods on this subset. As shown in Figure 4, we repeat this evaluation process 10 times and take the average accuracy as the success rate of different methods on each benchmark. Compared with other methods, our BoT consistently maintains a higher success rate across various tasks, surpassing the second-best by 10% in average success rate. We attribute our outstanding robustness to the great generalization ability of our distilled thought-templates during reasoning across different tasks. By offering high-level thought from the suitable thought-templates, the stability of our method across different tasks is greatly enhanced. 5 Model Analysis Distribution Analysis of Thought-Templates As depicted in the left figure of Figure 5, we choose six different benchmarks, each sampled with 100 distinct tasks. We update the meta-buffer from scratch, and after completing all sampled tasks, we display the number of derived thought-templates. We can observe that our BoT generates a greater number of thought-templates in the MGSM tasks that contain more diverse scenarios. In tasks with relatively simple requirements, such as Checkmate-in-One and Penguins, BoT produces more fixed thought-templates tailored for those specific issues. The distribution of templates indicates that our BoT can effectively discover appropriate thought templates for different benchmarks. <details> <summary>x5.png Details</summary>  ### Visual Description \n ## Pie Chart: Template Distribution Across Different Tasks ### Overview This image presents a pie chart illustrating the distribution of templates used across various tasks. The chart displays the proportion of each task represented by a slice of the pie, with labels indicating the task name and numerical values representing the count or percentage. ### Components/Axes * **Title:** "Template distribution across different tasks" (positioned at the top-center) * **Legend:** Located at the bottom-right of the chart. It maps colors to task names: * Sonnet writing (Blue) * Table of Penguins (Yellow) * Date understanding (Orange) * MGSM (Red) * Python Puzzles (Light Blue) * Checkmate-in-One (Green) * **Pie Slices:** Each slice represents a task, with the size of the slice proportional to its value. Labels are positioned near each slice, indicating the task name and its corresponding value. ### Detailed Analysis The pie chart shows the following distribution of templates across tasks: * **MGSM:** 78 (Red slice, largest portion of the pie, located at the bottom) * **Python Puzzles:** 37 (Light Blue slice, positioned to the left of MGSM) * **Sonnet writing:** 37 (Blue slice, positioned at the top-left) * **Date understanding:** 14 (Orange slice, positioned to the right of Sonnet writing) * **Table of Penguins:** 8 (Yellow slice, positioned below Date understanding) * **Checkmate-in-One:** 4 (Green slice, positioned between Python Puzzles and Table of Penguins) ### Key Observations * MGSM constitutes the largest portion of the template distribution, accounting for approximately 49% of the total. * Python Puzzles and Sonnet writing have the same count (37), representing approximately 23% each. * Checkmate-in-One has the smallest representation, with only 4 templates used. * The combined proportion of Date understanding, Table of Penguins, and Checkmate-in-One is relatively small compared to MGSM, Python Puzzles, and Sonnet writing. ### Interpretation The data suggests that the MGSM task is the most prevalent, utilizing nearly half of the templates. Python Puzzles and Sonnet writing are also significant tasks, each accounting for a substantial portion of the template usage. The smaller counts for Date understanding, Table of Penguins, and Checkmate-in-One indicate that these tasks are less frequently addressed or require fewer templates. The chart provides insight into the relative importance or frequency of different tasks within the context of template usage. This information could be valuable for resource allocation, task prioritization, or template development efforts. The large disparity between MGSM and other tasks suggests a potential focus area for further investigation – why is MGSM so dominant? Is it a particularly complex task, or is it simply the most common requirement? The relatively low usage of Checkmate-in-One might indicate a niche task or a need for more specialized templates. </details> <details> <summary>x6.png Details</summary>  ### Visual Description \n ## Pie Chart: Average Time Distribution for Each Part of Our BoT ### Overview This image presents a pie chart illustrating the average time distribution across different components of a "BoT" (likely a bot or automated system). The chart visually breaks down the percentage of time spent in each component: problem-distiller, reasoner, meta-buffer, and buffer-manager. ### Components/Axes * **Title:** Average time distribution for each part of our BoT * **Legend:** Located in the bottom-right corner. * problem-distiller (Blue) * reasoner (Green) * meta-buffer (Yellow) * buffer-manager (Red) * **Data Labels:** Displayed directly on each pie slice, indicating the percentage of time allocated to each component. ### Detailed Analysis The pie chart is divided into four segments, each representing a component of the BoT and its corresponding time allocation. * **Reasoner:** This component occupies the largest portion of the pie chart, representing approximately 52.7% of the total time. The segment is colored green. * **Buffer-manager:** This component occupies approximately 21.3% of the total time. The segment is colored red. * **Problem-distiller:** This component occupies approximately 15.6% of the total time. The segment is colored blue. * **Meta-buffer:** This component occupies the smallest portion of the pie chart, representing approximately 8.9% of the total time. The segment is colored yellow. ### Key Observations * The "Reasoner" component dominates the time distribution, consuming over half of the total time. * The "Buffer-manager" component accounts for a significant portion of the time, nearly a quarter. * The "Problem-distiller" and "Meta-buffer" components contribute relatively smaller percentages to the overall time distribution. ### Interpretation The data suggests that the "Reasoner" component is the most computationally intensive or frequently used part of the BoT. This could indicate that the core logic or decision-making process within the BoT relies heavily on this component. The substantial time spent in the "Buffer-manager" suggests that data handling or queuing is a significant aspect of the BoT's operation. The relatively low time allocation to the "Problem-distiller" and "Meta-buffer" might imply that these components are less frequently invoked or have simpler processing requirements. The chart provides insight into the performance bottlenecks or areas for optimization within the BoT. If the goal is to improve the overall efficiency, focusing on optimizing the "Reasoner" component would likely yield the most significant gains. Further investigation into the specific tasks performed by each component could reveal opportunities for streamlining processes and reducing time consumption. </details> Figure 5: Distribution Analysis of Thought-Templates and Time. Left: Distribution Analysis of Thought-Templates. Right: Distribution Analysis of Thought-Templates. Distribution Analysis of Time Cost As illustrated in Figure 5, we measured the average time cost for each component of BoT’s reasoning framework across different tasks. The time required for distilling task information and template retrieval is relatively short, whereas instantiated reasoning takes longer. Overall, considering the complexity of different components, our BoT achieves a relatively balanced distribution of time cost, demonstrating the efficiency of our BoT framework. Better Trade-off between Model Size and Performance As depicted in Figure 6, on Game of 24, word list sorting and Checkmate-in-One, Llama3-8B and Llama-70B models [6] may result in poor outcomes. However, equipped with our BoT, both models demonstrate a substantial accuracy improvement. Notably, BoT+Llama3-8B has the potential to surpass single Llama3-70B model. Our BoT enables smaller models to exhibit the capabilities that approximate or even surpass larger models, significantly bridging the gap between their reasoning abilities. Furthermore, it greatly diminishes the inference cost required by large language models when tackling complex problems. <details> <summary>x7.png Details</summary>  ### Visual Description \n ## Horizontal Bar Chart: Trade-off </details> Figure 6: We evaluate the trade-off between model size and performance with Llama3-8B and Llama3-70B models on three challenging benchmarks. 6 Ablation Study Impact of Problem-Distiller As illustrated in Figure 7, when the problem-distiller is disabled, both Llama3-70B and GPT-4 experience a certain degree of accuracy decline. More complex problems, such as Game of 24 and Checkmate-in-One, show a more significant accuracy reduction, whereas relatively simpler problems like word list sorting and MGSM exhibit smaller decreases. This is because LLMs can more easily extract key information in simpler tasks, making the impact of the problem-distiller less noticeable. In contrast, extracting key information and potential constraints in complex problems is more challenging, making the role of our problem-distiller more prominent, thereby explaining the differences depicted in the figure. <details> <summary>x8.png Details</summary>  ### Visual Description \n ## Bar Chart: Ablation study of problem-distiller ### Overview This bar chart presents a comparative analysis of model performance (Accuracy in percentage) across four different tasks: Game of 24, Word list sorting, Checkmate-in-One, and MGSM. The chart compares the performance of models with and without the "problem-distiller" component. The models tested are BoT+Llama-3-70B and BoT+GPT-4. ### Components/Axes * **Title:** "Ablation study of problem-distiller" (positioned at the top-center) * **X-axis:** Task names: "Game of 24", "Word list sorting", "Checkmate-in-One", "MGSM" (placed at the bottom) * **Y-axis:** Accuracy (%) - Scale ranges from 0 to 100 (placed on the left) * **Legend:** Located at the top of the chart, indicating the data series: * Blue: BoT+Llama-3-70B (w/o problem-distiller) * Orange: BoT+Llama-3-70B (w/ problem-distiller) * Red: BoT+GPT-4 (w/o problem-distiller) * Yellow: BoT+GPT-4 (w/ problem-distiller) ### Detailed Analysis The chart consists of four groups of bars, one for each task. Each group contains four bars representing the accuracy of each model configuration. **Game of 24:** * BoT+Llama-3-70B (w/o problem-distiller): Approximately 71.2% accuracy. * BoT+Llama-3-70B (w/ problem-distiller): Approximately 78.4% accuracy. * BoT+GPT-4 (w/o problem-distiller): Approximately 76.5% accuracy. * BoT+GPT-4 (w/ problem-distiller): Approximately 82.4% accuracy. **Word list sorting:** * BoT+Llama-3-70B (w/o problem-distiller): Approximately 89.5% accuracy. * BoT+Llama-3-70B (w/ problem-distiller): Approximately 92.3% accuracy. * BoT+GPT-4 (w/o problem-distiller): Approximately 97.3% accuracy. * BoT+GPT-4 (w/ problem-distiller): Approximately 99.6% accuracy. **Checkmate-in-One:** * BoT+Llama-3-70B (w/o problem-distiller): Approximately 64.3% accuracy. * BoT+Llama-3-70B (w/ problem-distiller): Approximately 75.6% accuracy. * BoT+GPT-4 (w/o problem-distiller): Approximately 78.9% accuracy. * BoT+GPT-4 (w/ problem-distiller): Approximately 86.4% accuracy. **MGSM:** * BoT+Llama-3-70B (w/o problem-distiller): Approximately 85.6% accuracy. * BoT+Llama-3-70B (w/ problem-distiller): Approximately 86.8% accuracy. * BoT+GPT-4 (w/o problem-distiller): Approximately 87.4% accuracy. * BoT+GPT-4 (w/ problem-distiller): Approximately 89.2% accuracy. ### Key Observations * The "problem-distiller" consistently improves the accuracy of both BoT+Llama-3-70B and BoT+GPT-4 across all tasks. * BoT+GPT-4 generally outperforms BoT+Llama-3-70B, both with and without the problem-distiller. * The largest performance gains from the problem-distiller are observed in the "Checkmate-in-One" task for BoT+Llama-3-70B (an increase of approximately 11.3 percentage points). * The smallest performance gains from the problem-distiller are observed in the "MGSM" task for BoT+Llama-3-70B (an increase of approximately 1.2 percentage points). ### Interpretation The data strongly suggests that the "problem-distiller" is an effective component for improving the performance of both model architectures (Llama-3-70B and GPT-4) across a variety of reasoning tasks. The consistent improvement across all tasks indicates that the problem-distiller is not task-specific but rather provides a general benefit to the models' reasoning capabilities. The larger gains observed in "Checkmate-in-One" might indicate that this task benefits more from the problem-distiller's ability to refine or structure the problem representation. The fact that GPT-4 consistently achieves higher accuracy, even without the problem-distiller, highlights its superior inherent reasoning abilities. The ablation study demonstrates the value added by the problem-distiller, quantifying its impact on model performance. The chart provides empirical evidence supporting the integration of the problem-distiller into these models to enhance their problem-solving skills. </details> Figure 7: We conduct ablation study on problem-distiller across four benchmarks, employing Llama3-70B and GPT-4 as the base models. Impact of Meta-Buffer As illustrated in Figure 8, when the meta-buffer is disabled, both Llama3-70B and GPT-4 models exhibit a noticeable decline in performance, particularly in benchmarks requiring complex reasoning, such as Game of 24 and Checkmate-in-One. This further underscores the superiority of our meta-buffer in addressing complex problems. <details> <summary>x9.png Details</summary>  ### Visual Description \n ## Bar Chart: Ablation study of meta-buffer ### Overview This bar chart presents a comparative analysis of the accuracy of different language model configurations on four distinct tasks: Game of 24, Word list sorting, Checkmate-in-One, and MGSM. The configurations include combinations of BoT (likely "Blend of Thoughts") with Llama-3-70B and GPT-4, both with and without a "meta-buffer." The chart aims to demonstrate the impact of the meta-buffer on the performance of these models. ### Components/Axes * **X-axis:** Represents the four tasks: "Game of 24", "Word list sorting", "Checkmate-in-One", and "MGSM". * **Y-axis:** Represents "Accuracy (%)", ranging from 0 to 100. * **Legend:** Located at the top-left corner, defines the four data series: * Blue: BoT + Llama-3-70B (w/o meta-buffer) * Red: BoT + Llama-3-70B (w/ meta-buffer) * Orange: BoT + GPT-4 (w/o meta-buffer) * Yellow: BoT + GPT-4 (w/ meta-buffer) ### Detailed Analysis The chart consists of four groups of bars, one for each task. Within each group, there are four bars representing the accuracy of each model configuration. **Game of 24:** * BoT + Llama-3-70B (w/o meta-buffer): Approximately 65.6% accuracy. * BoT + Llama-3-70B (w/ meta-buffer): Approximately 78.4% accuracy. * BoT + GPT-4 (w/o meta-buffer): Approximately 75.2% accuracy. * BoT + GPT-4 (w/ meta-buffer): Approximately 82.4% accuracy. **Word list sorting:** * BoT + Llama-3-70B (w/o meta-buffer): Approximately 81.7% accuracy. * BoT + Llama-3-70B (w/ meta-buffer): Approximately 92.3% accuracy. * BoT + GPT-4 (w/o meta-buffer): Approximately 95.4% accuracy. * BoT + GPT-4 (w/ meta-buffer): Approximately 99.6% accuracy. **Checkmate-in-One:** * BoT + Llama-3-70B (w/o meta-buffer): Approximately 27.4% accuracy. * BoT + Llama-3-70B (w/ meta-buffer): Approximately 56.7% accuracy. * BoT + GPT-4 (w/o meta-buffer): Approximately 75.6% accuracy. * BoT + GPT-4 (w/ meta-buffer): Approximately 86.4% accuracy. **MGSM:** * BoT + Llama-3-70B (w/o meta-buffer): Approximately 79.6% accuracy. * BoT + Llama-3-70B (w/ meta-buffer): Approximately 86.8% accuracy. * BoT + GPT-4 (w/o meta-buffer): Approximately 85.4% accuracy. * BoT + GPT-4 (w/ meta-buffer): Approximately 89.2% accuracy. ### Key Observations * The meta-buffer consistently improves the accuracy of both Llama-3-70B and GPT-4 across all four tasks. * GPT-4 generally outperforms Llama-3-70B, regardless of the presence of the meta-buffer. * The largest performance gains from the meta-buffer are observed in the "Word list sorting" and "Checkmate-in-One" tasks. * The "Checkmate-in-One" task has the lowest overall accuracy scores, indicating it is the most challenging task for these models. ### Interpretation The data strongly suggests that the meta-buffer is a beneficial component for improving the accuracy of these language models. The consistent performance gains across all tasks indicate that the meta-buffer provides a generalizable improvement, rather than being specific to a particular task. The larger gains observed in "Word list sorting" and "Checkmate-in-One" might indicate that these tasks benefit more from the additional contextual information or reasoning capabilities provided by the meta-buffer. The superior performance of GPT-4 suggests that larger and more capable models are better able to leverage the benefits of the meta-buffer. The relatively low accuracy on "Checkmate-in-One" could be due to the complexity of chess-related reasoning, or the limitations of the models in handling such specialized tasks. The chart provides empirical evidence supporting the integration of a meta-buffer into language model architectures to enhance their performance on a variety of tasks. </details> Figure 8: We conduct ablation study on meta-buffer across four benchmarks, employing Llama3-70B and GPT-4 as the base models. <details> <summary>x10.png Details</summary>  ### Visual Description \n ## Line Chart: Ablation study of buffer-manager -- Accuracy ### Overview This line chart presents the accuracy results of an ablation study comparing two configurations: "BoT+GPT4" and "BoT+GPT4 (w/o buffer-manager)" across four rounds. Accuracy is measured in percentage (%) and plotted against the round number. ### Components/Axes * **Title:** "Ablation study of buffer-manager -- Accuracy" (Top-center) * **X-axis:** "Round" with markers at Round 1, Round 2, Round 3, and Round 4. (Bottom-center) * **Y-axis:** "Accuracy (%)" with a scale ranging from 0 to 100, incrementing by 10. (Left-side) * **Legend:** Located at the bottom-right corner. * Blue Line: "BoT+GPT4" * Orange Line: "BoT+GPT4 (w/o buffer-manager)" ### Detailed Analysis * **BoT+GPT4 (Blue Line):** The blue line shows an upward trend, indicating increasing accuracy with each round. * Round 1: Approximately 52.8% * Round 2: Approximately 78.5% * Round 3: Approximately 87.4% * Round 4: Approximately 88.5% * **BoT+GPT4 (w/o buffer-manager) (Orange Line):** The orange line shows a fluctuating trend, with an initial increase followed by a decrease. * Round 1: Approximately 56.8% * Round 2: Approximately 53.6% * Round 3: Approximately 57.4% * Round 4: Approximately 54.1% ### Key Observations * The "BoT+GPT4" configuration consistently outperforms the "BoT+GPT4 (w/o buffer-manager)" configuration across all rounds. * The "BoT+GPT4" configuration demonstrates significant accuracy gains from Round 1 to Round 2, with smaller improvements in subsequent rounds. * The "BoT+GPT4 (w/o buffer-manager)" configuration shows a slight initial increase in accuracy from Round 1 to Round 3, but then declines in Round 4. * The difference in accuracy between the two configurations widens with each round, suggesting the buffer-manager becomes increasingly important as the process continues. ### Interpretation The data strongly suggests that the buffer-manager component significantly improves the accuracy of the "BoT+GPT4" system. The ablation study clearly demonstrates that removing the buffer-manager results in lower and more unstable accuracy scores. The initial performance of the configuration without the buffer-manager is slightly higher, but this advantage quickly diminishes as the rounds progress. This could indicate that the buffer-manager is crucial for maintaining consistency and preventing performance degradation over time. The rapid increase in accuracy for the "BoT+GPT4" configuration between Round 1 and Round 2 suggests that the buffer-manager has a substantial impact early in the process, potentially by stabilizing initial conditions or improving data handling. The diminishing returns in later rounds could be due to the system approaching its maximum achievable accuracy, or the buffer-manager's impact becoming less pronounced as the system converges. The consistent divergence between the two lines highlights the value of the buffer-manager as a critical component of the system. </details> Figure 9: We conduct ablation study on buffer-manager regarding reasoning accuracy across four tasks, employing Llama3-70B and GPT-4 as the base models. Impact of Buffer-Manager In this ablation study, we divide the entire process into four rounds. In each round, we randomly sample 50 questions from each benchmark and conduct reasoning. In the subsequent round, we continue to randomly sample another 50 questions from each benchmark. As depicted in Figure 9, with the increase of the number of rounds, the model with the buffer-manager continually expands the meta-buffer while also utilizing the thought-templates obtained from previously solved problems to help addressing subsequent similar problems. Therefore, we can observe that the accuracy of BoT steadily improves with each round. In contrast, the model without the buffer-manager fails to exhibit an upward trend. Additionally, we have also measured the reasoning time as depicted in Figure 10. when the number of rounds increases, the model with the buffer-manager will experience a continual improvement in reasoning efficiency. This is because, with the continual expansion of the meta-buffer, the likelihood of retrieving suitable thought-templates also increases. Consequently, models can avoid constructing reasoning structures from scratch, thereby enhancing the inference efficiency accordingly. <details> <summary>x11.png Details</summary>  ### Visual Description \n ## Line Chart: Ablation study of buffer-manager -- Time ### Overview This line chart presents the results of an ablation study examining the impact of a "buffer-manager" on inference time per problem. It compares two configurations: one *with* the buffer-manager (BoT+GPT4, represented by a blue line) and one *without* the buffer-manager (BoT+GPT4 (w/o buffer-manager), represented by an orange line) across four rounds. The y-axis represents the average inference time in seconds, while the x-axis represents the round number. ### Components/Axes * **Title:** "Ablation study of buffer-manager -- Time" (centered at the top) * **X-axis Label:** "Round" (centered at the bottom) * **Markers:** Round 1, Round 2, Round 3, Round 4 (equally spaced) * **Y-axis Label:** "Average inference time per problem (s)" (left side, vertical) * **Scale:** 0 to 350 seconds, with increments of 50. * **Legend:** Located at the bottom-center of the chart. * **BoT+GPT4:** Blue line * **BoT+GPT4 (w/o buffer-manager):** Orange line ### Detailed Analysis **BoT+GPT4 (Blue Line):** The blue line representing BoT+GPT4 exhibits a strong downward trend. It starts at approximately 297 seconds in Round 1 and steadily decreases to approximately 78.5 seconds in Round 4. * Round 1: ~297 seconds * Round 2: ~205 seconds * Round 3: ~128 seconds * Round 4: ~78.5 seconds **BoT+GPT4 (w/o buffer-manager) (Orange Line):** The orange line representing BoT+GPT4 without the buffer-manager shows a relatively flat trend, with slight fluctuations. It begins at approximately 308 seconds in Round 1 and ends at approximately 306 seconds in Round 4. * Round 1: ~308 seconds * Round 2: ~317 seconds * Round 3: ~304 seconds * Round 4: ~306 seconds ### Key Observations * The inference time for BoT+GPT4 *decreases significantly* across the four rounds, indicating improved performance with each round. * The inference time for BoT+GPT4 without the buffer-manager remains relatively constant across the four rounds. * In all four rounds, the inference time for BoT+GPT4 (with buffer-manager) is *lower* than that of BoT+GPT4 (without buffer-manager). * The difference in inference time between the two configurations is most pronounced in Round 1 (approximately 11 seconds) and Round 4 (approximately 227.5 seconds). ### Interpretation The data strongly suggests that the buffer-manager significantly improves the inference time of BoT+GPT4. The consistent decrease in inference time for the configuration *with* the buffer-manager indicates that it becomes more effective as the process iterates through rounds. The relatively stable inference time for the configuration *without* the buffer-manager suggests that it does not benefit from the iterative process in the same way. The ablation study demonstrates the value of the buffer-manager component. The large difference in inference time, particularly in later rounds, highlights the potential for substantial performance gains by including this feature. The consistent downward trend of the blue line suggests that the buffer-manager is effectively optimizing the process over time, leading to faster inference speeds. The orange line's flatness indicates that the absence of the buffer-manager results in a plateau in performance, unable to capitalize on iterative improvements. </details> Figure 10: We conduct ablation study on buffer-manager regarding reasoning efficiency across four tasks, employing Llama3-70B and GPT-4 as the base models. 7 Discussion Limitations and Future Directions Despite our method’s significant improvement in accuracy while maintaining reasoning efficiency and robustness, our method’s enhancements are limited when addressing problems requiring human-like creativity, as this issue often does not rely on a specific thought-template. Besides, if our BoT initializes the meta-buffer with a weaker model, the quality of the derived thought-templates may be suboptimal due to the weaker model’s limited reasoning ability and instruction-following capability. Overall, our BoT brings out a set of future directions: 1. integrating external resources with BoT to build a open-domain system like agent models [54, 55]. 2. making the distillation of thought-templates optimizable, which may significantly enhance their template qualities for more complex tasks. Conclusion In this work, we introduce Buffer of Thoughts, a novel beffered reasoning framework that employs LLMs to utilize pre-accumulated experiences and methodologies from prior tasks as thought-templates stored within a meta-buffer. We further design buffer-manager to continuously refine the problem-solving processes and dynamically distill thought-templates, thereby progressively raising the LLM’s reasoning capacity. Our BoT demonstrates SOTA performance on 10 challenging tasks, and offers promising prospects for future research and application. References - [1] T. Brown, B. Mann, N. Ryder, M. Subbiah, J. D. Kaplan, P. Dhariwal, A. Neelakantan, P. Shyam, G. Sastry, A. Askell, et al., “Language models are few-shot learners,” Advances in neural information processing systems, vol. 33, pp. 1877–1901, 2020. - [2] R. Anil, A. M. Dai, O. Firat, M. Johnson, D. Lepikhin, A. Passos, S. Shakeri, E. Taropa, P. Bailey, Z. Chen, et al., “Palm 2 technical report,” arXiv preprint arXiv:2305.10403, 2023. - [3] J. Achiam, S. Adler, S. Agarwal, L. Ahmad, I. Akkaya, F. L. Aleman, D. Almeida, J. Altenschmidt, S. Altman, S. Anadkat, et al., “Gpt-4 technical report,” arXiv preprint arXiv:2303.08774, 2023. - [4] Z. Du, Y. Qian, X. Liu, M. Ding, J. Qiu, Z. Yang, and J. Tang, “Glm: General language model pretraining with autoregressive blank infilling,” in Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pp. 320–335, 2022. - [5] A. Q. Jiang, A. Sablayrolles, A. Roux, A. Mensch, B. Savary, C. Bamford, D. S. Chaplot, D. d. l. Casas, E. B. Hanna, F. Bressand, et al., “Mixtral of experts,” arXiv preprint arXiv:2401.04088, 2024. - [6] H. Touvron, T. Lavril, G. Izacard, X. Martinet, M.-A. Lachaux, T. Lacroix, B. Rozière, N. Goyal, E. Hambro, F. Azhar, et al., “Llama: Open and efficient foundation language models,” arXiv preprint arXiv:2302.13971, 2023. - [7] H. Touvron, L. Martin, K. Stone, P. Albert, A. Almahairi, Y. Babaei, N. Bashlykov, S. Batra, P. Bhargava, S. Bhosale, et al., “Llama 2: Open foundation and fine-tuned chat models,” arXiv preprint arXiv:2307.09288, 2023. - [8] J. Wei, X. Wang, D. Schuurmans, M. Bosma, F. Xia, E. Chi, Q. V. Le, D. Zhou, et al., “Chain-of-thought prompting elicits reasoning in large language models,” Advances in neural information processing systems, vol. 35, pp. 24824–24837, 2022. - [9] B. Xu, A. Yang, J. Lin, Q. Wang, C. Zhou, Y. Zhang, and Z. Mao, “Expertprompting: Instructing large language models to be distinguished experts,” arXiv preprint arXiv:2305.14688, 2023. - [10] L. Gao, A. Madaan, S. Zhou, U. Alon, P. Liu, Y. Yang, J. Callan, and G. Neubig, “Pal: Program-aided language models,” in International Conference on Machine Learning, pp. 10764–10799, PMLR, 2023. - [11] X. Wang, J. Wei, D. Schuurmans, Q. V. Le, E. H. Chi, S. Narang, A. Chowdhery, and D. Zhou, “Self-consistency improves chain of thought reasoning in language models,” in The Eleventh International Conference on Learning Representations, 2022. - [12] M. Yasunaga, X. Chen, Y. Li, P. Pasupat, J. Leskovec, P. Liang, E. H. Chi, and D. Zhou, “Large language models as analogical reasoners,” International Conference on Learning Representations, 2024. - [13] Z. Zhang, A. Zhang, M. Li, and A. Smola, “Automatic chain of thought prompting in large language models,” in The Eleventh International Conference on Learning Representations, 2022. - [14] S. Yao, D. Yu, J. Zhao, I. Shafran, T. Griffiths, Y. Cao, and K. Narasimhan, “Tree of thoughts: Deliberate problem solving with large language models,” Advances in Neural Information Processing Systems, vol. 36, 2024. - [15] M. Suzgun and A. T. Kalai, “Meta-prompting: Enhancing language models with task-agnostic scaffolding,” arXiv preprint arXiv:2401.12954, 2024. - [16] D. Zhou, N. Schärli, L. Hou, J. Wei, N. Scales, X. Wang, D. Schuurmans, C. Cui, O. Bousquet, Q. V. Le, et al., “Least-to-most prompting enables complex reasoning in large language models,” in The Eleventh International Conference on Learning Representations, 2022. - [17] M. Besta, N. Blach, A. Kubicek, R. Gerstenberger, M. Podstawski, L. Gianinazzi, J. Gajda, T. Lehmann, H. Niewiadomski, P. Nyczyk, et al., “Graph of thoughts: Solving elaborate problems with large language models,” in Proceedings of the AAAI Conference on Artificial Intelligence, vol. 38, pp. 17682–17690, 2024. - [18] A. Asai, S. Min, Z. Zhong, and D. Chen, “Retrieval-based language models and applications,” in Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 6: Tutorial Abstracts), pp. 41–46, 2023. - [19] G. Mialon, R. Dessi, M. Lomeli, C. Nalmpantis, R. Pasunuru, R. Raileanu, B. Roziere, T. Schick, J. Dwivedi-Yu, A. Celikyilmaz, et al., “Augmented language models: a survey,” Transactions on Machine Learning Research, 2023. - [20] W. Shi, S. Min, M. Yasunaga, M. Seo, R. James, M. Lewis, L. Zettlemoyer, and W.-t. Yih, “Replug: Retrieval-augmented black-box language models,” arXiv preprint arXiv:2301.12652, 2023. - [21] Y. Gao, Y. Xiong, X. Gao, K. Jia, J. Pan, Y. Bi, Y. Dai, J. Sun, and H. Wang, “Retrieval-augmented generation for large language models: A survey,” arXiv preprint arXiv:2312.10997, 2023. - [22] P. Zhao, H. Zhang, Q. Yu, Z. Wang, Y. Geng, F. Fu, L. Yang, W. Zhang, and B. Cui, “Retrieval-augmented generation for ai-generated content: A survey,” arXiv preprint arXiv:2402.19473, 2024. - [23] S. Borgeaud, A. Mensch, J. Hoffmann, T. Cai, E. Rutherford, K. Millican, G. B. Van Den Driessche, J.-B. Lespiau, B. Damoc, A. Clark, et al., “Improving language models by retrieving from trillions of tokens,” in International conference on machine learning, pp. 2206–2240, PMLR, 2022. - [24] M. Yasunaga, A. Aghajanyan, W. Shi, R. James, J. Leskovec, P. Liang, M. Lewis, L. Zettlemoyer, and W.-T. Yih, “Retrieval-augmented multimodal language modeling,” in International Conference on Machine Learning, pp. 39755–39769, PMLR, 2023. - [25] G. Izacard, P. Lewis, M. Lomeli, L. Hosseini, F. Petroni, T. Schick, J. Dwivedi-Yu, A. Joulin, S. Riedel, and E. Grave, “Atlas: Few-shot learning with retrieval augmented language models,” Journal of Machine Learning Research, vol. 24, no. 251, pp. 1–43, 2023. - [26] Z. Wang, W. Nie, Z. Qiao, C. Xiao, R. Baraniuk, and A. Anandkumar, “Retrieval-based controllable molecule generation,” in The Eleventh International Conference on Learning Representations, 2022. - [27] L. Yang, Z. Huang, X. Zhou, M. Xu, W. Zhang, Y. Wang, X. Zheng, W. Yang, R. O. Dror, S. Hong, et al., “Prompt-based 3d molecular diffusion models for structure-based drug design,” 2023. - [28] T. Kojima, S. S. Gu, M. Reid, Y. Matsuo, and Y. Iwasawa, “Large language models are zero-shot reasoners,” Advances in neural information processing systems, vol. 35, pp. 22199–22213, 2022. - [29] O. Press, M. Zhang, S. Min, L. Schmidt, N. A. Smith, and M. Lewis, “Measuring and narrowing the compositionality gap in language models,” in Findings of the Association for Computational Linguistics: EMNLP 2023, pp. 5687–5711, 2023. - [30] S. Arora, A. Narayan, M. F. Chen, L. Orr, N. Guha, K. Bhatia, I. Chami, and C. Re, “Ask me anything: A simple strategy for prompting language models,” in The Eleventh International Conference on Learning Representations, 2022. - [31] T. Khot, H. Trivedi, M. Finlayson, Y. Fu, K. Richardson, P. Clark, and A. Sabharwal, “Decomposed prompting: A modular approach for solving complex tasks,” in The Eleventh International Conference on Learning Representations, 2022. - [32] J. Wei, Y. Tay, R. Bommasani, C. Raffel, B. Zoph, S. Borgeaud, D. Yogatama, M. Bosma, D. Zhou, D. Metzler, et al., “Emergent abilities of large language models,” Transactions on Machine Learning Research, 2022. - [33] F. Shi, M. Suzgun, M. Freitag, X. Wang, S. Srivats, S. Vosoughi, H. W. Chung, Y. Tay, S. Ruder, D. Zhou, et al., “Language models are multilingual chain-of-thought reasoners,” in The Eleventh International Conference on Learning Representations, 2022. - [34] Y. Fu, H. Peng, A. Sabharwal, P. Clark, and T. Khot, “Complexity-based prompting for multi-step reasoning,” in The Eleventh International Conference on Learning Representations, 2022. - [35] M. Suzgun, N. Scales, N. Schärli, S. Gehrmann, Y. Tay, H. W. Chung, A. Chowdhery, Q. Le, E. Chi, D. Zhou, et al., “Challenging big-bench tasks and whether chain-of-thought can solve them,” in Findings of the Association for Computational Linguistics: ACL 2023, pp. 13003–13051, 2023. - [36] H. S. Zheng, S. Mishra, X. Chen, H.-T. Cheng, E. H. Chi, Q. V. Le, and D. Zhou, “Take a step back: Evoking reasoning via abstraction in large language models,” arXiv preprint arXiv:2310.06117, 2023. - [37] P. Zhou, J. Pujara, X. Ren, X. Chen, H.-T. Cheng, Q. V. Le, E. H. Chi, D. Zhou, S. Mishra, and H. S. Zheng, “Self-discover: Large language models self-compose reasoning structures,” arXiv preprint arXiv:2402.03620, 2024. - [38] W. Chen, X. Ma, X. Wang, and W. W. Cohen, “Program of thoughts prompting: Disentangling computation from reasoning for numerical reasoning tasks,” Transactions on Machine Learning Research, 2023. - [39] X. Ning, Z. Lin, Z. Zhou, Z. Wang, H. Yang, and Y. Wang, “Skeleton-of-thought: Large language models can do parallel decoding,” in The Twelfth International Conference on Learning Representations, 2023. - [40] Y. Zhang, “Meta prompting for agi systems,” arXiv preprint arXiv:2311.11482, 2023. - [41] J. Chen, R. Xu, Z. Fu, W. Shi, Z. Li, X. Zhang, C. Sun, L. Li, Y. Xiao, and H. Zhou, “E-kar: A benchmark for rationalizing natural language analogical reasoning,” in Findings of the Association for Computational Linguistics: ACL 2022, pp. 3941–3955, 2022. - [42] O. Sultan and D. Shahaf, “Life is a circus and we are the clowns: Automatically finding analogies between situations and processes,” in Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing, pp. 3547–3562, 2022. - [43] N. Zhang, L. Li, X. Chen, X. Liang, S. Deng, and H. Chen, “Multimodal analogical reasoning over knowledge graphs,” in The Eleventh International Conference on Learning Representations, 2022. - [44] B. Bhavya, J. Xiong, and C. Zhai, “Analogy generation by prompting large language models: A case study of instructgpt,” in Proceedings of the 15th International Conference on Natural Language Generation, pp. 298–312, 2022. - [45] B. Bhavya, J. Xiong, and C. Zhai, “Cam: A large language model-based creative analogy mining framework,” in Proceedings of the ACM Web Conference 2023, pp. 3903–3914, 2023. - [46] Z. Zhang, A. Zhang, M. Li, and A. Smola, “Automatic chain of thought prompting in large language models,” in The Eleventh International Conference on Learning Representations, 2022. - [47] T. Webb, K. J. Holyoak, and H. Lu, “Emergent analogical reasoning in large language models,” Nature Human Behaviour, vol. 7, no. 9, pp. 1526–1541, 2023. - [48] J. Yu, R. He, and Z. Ying, “Thought propagation: An analogical approach to complex reasoning with large language models,” in International Conference on Learning Representations, 2024. - [49] T. Feng, P. Han, G. Lin, G. Liu, and J. You, “Thought-retriever: Don’t just retrieve raw data, retrieve thoughts,” in ICLR 2024 Workshop: How Far Are We From AGI. - [50] B. bench authors, “Beyond the imitation game: Quantifying and extrapolating the capabilities of language models,” Transactions on Machine Learning Research, 2023. - [51] T. Schuster, A. Kalyan, A. Polozov, and A. T. Kalai, “Programming puzzles,” in Thirty-fifth Conference on Neural Information Processing Systems Datasets and Benchmarks Track, 2021. - [52] A. T. K. Patrick Haluptzok, Matthew Bowers, “Language models can teach themselves to program better,” in Eleventh International Conference on Learning Representations (ICLR), 2023. - [53] K. Cobbe, V. Kosaraju, M. Bavarian, M. Chen, H. Jun, L. Kaiser, M. Plappert, J. Tworek, J. Hilton, R. Nakano, C. Hesse, and J. Schulman, “Training verifiers to solve math word problems,” arXiv preprint arXiv:2110.14168, 2021. - [54] G. Chen, S. Dong, Y. Shu, G. Zhang, J. Sesay, B. F. Karlsson, J. Fu, and Y. Shi, “Autoagents: A framework for automatic agent generation,” arXiv preprint arXiv:2309.17288, 2023. - [55] Q. Wu, G. Bansal, J. Zhang, Y. Wu, S. Zhang, E. Zhu, B. Li, L. Jiang, X. Zhang, and C. Wang, “Autogen: Enabling next-gen llm applications via multi-agent conversation framework,” arXiv preprint arXiv:2308.08155, 2023. Appendix A Additional Method Details A.1 Detailed Thought-Templates Here we show six example thought-templates in six different categories: A.1.1 Text Comprehension Task Description: The task involves analyzing a table with various attributes of penguins, such as name, age, height, and weight, and answering questions about these attributes. The table may be updated with new entries, and additional context or comparisons may be provided in natural language. Solution Description: To accurately answer questions about the penguins’ attributes, one must be able to interpret the data presented in tabular form, understand any additional information provided in natural language, and apply logical reasoning to identify the correct attribute based on the question asked. Thought Template: Step 1: Parse the initial table, extracting the header information and each penguin’s attributes into a structured format (e.g., a list of dictionaries). Step 2: Read and integrate any additional natural language information that updates or adds to the table, ensuring the data remains consistent. Step 3: Identify the attribute in question (e.g., oldest penguin, heaviest penguin) and the corresponding column in the table. Step 4: Apply logical reasoning to compare the relevant attribute across all entries to find the correct answer (e.g., the highest age for the oldest penguin). Step 5: Select the answer from the provided options that matches the result of the logical comparison. A.1.2 Creative Language Generation Task Description: The task is to generate a sonnet that adheres to the traditional English sonnet rhyme scheme of "ABAB CDCD EFEF GG" and includes three specific words verbatim in the text. Solution Description: Writing a sonnet involves crafting 14 lines of poetry that follow a specific rhyme pattern. The lines are typically in iambic pentameter, though flexibility in rhythm can be allowed for creative reasons. The given rhyme scheme dictates the end sounds of each line, ensuring a structured poetic form. Incorporating the three provided words verbatim requires strategic placement within the lines to maintain the poem’s coherence and thematic unity. Thought Template: Step 1: Identify the three words that must be included in the sonnet. Step 2: Understand the rhyme scheme "ABAB CDCD EFEF GG" and prepare a list of rhyming words that could be used. Step 3: Develop a theme or story for the sonnet that can naturally incorporate the three provided words. Step 4: Begin drafting the sonnet by writing the first quatrain (four lines) following the "ABAB" rhyme scheme, ensuring one or more of the provided words are included. Step 5: Continue with the second quatrain "CDCD," the third quatrain "EFEF," and finally the closing couplet "GG," each time incorporating the provided words as needed. Step 6: Review the sonnet for coherence, flow, and adherence to the rhyme scheme, making adjustments as necessary. A.1.3 Common Sense Reasoning Task Description: Given a specific date and an event, such as a holiday or historical event, determine the following date. Solution Description: To determine the next date, we need to consider the structure of the calendar, the number of days in each month, and whether it’s a leap year. Typically, the number of days in a month is fixed, except February may vary due to leap years. The next day in a year is usually the date increased by one day unless it’s the end of the month, then the next day will be the first day of the following month. For the end of the year, the next day will be January 1st of the following year. Thought Template: Step 1: Identify the given date’s month and day number. Step 2: Check if it’s the end of the month; if so, confirm the start date of the next month. Step 3: If it’s not the end of the month, simply add one to the day number. Step 4: Pay special attention to the end of the year, ensuring the year increments. A.1.4 Mathematical Reasoning Task Description: Solve an quadratic equation of the form $ax^{2}+bx+c=0$ considering any situations. Solution Description: To solve any quadratic equation of the form $ax^{2}+bx+c=0$ , we can follow a general approach based on the method described. Here is the structured template for solving such equations: Thought Template: Step 1: Calculate the Discriminant - Compute the discriminant $D$ using the formula $D=b^{2}-4ac$ . Step 2: Determine the Nature of the Roots - If $D>0$ , the equation has two distinct real roots. - If $D=0$ , the equation has exactly one real root (also known as a repeated or double root). - If $D<0$ , the equation has two complex roots. Step 3: Compute the Roots - For $D≥ 0$ , calculate the roots using the formula $x=\frac{-b±\sqrt{D}}{2a}$ . - For $D<0$ , calculate the real and imaginary parts of the complex roots using the formula $x=\frac{-b}{2a}±\frac{\sqrt{-D}}{2a}i$ , where $i$ is the imaginary unit. A.1.5 Code Programming Task Description: When given a list of numbers, try to utilize 4 basic mathematical operations (+-*/) to get a target number. Thought Template: Listing 1: Python template ⬇ from itertools import permutations, product def perform_operation (a, b, operation): # Define the operation logic (e.g., addition, subtraction, etc.). pass def evaluate_sequence (sequence, operations): # Apply operations to the sequence and check if the result meets the criteria. pass def generate_combinations (elements, operations): # Generate all possible combinations of elements and operations. pass def format_solution (sequence, operations): # Format the sequence and operations into a human-readable string. pass def find_solution (input_elements, target_result): # Data Input Handling # Validate and preprocess input data if necessary. # Core Algorithm Logic for sequence in permutations (input_elements): for operation_combination in generate_combinations (sequence, operations): try: if evaluate_sequence (sequence, operation_combination) == target_result: # Data Output Formatting return format_solution (sequence, operation_combination) except Exception as e: # Error Handling # Handle specific exceptions that may occur during evaluation. continue # If no solution is found after all iterations, return a default message. # return No solution found message return # Example usage: input_elements = [1, 7, 10, 3] target_result = 24 print (find_solution (input_elements, target_result)) A.1.6 Application Scheduling Task Description: Given some Chess moves in SAN, update the chess board state. Listing 2: Python template ⬇ import chess def find_checkmate_move (moves_san): # Initialize a new chess board board = chess. Board () # Apply the moves to the board for move_san in moves_san: # Remove move numbers and periods (e.g., "1." or "2.") if len (move_san. split (’. ␣ ’)) > 1: move_san = move_san. split (’. ␣ ’)[1] # Skip empty strings resulting from the removal if move_san: # Apply each move in SAN format to the board move = board. parse_san (move_san) board. push (move) # Generate all possible legal moves from the current position for move in board. legal_moves: # Make the move on a copy of the board to test the result board_copy = board. copy () board_copy. push (move) # Check if the move results in a checkmate if board_copy. is_checkmate (): # Return the move that results in checkmate in SAN format return board. san (move) # return No solution found message return #Example usage: input = ’......’ # Check input format and transform the input into legal format # Remove move numbers and periods (e.g., "1." or "2.") checkmate_move = find_checkmate_move (moves_san) print (checkmate_move) A.2 Prompt for Problem Distiller [Problem Distiller]: As a highly professional and intelligent expert in information distillation, you excel at extracting essential information to solve problems from user input queries. You adeptly transform this extracted information into a suitable format based on the respective type of the issue. Please categorize and extract the crucial information required to solve the problem from the user’s input query, the distilled information should include. 1. Key information: Values and information of key variables extracted from user input, which will be handed over to the respective expert for task resolution, ensuring all essential information required to solve the problem is provided. 2. Restrictions: The objective of the problem and corresponding constraints. 3. Distilled task: Extend the problem based on 1 and 2, summarize a meta problem that can address the user query and handle more input and output variations. Incorporate the real-world scenario of the extended problem along with the types of key variables and information constraints from the original problem to restrict the key variables in the extended problem. After that, use the user query input key information as input to solve the problem as an example. A.3 Prompt for Instantiated Reasoning [Meta Reasoner] You are a Meta Reasoner who are extremely knowledgeable in all kinds of fields including Computer Science, Math, Physics, Literature, History, Chemistry, Logical reasoning, Culture, Language….. You are also able to find different high-level thought for different tasks. Here are three reasoning sturctures: i) Prompt-based structure: It has a good performance when dealing with problems like Common Sense Reasoning, Application Scheduling ii) Procedure-based structure It has a good performance when dealing with creative tasks like Creative Language Generation, and Text Comprehension iii) Programming-based: It has a good performance when dealing with Mathematical Reasoning and Code Programming, it can also transform real-world problems into programming problem which could be solved efficiently. (Reasoning instantiation) Your task is: 1. Deliberately consider the context and the problem within the distilled respond from problem distiller and use your understanding of the question within the distilled respond to find a domain expert who are suitable to solve the problem. 2. Consider the distilled information, choose one reasoning structures for the problem. 3. If the thought-template is provided, directly follow the thought-template to instantiate for the given problem.