2312.08935

# MATH-SHEPHERD: VERIFY AND REINFORCE LLMS STEP-BY-STEP WITHOUT HUMAN ANNOTATIONS ## MATH-SHEPHERD: VERIFY AND REINFORCE LLMS STEP-BY-STEP WITHOUT HUMAN ANNOTATIONS Peiyi Wang 1 † Lei Li 3 Zhihong Shao 4 R.X. Xu 2 Damai Dai 1 Yifei Li 5 2 2 1 Deli Chen Y. Wu Zhifang Sui 1 National Key Laboratory for Multimedia Information Processing, Peking University 2 DeepSeek-AI 3 The University of Hong Kong 4 Tsinghua University 5 The Ohio State University { wangpeiyi9979, nlp.lilei } @gmail.com li.14042@osu.edu szf@pku.edu.cn <details> <summary>Image 1 Details</summary>  ### Visual Description Icon/Small Image (46x43) </details> Project Page: MATH-SHEPHERD ## ABSTRACT In this paper, we present an innovative process-oriented math process reward model called MATH-SHEPHERD , which assigns a reward score to each step of math problem solutions. The training of MATH-SHEPHERD is achieved using automatically constructed process-wise supervision data, breaking the bottleneck of heavy reliance on manual annotation in existing work. We explore the effectiveness of MATH-SHEPHERD in two scenarios: 1) Verification : MATH-SHEPHERD is utilized for reranking multiple outputs generated by Large Language Models (LLMs); 2) Reinforcement Learning : MATH-SHEPHERD is employed to reinforce LLMs with step-by-step Proximal Policy Optimization (PPO). With MATH-SHEPHERD, a series of open-source LLMs demonstrates exceptional performance. For instance, the step-by-step PPO with MATH-SHEPHERD significantly improves the accuracy of Mistral-7B (77.9% → 84.1% on GSM8K and 28.6% → 33.0% on MATH). The accuracy can be further enhanced to 89.1% and 43.5% on GSM8K and MATH with the verification of MATH-SHEPHERD, respectively. We believe that automatic process supervision holds significant potential for the future evolution of LLMs. Figure 1: We evaluate the performance of various LLMs with MATH-SHEPHERD on the GSM8K and MATH datasets. All base models are finetuned with the MetaMath dataset (Yu et al., 2023b). The +SHEPHERD results are obtained by selecting the best one from 256 candidates using MATHSHEPHERD. We observe that MATH-SHEPHERD is compatible with different LLMs. The results of GPT-4 (early) are from Bubeck et al. (2023). <details> <summary>Image 2 Details</summary>  ### Visual Description ## Bar Chart: Model Accuracy Comparison on GSM8K Benchmark ### Overview The chart compares the accuracy of different large language models (LLMs) on the GSM8K benchmark, evaluating two approaches: "Fine-tuned LLMs" (blue) and "+SHEPHERD" (orange). The y-axis represents accuracy in percentage, while the x-axis lists five models: LLama2-70B MAmmoTH, LLama2-70B WizardMATH, LLama2-70B MetaMATH*, LLaMA-34B MetaMATH*, and DeepSeek-67B MetaMATH*. Two reference lines are included: "GPT-4 (early): 92.0" (red) and "GPT-4-0613*: 94.4" (gold). ### Components/Axes - **X-axis (Models)**: - LLama2-70B MAmmoTH - LLama2-70B WizardMATH - LLama2-70B MetaMATH* - LLaMA-34B MetaMATH* - DeepSeek-67B MetaMATH* - **Y-axis (Accuracy %)**: Ranges from 70% to 95%, with gridlines at 5% intervals. - **Legend**: - Blue: Fine-tuned LLMs - Orange: +SHEPHERD - **Title**: "GSM8K" (bottom center). ### Detailed Analysis - **LLama2-70B MAmmoTH**: - Fine-tuned LLMs: 72.4% (blue) - +SHEPHERD: 93.2% (orange) - **LLama2-70B WizardMATH**: - Fine-tuned LLMs: 81.6% (blue) - +SHEPHERD: 80.4% (orange) - **LLama2-70B MetaMATH***: - Fine-tuned LLMs: 75.8% (blue) - +SHEPHERD: 90.9% (orange) - **LLaMA-34B MetaMATH***: - Fine-tuned LLMs: 82.8% (blue) - +SHEPHERD: Not explicitly labeled (implied by bar height). - **DeepSeek-67B MetaMATH***: - Fine-tuned LLMs: 82.8% (blue) - +SHEPHERD: 93.3% (orange) ### Key Observations 1. **+SHEPHERD consistently improves accuracy** across all models, with gains ranging from +17.4% (LLama2-70B MAmmoTH) to +7.5% (LLama2-70B WizardMATH). 2. **DeepSeek-67B MetaMATH*** achieves the highest accuracy (93.3%) with +SHEPHERD, surpassing all other models. 3. **LLama2-70B MAmmoTH** shows the largest improvement (+20.8%) when +SHEPHERD is applied. 4. **GPT-4 benchmarks** (92.0% and 94.4%) are not part of the chart but are referenced as external standards. ### Interpretation The data demonstrates that the +SHEPHERD method significantly enhances the performance of LLMs on the GSM8K benchmark, particularly for models with lower baseline accuracy (e.g., LLama2-70B MAmmoTH). This suggests that +SHEPHERD may address limitations in fine-tuned models, such as reasoning gaps or domain-specific knowledge deficits. The consistent improvement across models implies that +SHEPHERD is a robust augmentation technique, though its effectiveness varies slightly depending on the base model. The absence of explicit values for LLaMA-34B MetaMATH* with +SHEPHERD introduces minor uncertainty, but its bar height aligns with the trend of higher accuracy for +SHEPHERD. The GPT-4 benchmarks highlight the gap between current LLMs and state-of-the-art performance, emphasizing the need for further optimization. </details> <details> <summary>Image 3 Details</summary>  ### Visual Description ## Bar Chart: Model Accuracy on MATH Dataset ### Overview The chart compares the accuracy of various large language models (LLMs) on the MATH dataset, with and without the "+SHEPHERD" enhancement. It includes two horizontal reference lines: one at 42.5% labeled "GPT-4 (early)" and another at 56.2% labeled "GPT-4-0613*". ### Components/Axes - **X-axis**: Model names (LLama2-70B MAmmoTH, LLama2-70B WizardMATH, LLama2-70B MetaMATH*, LLeMma-34B MetaMATH*, DeepSeek-67B MetaMATH*). - **Y-axis**: Accuracy (%) ranging from 10% to 60%. - **Legend**: - Blue: "Fine-tuned LLMs" (base accuracy). - Orange: "+SHEPHERD" (additional accuracy from the enhancement). - **Horizontal Lines**: - Red line at 42.5% (GPT-4 early). - Green line at 56.2% (GPT-4-0613*). ### Detailed Analysis - **LLama2-70B MAmmoTH**: - Base accuracy: 21.1% (blue). - +SHEPHERD: 22.7% (orange). - **LLama2-70B WizardMATH**: - Base accuracy: 22.7% (blue). - +SHEPHERD: 29.8% (orange). - **LLama2-70B MetaMATH***: - Base accuracy: 34.8% (blue). - +SHEPHERD: 45.2% (orange). - **LLeMma-34B MetaMATH***: - Base accuracy: 34.8% (blue). - +SHEPHERD: 47.3% (orange). - **DeepSeek-67B MetaMATH***: - Base accuracy: 36.8% (blue). - +SHEPHERD: 48.1% (orange). ### Key Observations 1. **SHEPHERD Enhancement**: All models show improved accuracy when combined with SHEPHERD, with the largest gains in LLama2-70B WizardMATH (+7.1%) and DeepSeek-67B MetaMATH* (+11.3%). 2. **GPT-4 Benchmarks**: - GPT-4 (early) at 42.5% is surpassed by all models with SHEPHERD. - GPT-4-0613* at 56.2% remains the highest accuracy, but only DeepSeek-67B MetaMATH* (+SHEPHERD) approaches this value (48.1%). 3. **Model Performance**: - LLama2-70B MAmmoTH and WizardMATH have the lowest base accuracies but show moderate improvements with SHEPHERD. - LLeMma-34B and DeepSeek-67B MetaMATH* achieve the highest combined accuracies. ### Interpretation The chart demonstrates that the "+SHEPHERD" enhancement significantly boosts the performance of all tested models on the MATH dataset. While GPT-4-0613* remains the top performer, the integration of SHEPHERD with models like DeepSeek-67B MetaMATH* brings their accuracy closer to GPT-4's baseline. This suggests that SHEPHERD is a critical component for improving mathematical reasoning capabilities in LLMs, particularly for models with lower initial performance. The data highlights the importance of hybrid approaches (fine-tuning + external enhancements) in advancing LLM accuracy for complex tasks like mathematical problem-solving. </details> † Contribution during internship at DeepSeek-AI. ## 1 INTRODUCTION Large language models (LLMs) have demonstrated remarkable capabilities across various tasks (Park et al., 2023; Kaddour et al., 2023; Song et al.; Li et al., 2023a; Wang et al., 2023a; Chen et al., 2023; Zheng et al., 2023; Wang et al., 2023c), However, even the most advanced LLMs face challenges in complex multi-step mathematical reasoning problems (Lightman et al., 2023; Huang et al., 2023). To address this issue, prior research has explored different methodologies, such as pretraining (Azerbayev et al., 2023), fine-tuning (Luo et al., 2023; Yu et al., 2023b; Wang et al., 2023b), prompting (Wei et al., 2022; Fu et al., 2022), and verification (Wang et al., 2023d; Li et al., 2023b; Zhu et al., 2023; Leviathan et al., 2023). Among these techniques, verification has recently emerged as a favored method. The motivation behind verification is that relying solely on the top-1 result may not always produce reliable outcomes. A verification model can rerank candidate responses, ensuring higher accuracy and consistency in the outputs of LLMs. In addition, a good verification model can also offer invaluable feedback for further improvement of LLMs (Uesato et al., 2022; Wang et al., 2023b; Pan et al., 2023). The verification models generally fall into the outcome reward model (ORM) (Cobbe et al., 2021; Yu et al., 2023a) and process reward model (PRM) (Li et al., 2023b; Uesato et al., 2022; Lightman et al., 2023; Ma et al., 2023). The ORM assigns a confidence score based on the entire generation sequence, whereas the PRM evaluates the reasoning path step-by-step. PRM is advantageous due to several compelling reasons. One major benefit is its ability to offer precise feedback by identifying the specific location of any errors that may arise, which is a valuable signal in reinforcement learning and automatic correction. Besides, The PRM exhibits similarities to human behavior when assessing a reasoning problem. If any steps contain an error, the final result is more likely to be incorrect, mirroring the way human judgment works. However, gathering data to train a PRM can be an arduous process. Uesato et al. (2022) and Lightman et al. (2023) utilize human annotators to provide process supervision annotations, enhancing the performance of PRM. Nevertheless, annotation by humans, particularly for intricate multi-step reasoning tasks that require advanced annotator skills, can be quite costly, which hinders the advancement and practical application of PRM. To tackle the problem, in this paper, we propose an automatic process annotation framework. Inspired by Monte Carlo Tree Search (Kocsis & Szepesv´ ari, 2006; Coulom, 2006; Silver et al., 2016; ´ Swiechowski et al., 2023), we define the quality of an intermediate step as its potential to deduce the correct final answer. By leveraging the correctness of the answer, we can automatically gather step-wise supervision. Specifically, given a math problem with a golden answer and a step-by-step solution, to achieve the label of a specific step, we utilize a fine-tuned LLM to decode multiple subsequent reasoning paths from this step. We further validate whether the decoded final answer matches with the golden answer. If a reasoning step can deduce more correct answers than another, it would be assigned a higher correctness score. We use this automatic way to construct the training data for MATH-SHEPHERD, and verify our ideas on two widely used mathematical benchmarks, GSM8K (Cobbe et al., 2021) and MATH (Hendrycks et al., 2021). We explore the effectiveness of MATH-SHEPHERD in two scenarios: 1) verification: MATH-SHEPHERD is utilized for reranking multiple outputs generated by LLMs; 2) reinforcement learning: MATH-SHEPHERD is employed to reinforce LLMs with step-by-step Proximal Policy Optimization (PPO). With the verification of MATH-SHEPHERD, a series of open-source LLMs from 7B to 70B demonstrates exceptional performance. For instance, the step-by-step PPO with MATHSHEPHERD significantly improves the accuracy of Mistral-7B (77.9% → 84.1% on GSM8K and 28.6% → 33.0% on MATH). The accuracy can be further enhanced to 89.1% and 43.5% on GSM8K and MATH with verification. DeepSeek 67B (DeepSeek, 2023) achieves accuracy rates of 93.3% on the GSM8K dataset and 48.1% on the MATH dataset with verification of MATH-SHEPHERD. To the best of our knowledge, these results are unprecedented for open-source models that do not rely on additional tools. Our main contributions are as follows: 1) We propose a framework to automatically construct process supervision datasets without human annotations for math reasoning tasks. - 2) We evaluate our method on both step-by-step verification and reinforcement learning scenarios. Extensive experiments on two widely used mathematical benchmarks - GSM8K and MATH, in addition to a series of LLMs ranging from 7B to 70B, demonstrate the effectiveness of our method. - 3) We empirically analyze the key factors for training high-performing process reward models, shedding light on future directions toward improving reasoning capability with automatic step-bystep verification and supervision. ## 2 RELATED WORKS Improving and eliciting mathematical reasoning abilities of LLMs. Mathematical reasoning tasks are one of the most challenging tasks for LLMs. Researchers have proposed various methods to improve or elicit the mathematical reasoning ability of LLMs, which can be broadly divided into three groups: 1) pre-training : The pre-training methods (OpenAI, 2023; Anil et al., 2023; Touvron et al., 2023; Azerbayev et al., 2023) pre-train LLMs on a vast of datasets that are related to math problems, such as the Proof-Pile and ArXiv (Azerbayev et al., 2023) with a simple next token prediction objective. 2) fine-tuning : The fine-tuning methods (Yu et al., 2023b; Luo et al., 2023; Yue et al., 2023; Wang et al., 2023b; Gou et al., 2023) can also enhance the mathematical reasoning ability of LLMs. The core of fine-tuning usually lies in constructing high-quality question-response pair datasets with a chain-of-thought reasoning process. and 3) prompting : The prompting methods (Wei et al., 2022; Zhang et al., 2023; Fu et al., 2022; Bi et al., 2023) aim to elicit the mathematical reasoning ability of LLMs by designing prompting strategy without updating the model parameters, which is very convenient and practical. Mathematical reasoning verification for LLMs. Except for directly improving and eliciting the mathematical reasoning potential of LLMs, the reasoning results can be boosted via an extra verifier for selecting the best answer from multiple decoded candidates. There are two primary types of verifiers: the Outcome Reward Model (ORM) and the Process Reward Model (PRM). The ORM allocates a score to the entire solution while the PRM assigns a score to each individual step in the reasoning process. Recent findings by (Lightman et al., 2023) suggest that PRM outperforms ORM. In addition to verification, reward models can offer invaluable feedback for further training of generators (Uesato et al., 2022; Pan et al., 2023). Compared to ORM, PRM provides more detailed feedback, demonstrating greater potential to enhance generator (Wu et al., 2023). However, training a PRM requires access to expensive human-annotated datasets (Uesato et al., 2022; Lightman et al., 2023), which hinders the advancement and practical application of PRM. Therefore, in this paper, we aim to build a PRM for mathematical reasoning without human annotation, and we explore the effectiveness of the automatic PRM with both verification and reinforcement learning scenarios. ## 3 METHODOLOGY In this section, we first present our task formulation to evaluate the performance of reward models (§3.1). Subsequently, we outline two typical categories of reward models, ORM and PRM(§3.2). Then, we introduce our methodology to automatically build the training dataset for PRM(§3.3), breaking the bottleneck of heavy reliance on manual annotation in existing work (Uesato et al., 2022; Lightman et al., 2023). ## 3.1 TASK FORMULATION We evaluate the performance of the reward model in two scenarios: Verification Following (Lightman et al., 2023), we consider a best-of-N selection evaluation paradigm. Specifically, given a problem p in the testing set, we sample N candidate solutions from a generator. These candidates are then scored using a reward model, and the highest-scoring solution is selected as the final answer. An enhanced reward model elevates the likelihood of selecting the solution containing the correct answer, consequently raising the success rate in solving mathematical problems for LLMs. Reinforcement learning We also use the automatically constructed PRM to supervise LLMs with step-by-step PPO. In this scenario, we evaluate the accuracy of the LLMs' greedy decoding output. An enhanced reward model is instrumental in training higher-performing LLMs. ## 3.2 REWARD MODELS FOR MATHEMATICAL PROBLEM ORM Given a mathematical problem p and its solution s , ORM ( P × S → R ) assigns a single real-value to s to indicate whether s is correct. ORM is usually trained with a cross-entropy loss (Cobbe et al., 2021; Li et al., 2023b): $$\displaystyle \sum _ { i = 1 } ^ { n } L _ { i s } ( 1 )$$ where y s is the golden answer of the solution s , y s = 1 if s is correct, otherwise y s = 0 . r s is the sigmoid score of s assigned by ORM. The success of the reward model hinges on the effective construction of the high-quality training dataset. As the math problem usually has a certain answer, we can automatically construct the training set of ORM by two steps: 1) sampling some candidate solutions for a problem from a generator; 2) assigning the label to each sampling solution by checking whether its answer is correct. Although false positives solutions that reach the correct answer with incorrect reasoning will be misgraded, previous studies have proven that it is still effective for training a good ORM (Lightman et al., 2023; Yu et al., 2023a). PRM Take a step further, PRM ( P × S → R + ) assigns a score to each reasoning step of s , which is usually trained with: $$\sum _ { i = 1 } ^ { K } y _ { s _ { i } } \log r _ { s _ { i } } + ( 1 - y _ { s _ { i } } )$$ where y s i is the golden answer of s i (the i -th step of s ), r s i is the sigmoid score of s i assigned by PRM and K is the number of reasoning steps for s . (Lightman et al., 2023) also conceptualizes the PRM training as a three-class classification problem, in which each step is classified as either 'good', 'neutral', or 'bad'. In this paper, we found that there is not much difference between the binary and the three classifications, and we regard PRM training as the binary classification. Compared to ORM, PRM can provide more detailed and reliable feedback (Lightman et al., 2023). However, there are currently no automated methods available for constructing high-quality PRM training datasets. Previous works (Uesato et al., 2022; Lightman et al., 2023) typically resort to costly human annotations. While PRM manages to outperform ORM (Lightman et al., 2023), the annotation cost invariably impedes both the development and application of PRM. ## 3.3 AUTOMATIC PROCESS ANNOTATION In this section, we propose an automatic process annotation framework to mitigate the annotation cost issues associated with PRM. We first define the quality of a reasoning step, followed by the introduction of our solution that obviates the necessity for human annotation. ## 3.3.1 DEFINITION Inspired by Monto Carlo Tree Search (Kocsis & Szepesv´ ari, 2006; Coulom, 2006; Silver et al., 2016; ´ Swiechowski et al., 2023), we define the quality of a reasoning step as its potential to deduce the correct answer. This criterion stems from the primary objective of the reasoning process, which essentially is a cognitive procedure aiding humans or intelligent agents in reaching a well-founded outcome (Huang & Chang, 2023). Therefore, a step that has the potential to deduce a well-founded result can be considered a good reasoning step. Analogous to ORM, this definition also introduces some degree of noise. Nevertheless, we find that it is beneficial for effectively training a good PRM. ## 3.3.2 SOLUTION Completion To quantify and estimate the potential for a give reasoning step s i , as shown in Figure 2, we use a 'completer' to finalize N subsequent reasoning processes from this step: { ( s i +1 ,j , · · · , s K j ,j , a j ) } N j =1 , where a j and K j are the decoded answer and the total number of steps for the j -th finalized solution, respectively. Then, we estimate the potential of this step based on the correctness of all decoded answers A = { a j } N j =1 . Figure 2: Comparison for previous automatic outcome annotation and our automatic process annotation. (a): automatic outcome annotation assigns a label to the entire solution S , dependent on the correctness of the answer; (b) automatic process annotation employs a 'completer' to finalize N reasoning processes (N=3 in this figure) for an intermediate step ( s 1 in this figure), subsequently use hard estimation (HE) and soft estimation (SE) to annotate this step based on all decoded answers. <details> <summary>Image 4 Details</summary>  ### Visual Description ## Diagram: Polynomial Root Problem Solution Process ### Overview The diagram illustrates a step-by-step solution process for a mathematical problem involving a monic polynomial of degree 4 with known roots (1, 2, 3) and an unknown root (r). It compares two solution paths: one leading to an incorrect answer (20) and another to the correct answer (24), annotated with process variables and outcome indicators. ### Components/Axes 1. **Problem Statement** (Top-left): - Text: "Let p(x) be a monic polynomial of degree 4. Three of the roots of p(x) are 1, 2, and 3. Find p(0) + p(4)." - Golden Answer: 24 (Top-right). 2. **Solution Structure** (Middle): - **Solution Path**: - Labeled steps: S = S₁, S₂, S₃, ..., Sₖ. - Outcome Annotation: "Answer: 20 ❌" (yₛ = 0). - **Process Annotation** (Bottom): - Detailed steps: S₁ → S₂,₁ → S₃,₁ → ... → Sₖ,₁ (Answer: 24 ✔️), S₂,₂ → S₂,₃ → ... → Sₖ,₂ (Answer: 24 ✔️), S₂,₃ → Sₖ,₃ (Answer: 20 ❌). - Variables: yₛᴱ = 2/3; yₛ₁ = 1. 3. **Legend** (Bottom-center): - Definitions: - sᵢ: The i-th step of the solution S. - sᵢⱼ: The i-th step of the j-th finalized solution. ### Detailed Analysis - **Problem Statement**: - Polynomial p(x) is monic (leading coefficient = 1) with roots 1, 2, 3, and r. The task is to compute p(0) + p(4). - **Solution Path**: - Initial steps (S₁, S₂, ..., Sₖ) lead to an incorrect answer (20), marked with a red X and yₛ = 0. - **Process Annotation**: - **S₁**: Explicitly defines p(x) = (x-1)(x-2)(x-3)(x-r), establishing the polynomial's form. - **S₂,₁ → Sₖ,₁**: Correct path leading to Answer: 24 ✔️. - **S₂,₂ → Sₖ,₂**: Alternative correct path (Answer: 24 ✔️). - **S₂,₃ → Sₖ,₃**: Incorrect path (Answer: 20 ❌). - **Legend**: - yₛᴱ = 2/3: Likely represents a weighting factor for solution steps. - yₛ₁ = 1: Indicates full confidence in the first step (S₁). ### Key Observations 1. **Divergent Outcomes**: Two solution paths yield the same correct answer (24), while one leads to an incorrect answer (20). 2. **Step Finalization**: The process annotation distinguishes between "finalized" solutions (S₂,₁/Sₖ,₁ and S₂,₂/Sₖ,₂) and non-finalized steps (S₂,₃/Sₖ,₃). 3. **Confidence Metrics**: yₛ₁ = 1 suggests S₁ is a foundational, error-free step, while yₛᴱ = 2/3 implies partial confidence in subsequent steps. ### Interpretation The diagram demonstrates a structured approach to solving polynomial problems, emphasizing: - **Root Identification**: Explicitly defining the polynomial's form using known roots (S₁). - **Solution Branching**: Multiple paths (S₂,₁/S₂,₂/S₂,₃) represent different methodological approaches, with some leading to correct results and others to errors. - **Confidence Indicators**: The legend's variables (yₛᴱ, yₛ₁) quantify confidence in solution steps, suggesting a probabilistic or iterative refinement process. The correct answer (24) arises from properly accounting for the unknown root (r) in the polynomial's expansion, while the incorrect answer (20) likely stems from an incomplete or misapplied step (e.g., omitting r in calculations). The process annotation highlights the importance of validating intermediate steps to avoid errors in final results. </details> Estimation In this paper, we use two methods to estimate the quality y s i for the step s i , hard estimation (HE) and soft estimation (SE). HE supposes that a reasoning step is good as long as it can reach the correct answer a ∗ : $$y _ { s _ { 1 } } = \int _ { 0 } ^ { 1 } z _ { a _ { j } } \in A , y _ { j } = a *$$ SE assumes the quality of a step as the frequency with which it reaches the correct answer: $$y _ { s } = \frac { \sum _ { j = 1 } ^ { N } y _ { j } = 1 ( a _ { j } = a * ) } { N } .$$ Once we gather the label of each step, we can train PRM with the cross-entropy loss. In conclusion, our automatic process annotation framework defines the quality of a step as its potential to deduce the correct answer and achieve the label of each step by completion and estimation. ## 3.4 RANKING FOR VERIFICATION Following (Lightman et al., 2023), we use the minimum score across all steps to represent the final score of a solution assigned by PRM. We also explore the combination of self-consistency and reward models following (Li et al., 2023b). In this context, we initially classify solutions into distinct groups according to their final answers. Following that, we compute the aggregate score for each group. Formally, the final prediction answer based on N candidate solutions is: $$a _ { sc + r m } = \arg _ { a } \max _ { i = 1 } ^ { N } \sum _ { i = 1 } ^ { N } I ( a _ { i } = a ) .$$ Where RM ( p, S i ) is the score of the i -th solution assigned by ORM or PRM for problem p . ## 3.5 REINFORCE LEARNING WITH PROCESS SUPERVISION Upon achieving PRM, we employ reinforcement learning to train LLMs. We implement Proximal Policy Optimization (PPO) in a step-by-step manner. This method differs from the conventional strategy that utilizes PPO with ORM, which only offers a reward at the end of the response. Conversely, our step-by-step PPO offers rewards at the end of each reasoning step. Table 1: Performances of different LLMs on GSM8K and MATH with different verification strategies. The reward models are trained based on LLama2-70B and LLemma-34B on GSM8K and MATH, respectively. The verification is based on 256 outputs. | Models | Verifiers | GSM8K | MATH500 | |------------------------|-----------------------------------------|---------|-----------| | LLaMA2-70B: MetaMATH | Self-Consistency | 88 | 39.4 | | LLaMA2-70B: MetaMATH | ORM | 91.8 | 40.4 | | LLaMA2-70B: MetaMATH | Self-Consistency+ORM | 92 | 42 | | LLaMA2-70B: MetaMATH | MATH-SHEPHERD (Ours) | 93.2 | 44.5 | | LLaMA2-70B: MetaMATH | Self-Consistency + MATH-SHEPHERD (Ours) | 92.4 | 45.2 | | LLemma-34B: MetaMATH | Self-Consistency | 82.6 | 44.2 | | LLemma-34B: MetaMATH | ORM | 90 | 43.7 | | LLemma-34B: MetaMATH | Self-Consistency+ORM | 89.6 | 45.4 | | LLemma-34B: MetaMATH | MATH-SHEPHERD (Ours) | 90.9 | 46 | | LLemma-34B: MetaMATH | Self-Consistency + MATH-SHEPHERD (Ours) | 89.7 | 47.3 | | DeepSeek-67B: MetaMATH | Self-Consistency | 88.2 | 45.4 | | DeepSeek-67B: MetaMATH | ORM | 92.6 | 45.3 | | DeepSeek-67B: MetaMATH | Self-Consistency+ORM | 92.4 | 47 | | DeepSeek-67B: MetaMATH | MATH-SHEPHERD (Ours) | 93.3 | 47 | | DeepSeek-67B: MetaMATH | Self-Consistency + MATH-SHEPHERD (Ours) | 92.5 | 48.1 | ## 4 EXPERIMENTS Datasets We conduct our experiments using two widely used math reasoning datasets, GSM8K (Cobbe et al., 2021) and MATH (Hendrycks et al., 2021). For the GSM8K dataset, we leverage the whole test set in both verification and reinforcement learning scenarios. For the MATH dataset, in the verification scenario, due to the computation cost, we employ a subset MATH500 that is identical to the test set of Lightman et al. (2023). The subset consists of 500 representative problems, and we find that the subset evaluation produces similar results to the full-set evaluation. To assess different verification methods, we generate 256 candidate solutions for each test problem. We report the mean accuracy of 3 groups of sampling results. In the reinforcement learning scenario, we use the whole test set to evaluate the model performance. We train LLMs with MetaMATH (Yu et al., 2023b). Parameter Setting Our experiments are based on a series of large language models, LLaMA27B/13B/70B (Touvron et al., 2023), LLemma-7B/34B (Azerbayev et al., 2023), Mistral-7B (Jiang et al., 2023) and DeepSeek-67B (DeepSeek, 2023). We train the generator and completer for 3 epochs on MetaMATH. We train the Mistral-7B with a learning rate of 5e-6. For other models, The learning rates are set to 2e-5, 1e-5, and 6e-6 for the 7B/13B, 34B, and 67B/70B LLMs, respectively. To construct the training dataset of ORM and PRM, we train 7B and 13B models for a single epoch on the GSM8K and MATH training sets. Subsequently, we sample 15 solutions per problem from each model for the training set. Following this, we eliminate duplicate solutions and annotate the solutions at each step. We use LLemma-7B as the completer with the decoded number N=8. Consequently, we obtain around 170k solutions for GSM8K and 270k solutions for MATH. For verification, we choose LLaMA2-70B and LLemma-34B as the base models to train reward models for GSM8K and MATH, respectively. For reinforcement learning, we choose Mistral-7B as the base model to train reward models and use it to supervise LLama2-7B and Mistral-7B generators. The reward model is trained in 1 epoch with a learning rate 1e-6. For the sake of convenience, we train the PRM using the hard estimation version because it allows us to utilize a standard language modeling pipeline by selecting two special tokens to represent 'has potential' and 'no potential' labels, thereby eliminating the need for any specific model adjustments. In reinforcement learning, the learning rate is 4e-7 and 1e-7 for LLaMA2-7B and Mistral-7B, respectively. The Kullback-Leibler coefficient is set to 0.04. We implement a cosine learning rate scheduler, employing a minimal learning rate set to 1e-8. We use 3D parallelism provided by hfai 1 to train all models with the max sequence length of 512. 1 https://doc.hfai.high-flyer.cn/index.html Table 2: Performances of different 7B models on GSM8K and MATH with greedy decoding. We use the questions in MetaMATH for RFT and PPO training. Both LLaMA2-7B and Mistral-7B are supervised by Mistral-7B-ORM and -MATH-SHEPHERD. | Models | GSM8K | MATH | |-----------------------------------------|---------|--------| | LLaMA2-7B: MetaMATH | 66.6 | 19.2 | | + RFT | 68.5 | 19.9 | | + ORM-PPO | 70.8 | 20.8 | | + MATH-SHEPHERD-step-by-step-PPO (Ours) | 73.2 | 21.6 | | Mistral-7B: MetaMATH | 77.9 | 28.6 | | + RFT | 79 | 29.9 | | + ORM-PPO | 81.8 | 31.3 | | + MATH-SHEPHERD-step-by-step-PPO (Ours) | 84.1 | 33 | Baselines and Metrics In the verification scenario, following (Lightman et al., 2023), we evaluate the performance of our reward model by comparing it against the Self-consistency (majority voting) and outcome reward model. The accuracy of the best-of-N solution is utilized as the evaluation metric. For PRM, the minimum score across all steps is adopted to represent the final score of a solution. In the reinforcement scenario, we compare our step-by-step supervision with the outcome supervision provided by ORM, and Rejective Sampling Fine-tuning (RFT) (Yuan et al., 2023), we sample 8 responses for each question in MetaMATH for RFT. We use the accuracy of LLMs' greedy decoding output to assess the performance. ## 4.1 MAIN RESULTS MATH-SHEPHERD as verifier Table 1 presents the performance comparison of various methods on GSM8K and MATH. We find that: 1) As the verifier, MATH-SHEPHERD consistently outperforms self-consistency and ORM on two datasets with all generators. Specifically, enhanced by MATHSHEPHERD, DeepSeek-67B achieves 93.3% and 48.1% accuracy on GSM8K and MATH; 2) In comparison to GSM8K, PRM achieves a greater advantage over ORM on the more challenging MATH dataset; This outcome aligns with the findings in Uesato et al. (2022) and Lightman et al. (2023). The former discovers that PRM and ORM yield similar results on GSM8K, whereas the latter shows that PRM significantly outperforms ORM on the MATH dataset. This could be attributed to the relative simplicity of the GSM8K dataset compared to MATH, i.e., the GSM8K dataset necessitates fewer steps for problem-solving. As a result, ORM operates efficiently when handling this particular dataset; 3) In GSM8K, when combined with self-consistency, there's a drop in performance, whereas in MATH, performance improves. These results indicate that if the reward model is sufficiently powerful for a task, combining it with self-consistency may harm the verification performance. MATH-SHEPHERD as reward model on reinforcement learning Table 2 presents the performance of different LLMs with greedy decoding outputs. As is shown: 1) step-by-step PPO significantly improves the performance of two supervised fine-tuned models. For example, Mistral-7B with step-by-step PPO achieves 84.1% and 33.0% on the GSM8K and MATH datasets, respectively; 2) RFT only slightly improves the model performance, we believe this is because MetaMATH already has conducted some data augmentation strategies like RFT; 3) the vanilla PPO with ORM can also enhance the model performance. However, it does not perform as well as the step-by-step PPO supervised by MATH-SHEPHERD, demonstrating the potential of step-by-step supervision. MATH-SHEPHERD as both reward models and verifiers We also combine the reinforcement learning and the verification. As shown in Table 3: 1) reinforcement learning and verification are complementary. For example, in MATH, step-by-step PPO Mistral-7B outperforms supervised fine-tuning Mistral-7B 7.2% accuracy with self-consistency as the verifier; The performance gap is even larger than that of greedy decoding results, i.e., 4.4%; 2) after reinforcement learning, the vanilla verification methods with only reward models is inferior to self-consistency, we think the Table 3: Results of reinforcement learning and verification combination. The reward models are trained based on Mistral-7B. The verification is based on 256 outputs. | Models | Verifiers | GSM8K | MATH500 | |--------------------------|-----------------------------------------|---------|-----------| | Mistral-7B: MetaMATH | Self-Consistency | 83.9 | 35.1 | | Mistral-7B: MetaMATH | ORM | 86.2 | 36.4 | | Mistral-7B: MetaMATH | Self-Consistency+ORM | 86.6 | 38 | | Mistral-7B: MetaMATH | MATH-SHEPHERD (Ours) | 87.1 | 37.3 | | Mistral-7B: MetaMATH | Self-Consistency + MATH-SHEPHERD (Ours) | 86.3 | 38.3 | | Mistral-7B: MetaMATH | Self-Consistency | 87.4 | 42.3 | | Mistral-7B: MetaMATH | ORM | 87.6 | 41.3 | | +step-by-step PPO (Ours) | Self-Consistency+ORM | 89 | 43.1 | | +step-by-step PPO (Ours) | MATH-SHEPHERD (Ours) | 88.4 | 41.1 | | +step-by-step PPO (Ours) | Self-Consistency + MATH-SHEPHERD (Ours) | 89.1 | 43.5 | reason is that the initial reward model is not sufficient to supervise the more powerful model after PPO. These results can also show the potential of iterative reinforcement learning, which we leave for future work. ## 5 ANALYSIS ## 5.1 PERFORMANCE WITH DIFFERENT NUMBER OF CANDIDATE SOLUTIONS Figure 3 illustrates the performance comparison of various strategies when applied to different numbers of candidates ranging from 1 to 256 on two benchmarks. The key observations are as follows: 1) PRM exhibits consistent superior performance when compared to both ORM and majority voting, with the degree of this superiority becoming more pronounced as N escalates. 2) In MATH, our automatically annotated datasets outperform the human-annotated PRM800K (Lightman et al., 2023). We ascribe this superiority to the distribution gap and the data quantity. Specifically, PRM800K is annotated based on the output from GPT-4, and consequently, a discrepancy arises for the output of open-source LLaMA models fine-tuned on MetaMATH. Furthermore, when considering the quantity of data, our automated reward model data exhibits both high scalability and a reduced labeling cost. Consequently, our dataset is four times larger than that provided in PRM800K. Overall, these results further underscore the effectiveness and potential of our method. ## 5.2 QUALITY OF THE AUTOMATIC PROCESS ANNOTATIONS In this section, we explore the quality of our automatic PRM dataset. To achieve this, we manually annotate 160 steps sampled from the training set of GSM8K and use different completers to infer from each step to achieve their label. We find that: Automatic process annotation exhibits satisfactory quality. Figure 4(a) demonstrates that utilizing LLaMA2-70B trained on MetaMATH as the completer, the accuracy of the hard estimation (HE) reaches 86% when N equals 4. This suggests that our automatically constructed dataset is of high quality. However, we observed a decline in the accuracy of the constructed dataset with further increases in N. Our analysis indicates that larger values for N may lead to false positives. Figure 4(b) shows the cross-entropy loss between SE and HE labels compared to the human-annotated distribution: as N increases, SE progressively aligns closer to the standard distribution, in contrast to HE which does not exhibit similar behavior. It is essential to note that at N=4, HE achieves an accuracy of 86%. We can theoretically attain higher quality data exceeding 86% accuracy by utilizing SE. However, we discovered that the performance of the verifier exhibits no substantial divergence whether trained with either SE or HE. This may be attributable to the already high-quality annotations provided by HE. Furthermore, we also delve into other automatic process annotation methodologies. For instance, (Li et al., 2023b) employs a natural language inference (NLI) model and a string match rule to annotate a Figure 3: Performance of LLaMA2-70B using different verification strategies across different numbers of solution candidates on GSM8K and MATH. <details> <summary>Image 5 Details</summary>  ### Visual Description ## Line Graphs: Performance of Problem-Solving Methods Across N Solutions ### Overview The image contains two line graphs comparing the performance of different problem-solving methods (SC, ORM, PRM800K, SHEPHERD) on two datasets: **GSM8K** (left) and **MATH** (right). The x-axis represents the number of solutions per problem (N = 1, 4, 16, 64, 256), and the y-axis shows the percentage of problems solved. Each method is represented by a distinct colored line, with shaded regions indicating confidence intervals. --- ### Components/Axes #### **GSM8K Graph (Left)** - **X-axis**: N = number of solutions per problem (1, 4, 16, 64, 256) - **Y-axis**: % Problems Solved (Best-of-N) (80% to 92.5%) - **Legend**: - **SC** (red) - **ORM** (blue) - **SHEPHERD** (green) - **Shading**: Confidence intervals around each line. #### **MATH Graph (Right)** - **X-axis**: N = number of solutions per problem (1, 4, 16, 64, 256) - **Y-axis**: % Problems Solved (Best-of-N) (30% to 45%) - **Legend**: - **SC** (red) - **ORM** (blue) - **PRM800K** (purple) - **SHEPHERD** (green) - **Shading**: Confidence intervals around each line. --- ### Detailed Analysis #### **GSM8K Graph** - **SC (Red)**: - Starts at **80%** (N=1), increases to **82.5%** (N=4), **85%** (N=16), **87.5%** (N=64), and plateaus at **87.5%** (N=256). - **Trend**: Gradual, linear improvement with diminishing returns. - **ORM (Blue)**: - Starts at **80%** (N=1), rises to **87.5%** (N=4), **91%** (N=16), **92%** (N=64), and stabilizes at **92%** (N=256). - **Trend**: Steeper initial growth, then plateaus. - **SHEPHERD (Green)**: - Starts at **80%** (N=1), jumps to **87.5%** (N=4), **92%** (N=16), **93%** (N=64), and **93.5%** (N=256). - **Trend**: Rapid early improvement, sustained high performance. #### **MATH Graph** - **SC (Red)**: - Starts at **30%** (N=1), increases to **32%** (N=4), **35%** (N=16), **37.5%** (N=64), and **39%** (N=256). - **Trend**: Slow, linear growth with minimal gains at higher N. - **ORM (Blue)**: - Starts at **30%** (N=1), rises to **35%** (N=4), **39%** (N=16), **40%** (N=64), and **40%** (N=256). - **Trend**: Moderate improvement, plateaus after N=16. - **PRM800K (Purple)**: - Starts at **30%** (N=1), increases to **35%** (N=4), **39%** (N=16), **40.5%** (N=64), and **41%** (N=256). - **Trend**: Similar to ORM but slightly higher at N=64 and N=256. - **SHEPHERD (Green)**: - Starts at **30%** (N=1), jumps to **35%** (N=4), **40%** (N=16), **43%** (N=64), and **44%** (N=256). - **Trend**: Strong early gains, sustained outperformance. --- ### Key Observations 1. **SHEPHERD Dominates**: In both datasets, SHEPHERD consistently outperforms other methods, especially at higher N values (e.g., 93.5% vs. 87.5% in GSM8K at N=256). 2. **SC Lags Behind**: SC shows the slowest improvement across all N values, suggesting lower efficiency. 3. **ORM vs. PRM800K**: ORM and PRM800K perform similarly in MATH, but PRM800K edges out ORM at N=64 and N=256. 4. **Dataset Differences**: GSM8K problems have higher baseline performance (80–93.5%) compared to MATH (30–44%), indicating potential differences in problem difficulty or method suitability. --- ### Interpretation - **Efficiency of SHEPHERD**: Its rapid improvement and high performance at large N suggest it is optimized for scaling, possibly leveraging parallel or hierarchical problem-solving strategies. - **SC’s Limitations**: The flat trend for SC implies it struggles with complex or large-scale problems, possibly due to a lack of adaptive mechanisms. - **ORM/PRM800K Trade-offs**: While ORM and PRM800K show moderate gains, their plateaus at higher N suggest diminishing returns, highlighting the need for methods that balance exploration and exploitation. - **Dataset-Specific Insights**: The stark performance gap between GSM8K and MATH may reflect differences in problem structure (e.g., GSM8K’s arithmetic vs. MATH’s reasoning tasks), emphasizing the importance of method adaptability. This analysis underscores the critical role of method design in scaling problem-solving capabilities, with SHEPHERD emerging as the most robust approach across both datasets. </details> Figure 4: Quality of process annotation on GSM8K. (a): Accuracy of the process annotation using different completer; (b): Loss of the process annotation using different completer; (c): Loss of the process annotation using the same completer with different training data. <details> <summary>Image 6 Details</summary>  ### Visual Description ## Line Graphs: Model Performance vs. Decoded Paths ### Overview The image contains three line graphs comparing model performance metrics (accuracy and loss) across different model sizes (7B, 13B, 70B), training methods (Soft vs. Hard), and data types (Normal, Weak, Augmented). Each graph plots performance against the number of decoded paths (N = 1, 4, 16, 64, 256). --- ### Components/Axes 1. **Left Graph: Accuracy (%)** - **X-axis**: Number of decoded paths (N) – logarithmic scale (1, 4, 16, 64, 256). - **Y-axis**: Accuracy (%) – linear scale (80–86%). - **Legend**: - Red: 7B - Blue: 13B - Green: 70B 2. **Middle Graph: Loss** - **X-axis**: Number of decoded paths (N) – same as above. - **Y-axis**: Loss – linear scale (1.0–3.0). - **Legend**: - Red: 7B:Soft - Blue: 13B:Soft - Green: 70B:Soft - Purple: 70B:Hard 3. **Right Graph: Loss** - **X-axis**: Number of decoded paths (N) – same as above. - **Y-axis**: Loss – linear scale (1.0–4.0). - **Legend**: - Red: Normal - Blue: Weak - Green: Augmented --- ### Detailed Analysis #### Left Graph: Accuracy - **7B (Red)**: Peaks at 84% accuracy at N=4, then declines to 81% at N=256. - **13B (Blue)**: Peaks at 85% at N=4, then drops to 83% at N=256. - **70B (Green)**: Peaks at 86% at N=4, stabilizes at 84% at N=256. #### Middle Graph: Loss (Soft Training) - **7B:Soft (Red)**: Starts at 2.5 loss at N=1, decreases to 1.8 at N=256. - **13B:Soft (Blue)**: Starts at 2.4, decreases to 1.6 at N=256. - **70B:Soft (Green)**: Starts at 2.2, decreases to 1.5 at N=256. - **70B:Hard (Purple)**: Starts at 2.5, decreases to 1.8 at N=256. #### Right Graph: Loss (Data Types) - **Normal (Red)**: Starts at 3.5 loss at N=1, decreases to 2.0 at N=256. - **Weak (Blue)**: Starts at 3.8, decreases to 1.8 at N=256. - **Augmented (Green)**: Starts at 2.8, decreases to 1.6 at N=256. --- ### Key Observations 1. **Accuracy Trends**: - All models show peak accuracy at N=4, followed by a decline. The 70B model maintains higher accuracy across all N values. - The 7B model’s accuracy drops sharply after N=4, suggesting overfitting or diminishing returns. 2. **Loss Trends**: - Loss decreases with increasing N for all models and training methods. The 70B:Soft model achieves the lowest loss (1.5 at N=256). - The 70B:Hard model has higher loss than 70B:Soft but still outperforms smaller models. 3. **Data Type Impact**: - Augmented data (Green) maintains the lowest loss across all N values, indicating better generalization. - Weak data (Blue) shows the steepest loss reduction, suggesting it benefits more from increased N. --- ### Interpretation - **Model Size vs. Performance**: Larger models (70B) achieve higher accuracy and lower loss, but require more computational resources. The 7B model’s performance degrades significantly at higher N, highlighting scalability challenges. - **Training Methods**: Soft training (70B:Soft) yields lower loss than Hard training (70B:Hard), suggesting softer optimization improves efficiency. - **Data Augmentation**: Augmented data (Green) reduces loss more effectively than Normal or Weak data, implying robustness to data variability. - **Anomalies**: The 7B model’s accuracy drop after N=4 is unusual, as larger models maintain performance. This may indicate architectural limitations or overfitting in smaller models. The data underscores the trade-offs between model size, training strategies, and data quality in optimizing performance. Augmented data and larger models (70B) appear most effective, but practical deployment must balance resource constraints. </details> given step. The NLI-based method annotates a step as correct if it is entailment with any step in the reference solutions. The Rule-based method annotates a step as correct if its support number precisely matches that of any steps in the reference solutions. As demonstrated in Table 4, our annotation strategy exhibits substantial superiority over the two approaches. The ability of the LLM completer plays an important role in the data quality. We employ a completer to finalize multiple subsequent reasoning processes for a given step. Therefore, we investigate the impact of the LLM completer. Figure 4(b) presents the cross-entropy loss across diverse completers trained on MetaMath. The results indicate that a larger completer is adept at generating superior-quality datasets. Figure 4(c) depicts the cross-entropy loss of LLaMA2-70B trained with different datasets. 'Normal' denotes the original GSM8K training dataset; 'Weak' refers to the Normal set excluding examples whose questions are in our 160 evaluation set; while 'Augmented' symbolizes MetaMath, an augmented version of the Normal set. The findings suggest that high-quality training sets allow the model to operate more proficiently as a completer. Importantly, the 'Weak' set exhibits a markedly larger loss than other datasets. This insight drives us to infer that LLMs should acquire the questions in advance to enhance their performance as completers. We can also conjecture that a stronger foundational model, coupled with superior training data, could further enhance the quality of automatic annotation. ## 5.3 INFLUENCE OF THE PRE-TRAINED BASE MODELS To conduct an exhaustive evaluation of MATH-SHEPHERD's effectiveness, we performed a diverse range of experiments using model sizes 7B, 13B, and 70B. Table 4: The comparison between NLI/Rule-based automatic process annotation methods from Li et al. (2023b) and our method. | Methods | Models | Accuracy (%) | Loss | |---------------------------------|---------------------------|----------------|--------| | DIVERSE-NLI (Li et al., 2023b) | DeBERTa (He et al., 2020) | 61.3 | 5.43 | | DIVERSE-NLI (Li et al., 2023b) | LLaMA2-13B | 75.6 | 3.27 | | DIVERSE-Rule (Li et al., 2023b) | - | 75 | 3.43 | | MATH-SHEPHERD | LLaMA2-13B (N = 4) | 85 | 2.05 | Figure 5: Performance of different verification strategies on different sizes of generators and verifiers. <details> <summary>Image 7 Details</summary>  ### Visual Description ## Line Graphs: Algorithm Performance Comparison Across Problem Sizes and Model Configurations ### Overview The image contains four line graphs comparing the performance of three algorithms (SC, ORM, SHEPHERD) across varying problem sizes (N = 1, 4, 16, 64, 256) and model configurations (Generator/Verifier: 7B, 13B, 70B). Each graph tracks the percentage of problems solved (% Problems Solved) as N increases, with distinct trends for each algorithm and model pair. --- ### Components/Axes 1. **Subplots**: - (a) Generator: 7B; Verifier: 7B - (b) Generator: 13B; Verifier: 13B - (c) Generator: 70B; Verifier: 7B - (d) Generator: 7B; Verifier: 70B 2. **Axes**: - **X-axis**: Number of solutions per problem (N), labeled as "N = number of solutions per problem". Values: 1, 4, 16, 64, 256. - **Y-axis**: Percentage of problems solved (% Problems Solved), ranging from ~62% to ~88% across subplots. 3. **Legend**: - Located in the bottom-right corner of each subplot. - **Colors**: - Red: SC - Blue: ORM - Green: SHEPHERD --- ### Detailed Analysis #### Subplot (a): Generator: 7B; Verifier: 7B - **SC (Red)**: Starts at ~62% (N=1), rises sharply to ~71% (N=4), then plateaus to ~72% (N=16–256). - **ORM (Blue)**: Begins at ~62% (N=1), peaks at ~73% (N=16), then declines slightly to ~72% (N=256). - **SHEPHERD (Green)**: Starts at ~62% (N=1), rises steadily to ~74% (N=256). #### Subplot (b): Generator: 13B; Verifier: 13B - **SC (Red)**: Starts at ~68% (N=1), increases to ~76% (N=16), then plateaus to ~77% (N=256). - **ORM (Blue)**: Begins at ~68% (N=1), rises to ~80% (N=16), then stabilizes at ~80% (N=256). - **SHEPHERD (Green)**: Starts at ~68% (N=1), increases to ~81% (N=16), then plateaus to ~82% (N=256). #### Subplot (c): Generator: 70B; Verifier: 7B - **SC (Red)**: Starts at ~81% (N=1), rises sharply to ~87% (N=16), then plateaus to ~88% (N=256). - **ORM (Blue)**: Begins at ~81% (N=1), increases to ~86% (N=16), then declines slightly to ~85% (N=256). - **SHEPHERD (Green)**: Starts at ~81% (N=1), rises to ~86% (N=16), then plateaus to ~86% (N=256). #### Subplot (d): Generator: 7B; Verifier: 70B - **SC (Red)**: Starts at ~65% (N=1), increases to ~70% (N=16), then plateaus to ~71% (N=256). - **ORM (Blue)**: Begins at ~65% (N=1), rises to ~80% (N=16), then plateaus to ~85% (N=256). - **SHEPHERD (Green)**: Starts at ~65% (N=1), increases to ~83% (N=16), then plateaus to ~86% (N=256). --- ### Key Observations 1. **Algorithm Performance**: - **SHEPHERD** consistently outperforms SC and ORM in most configurations, especially at larger N (e.g., N=256 in subplot d). - **SC** shows strong performance with larger generators (e.g., 70B in subplot c) but struggles with smaller generators (e.g., subplot a). - **ORM** performs best when the verifier matches the generator size (e.g., subplot b) but underperforms when the verifier is smaller (e.g., subplot c). 2. **Model Size Impact**: - Larger generators (70B) generally improve performance across all algorithms, particularly for SC (subplot c). - Mismatched generator-verifier sizes (e.g., subplot d: 7B generator, 70B verifier) reduce performance for SC but benefit ORM and SHEPHERD. 3. **Trends**: - All algorithms show diminishing returns at N=256, suggesting saturation. - SHEPHERD’s performance is less sensitive to N increases compared to SC and ORM. --- ### Interpretation The data suggests that **SHEPHERD** is the most robust algorithm, maintaining high performance across varying problem sizes and model configurations. **SC** excels with large generators but falters with smaller ones, likely due to its reliance on computational resources. **ORM** performs optimally when generator and verifier sizes align, highlighting its dependency on model consistency. The results imply that algorithm choice should consider both problem complexity (N) and model scale (generator/verifier size) for optimal outcomes. </details> Figures 5(a), 5(b), and 3(a) display the results from the 7B, 13B, and 70B generators paired with equal-sized reward models, respectively. It becomes evident that PRM exhibits superiority over self-consistency and ORM across all sizes of base models. Moreover, bigger reward models prove to be more robust; for instance, the accuracy of the 70B reward models escalates as the number of candidate solutions rises, while the 7B reward models show a decreasing trend. Figure 5(c) and 5(d) presents the performance of 7B and 70B generators interfaced with differentsized reward models. The findings illustrate that utilizing a larger reward model to validate the output of a smaller generator significantly enhances performance. Conversely, when a smaller reward model is employed to validate the output of a larger generator, the verification process adversely impacts the model's performance compared to SC. These results substantiate that we should utilize a more potent reward model for validating or supervising the generator. ## 5.4 INFLUENCE OF THE NUMBER OF DATA We delve deeper into the analysis of PRM and ORM by utilizing varying quantities of training data. As depicted in Figure 6(a), it is clear that PRM exhibits superior data efficiency. Specifically, it outperforms ORM by approximately 4% accuracy when applying a modestly sized training dataset (i.e., 10k instances). Furthermore, PRM seems to have a higher potential ceiling than ORM. These observations highlight the efficacy of PRM for verification purposes. ## 5.5 OUT-OF-DISTRIBUTION PERFORMANCE To further demonstrate the effectiveness of our method, we conduct an out-of-distribution evaluation on the Hungarian national final exam 2 , which consists of 33 questions. The total score of these questions is 100. We use the LLemma-34B trained on MetaMATH to serve as the generator and generate 256 candidate solutions for each question. We use LLemma-34B-ORM and LLemma34B-PRM to select the solution for each question. As shown in Figure 6(b): 1) both LLemma34B-ORM and LLemma-34B-PRM outperform the origin LLemma-34B, showing the reward model can generalize to other domains; 2) PRM outperforms ORM 9 scores, further demonstrating the superiority of PRM. 2 https://huggingface.co/datasets/keirp/hungarian\_national\_hs\_finals\_ exam Figure 6: (a): Performance of different reward models using different numbers of training data; (b) performance of different verification strategies on the out-of-distribution Hungarian national exam. <details> <summary>Image 8 Details</summary>  ### Visual Description ## Line Chart: % Problems Solved vs. Training Solutions ### Overview A line chart comparing three methods (SC, ORM, SHEPHERD) across varying numbers of training solutions (10k to 160k) in terms of "% Problems Solved (Best-of-256)". ### Components/Axes - **X-axis**: "Number of training solutions" (10k, 20k, 40k, 80k, 160k). - **Y-axis**: "% Problems Solved (Best-of-256)" (88% to 94%, with axis labels up to 92%). - **Legend**: - **SC**: Red line (flat at 88%). - **ORM**: Blue line (peaks at 92% at 20k, dips to ~91% at 80k, rises to 92% at 160k). - **SHEPHERD**: Green line (rises from 90% at 10k to 94% at 160k). - **Legend Position**: Bottom-right corner. ### Detailed Analysis - **SC**: Flat line at 88% across all training solutions. - **ORM**: - Starts at 86% (10k), jumps to 92% (20k), remains stable (~92%) at 40k and 160k, dips to ~91% at 80k. - **SHEPHERD**: - Starts at 90% (10k), increases steadily to 94% (160k). ### Key Observations - SHEPHERD consistently outperforms SC and ORM as training solutions increase. - ORM shows volatility (e.g., dip at 80k) but generally matches SHEPHERD’s performance at higher training volumes. - SC remains stagnant regardless of training scale. - SHEPHERD’s y-axis value (94%) exceeds the labeled axis maximum (92%), suggesting a potential axis truncation or data anomaly. ## Bar Chart: Method Scores ### Overview A bar chart comparing three methods (Greedy, ORM, SHEPHERD) by "Score" (30–70). ### Components/Axes - **X-axis**: Methods (Greedy, ORM, SHEPHERD). - **Y-axis**: "Score" (30–70). - **Legend**: - **Greedy**: Light blue bar (46.0). - **ORM**: Dark blue bar (54.0). - **SHEPHERD**: Green bar (63.0). - **Legend Position**: Top-right corner. ### Detailed Analysis - **Greedy**: Lowest score (46.0). - **ORM**: Mid-range score (54.0). - **SHEPHERD**: Highest score (63.0). ### Key Observations - SHEPHERD dominates in both charts, outperforming ORM and Greedy by significant margins. - ORM’s score (54.0) aligns with its line chart performance (~91–92% problem-solving). - Greedy’s low score (46.0) contrasts with its line chart baseline (88%), indicating a different evaluation metric. ## Interpretation 1. **Performance Trends**: - SHEPHERD demonstrates superior scalability and efficiency, achieving higher problem-solving rates and scores across all training scales. - ORM’s volatility (e.g., dip at 80k) suggests potential instability or sensitivity to training data size. - SC’s flat performance implies it is either capped or ineffective at leveraging additional training data. 2. **Method Comparison**: - SHEPHERD’s consistent dominance in both charts highlights its robustness, possibly due to advanced optimization or algorithmic design. - The bar chart’s "Score" metric (likely a composite or alternative evaluation) reinforces SHEPHERD’s superiority, even when compared to simpler methods like Greedy. 3. **Anomalies**: - SHEPHERD’s y-axis value (94%) exceeding the labeled maximum (92%) warrants investigation—potential axis mislabeling or data outlier. - ORM’s dip at 80k training solutions may indicate a temporary degradation or overfitting. 4. **Implications**: - SHEPHERD is the most reliable method for scaling with training data. - SC’s stagnation suggests it may not be suitable for dynamic or large-scale problem-solving tasks. - The bar chart’s "Score" metric could reflect real-world applicability, where SHEPHERD’s higher score translates to practical advantages. </details> Table 5: A case study from the Hungarian national exam. Red text denotes the mistake that ORM fails to detect. <details> <summary>Image 9 Details</summary>  ### Visual Description ## Textual Comparison: MATH-SHEPHERD vs. ORM Methodologies ### Overview The image compares two problem-solving approaches (MATH-SHEPHERD and ORM) for an arithmetic sequence problem. Both methods attempt to calculate the 13th term of a sequence where the first term is 18, the sum of the first six terms equals the sum of the first seven terms, and the sum of the first 13 terms is 0. Scores (MATH-SHEPHERD in teal, ORM in yellow) are provided for each step, indicating accuracy or confidence. ### Components/Axes - **Structure**: Vertical comparison with labeled steps (Step 1–9) for each method. - **Textual Elements**: - Problem statement at the top. - Step-by-step calculations for both methods. - Scores (e.g., "MATH-SHEPHERD: 0.85", "ORM Score: 0.73"). - **Color Coding**: - Teal for MATH-SHEPHERD scores. - Yellow for ORM scores. ### Detailed Analysis #### MATH-SHEPHERD Method 1. **Step 1**: Defines the common difference as `d`. Score: 0.99. 2. **Step 2**: Sum of first six terms: `108 + 30d`. Score: 0.85. 3. **Step 6**: Solves for `d = -3`. Score: 0.85. 4. **Step 8**: Calculates 13th term: `18 - 36 = -18`. Score: 0.93. 5. **Final Score**: 0.73. #### ORM Method 1. **Step 1**: Defines the common difference as `d`. Score: 0.99. 2. **Step 2**: Sum of first six terms: `90 + 15d`. Score: 0.53. 3. **Step 6**: Solves for `d = -2` (incorrectly divides by -6). Score: 0.38. 4. **Step 9**: Calculates 13th term: `18 - 26 = -8`. Score: 0.38. 5. **Final Score**: 0.84. ### Key Observations - **MATH-SHEPHERD** consistently achieves higher scores (0.85–0.93) compared to ORM (0.38–0.53). - **ORM** introduces errors in Step 6 (incorrect division by -6) and Step 9 (miscalculates 13th term). - Both methods agree on the first term (18) and initial steps but diverge in solving for `d` and the final term. ### Interpretation The MATH-SHEPHERD method demonstrates greater accuracy and consistency in solving the arithmetic sequence problem. Its higher scores suggest robust handling of summation formulas and algebraic manipulation. The ORM method’s lower scores stem from arithmetic errors (e.g., incorrect division in Step 6) and misapplication of sequence properties. Notably, the final scores (0.73 vs. 0.84) appear counterintuitive given the step-by-step performance, indicating potential scoring criteria that prioritize specific steps over others. The problem itself highlights the importance of precise algebraic reasoning: - The sum of the first six terms equals the sum of the first seven terms only if the seventh term is zero, implying `d = -3` (MATH-SHEPHERD’s result). - The 13th term calculation (`18 + 12d`) aligns with MATH-SHEPHERD’s `-18` but conflicts with ORM’s `-8`, underscoring the impact of errors in intermediate steps. </details> | Question | The first term of an arithmetic sequence is 18. The sum of the first six terms of the sequence is equal to the sum of the first seven terms. Show that the sum of the first 13 terms of the sequence is 0 and calculate the 13th term of the sequence. | |---------------|------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------| | MATH-SHEPHERD | Step 1: Let the common difference of the arithmetic sequence be d . MATH-SHEPHERD: 0.99. Step 2: The sum of the first six terms of the sequence is 18 +(18+ d ) +(18+2 d ) +(18+3 d ) +(18+4 d )+ (18+5 d ) = 108 +30 d . MATH-SHEPHERD: 0.85. ...... Step 6: Therefore, d = - 3 . MATH-SHEPHERD: 0.85. ...... Step 8: The 13th term of the sequence is 18 - 36 = - 18 . MATH-SHEPHERD: 0.93. ORMScore: 0.73. | | ORM | Step 1: Let the common difference of the arithmetic sequence be d . MATH-SHEPHERD: 0.99. Step 2: The sum of the first six terms is 18 +(18+ d ) +(18+2 d ) +(18+3 d ) +(18+4 d ) +(18+5 d ) = 90 + 15 d . MATH-SHEPHERD: 0.53. ...... Step 6: Dividing by - 6 , we find that d = - 2 . MATH-SHEPHERD: 0.38. ...... Step 9: The 13th term of the sequence is 18 - 26 = - 8 . MATH-SHEPHERD: 0.38. ORMScore: 0.84. | We also conduct a case study to intuitively demonstrate the effectiveness of MATH-SHEPHERD. As outlined in Table 5, when presented with a question from the Hungarian national final exam, our MATH-SHEPHERD accurately selected the correct solution from a pool of 256 potential solutions, which ORM failed. Moreover, MATH-SHEPHERD displayed superior discernment by precisely identifying incorrect steps within the solutions selected by ORM. Notably, it recognized errors in Step 2, Step 6, and Step 9 and so on, and subsequently assigned them lower scores relative to those for steps present in the correct solutions. ## 6 LIMITATIONS Our paper has some limitations, which we leave for future work: The computational cost of the completion process. To determine the label of each reasoning step, we utilize a 'completer' to decode N subsequent reasoning processes. We observe that as N increases, so does the quality of automatic annotations. However, this completion process demands a lot of computing resources, potentially imposing a limitation on the usage of our method. Despite this limitation, the cost remains significantly lower than human annotation. Furthermore, we are optimistic that advancements in efficient inference techniques such as speculative decoding (Xia et al., 2022; Leviathan et al., 2023) and vLLM (Kwon et al., 2023) could mitigate this limitation. The automatic process annotation consists of noise. Similar to the automatic outcome annotation, our automatic process annotation also has noise. Despite this, our experiments verify the efficacy of our method for training a PRM. In particular, the PRM trained on our dataset outperforms the human-annotated PRM800K dataset. However, a noticeable gap remains between PRM800K and the candidate responses generated by the open-source models utilized in this study, which may result in the invalidation of PRM800K. As a result, the impact of this potential noise on PRM performance is still undetermined. A comprehensive comparison between human and automated annotations is envisaged for future studies. Furthermore, we assert that integrating human and automated process annotations could play a vital role in constructing robust and efficient process supervision. ## 7 CONCLUSION In this paper, we introduce a process-oriented math verifier called MATH-SHEPHERD, which assigns a reward score to each step of the LLM's outputs on math problems. The training of MATH-SHEPHERD is achieved using automatically constructed process-wise supervision data, thereby eradicating the necessity for labor-intensive human annotation. Remarkably, this automatic methodology correlates strongly with human annotations. Extensive experiments in both verification and reinforcement learning scenarios demonstrate the effectiveness of our method. ## REFERENCES - Rohan Anil, Andrew M Dai, Orhan Firat, Melvin Johnson, Dmitry Lepikhin, Alexandre Passos, Siamak Shakeri, Emanuel Taropa, Paige Bailey, Zhifeng Chen, et al. Palm 2 technical report. arXiv preprint arXiv:2305.10403 , 2023. - Zhangir Azerbayev, Hailey Schoelkopf, Keiran Paster, Marco Dos Santos, Stephen McAleer, Albert Q Jiang, Jia Deng, Stella Biderman, and Sean Welleck. Llemma: An open language model for mathematics. arXiv preprint arXiv:2310.10631 , 2023. - Zhen Bi, Ningyu Zhang, Yinuo Jiang, Shumin Deng, Guozhou Zheng, and Huajun Chen. When do program-of-thoughts work for reasoning? arXiv preprint arXiv:2308.15452 , 2023. - S´ ebastien Bubeck, Varun Chandrasekaran, Ronen Eldan, Johannes Gehrke, Eric Horvitz, Ece Kamar, Peter Lee, Yin Tat Lee, Yuanzhi Li, Scott Lundberg, et al. Sparks of artificial general intelligence: Early experiments with gpt-4. arXiv preprint arXiv:2303.12712 , 2023. - Liang Chen, Yichi Zhang, Shuhuai Ren, Haozhe Zhao, Zefan Cai, Yuchi Wang, Peiyi Wang, Tianyu Liu, and Baobao Chang. Towards end-to-end embodied decision making via multi-modal large language model: Explorations with gpt4-vision and beyond. arXiv preprint arXiv:2310.02071 , 2023. - Karl Cobbe, Vineet Kosaraju, Mohammad Bavarian, Mark Chen, Heewoo Jun, Lukasz Kaiser, Matthias Plappert, Jerry Tworek, Jacob Hilton, Reiichiro Nakano, et al. Training verifiers to solve math word problems. arXiv preprint arXiv:2110.14168 , 2021. - R´ emi Coulom. Efficient selectivity and backup operators in monte-carlo tree search. In International conference on computers and games , pp. 72-83. Springer, 2006. - DeepSeek. Deepseek llm: Let there be answers. https://github.com/deepseek-ai/ DeepSeek-LLM , 2023. - Yao Fu, Hao Peng, Ashish Sabharwal, Peter Clark, and Tushar Khot. Complexity-based prompting for multi-step reasoning. arXiv preprint arXiv:2210.00720 , 2022. - Zhibin Gou, Zhihong Shao, Yeyun Gong, Yujiu Yang, Minlie Huang, Nan Duan, Weizhu Chen, et al. Tora: A tool-integrated reasoning agent for mathematical problem solving. arXiv preprint arXiv:2309.17452 , 2023. - Pengcheng He, Xiaodong Liu, Jianfeng Gao, and Weizhu Chen. Deberta: Decoding-enhanced bert with disentangled attention. arXiv preprint arXiv:2006.03654 , 2020. - Dan Hendrycks, Collin Burns, Saurav Kadavath, Akul Arora, Steven Basart, Eric Tang, Dawn Song, and Jacob Steinhardt. Measuring mathematical problem solving with the math dataset. arXiv preprint arXiv:2103.03874 , 2021. - Jie Huang and Kevin Chen-Chuan Chang. Towards reasoning in large language models: A survey. In Anna Rogers, Jordan Boyd-Graber, and Naoaki Okazaki (eds.), Findings of the Association for Computational Linguistics: ACL 2023 , pp. 1049-1065, Toronto, Canada, July 2023. Association for Computational Linguistics. doi: 10.18653/v1/2023.findings-acl.67. URL https://aclanthology.org/2023.findings-acl.67 . - Jie Huang, Xinyun Chen, Swaroop Mishra, Huaixiu Steven Zheng, Adams Wei Yu, Xinying Song, and Denny Zhou. Large language models cannot self-correct reasoning yet. arXiv preprint arXiv:2310.01798 , 2023. - Albert Q Jiang, Alexandre Sablayrolles, Arthur Mensch, Chris Bamford, Devendra Singh Chaplot, Diego de las Casas, Florian Bressand, Gianna Lengyel, Guillaume Lample, Lucile Saulnier, et al. Mistral 7b. arXiv preprint arXiv:2310.06825 , 2023. - Jean Kaddour, Joshua Harris, Maximilian Mozes, Herbie Bradley, Roberta Raileanu, and Robert McHardy. Challenges and applications of large language models. arXiv preprint arXiv:2307.10169 , 2023. - Levente Kocsis and Csaba Szepesv´ ari. Bandit based monte-carlo planning. In European conference on machine learning , pp. 282-293. Springer, 2006. - Woosuk Kwon, Zhuohan Li, Siyuan Zhuang, Ying Sheng, Lianmin Zheng, Cody Hao Yu, Joseph Gonzalez, Hao Zhang, and Ion Stoica. Efficient memory management for large language model serving with pagedattention. In Proceedings of the 29th Symposium on Operating Systems Principles , pp. 611-626, 2023. - Yaniv Leviathan, Matan Kalman, and Yossi Matias. Fast inference from transformers via speculative decoding. In International Conference on Machine Learning , pp. 19274-19286. PMLR, 2023. - Lei Li, Yuwei Yin, Shicheng Li, Liang Chen, Peiyi Wang, Shuhuai Ren, Mukai Li, Yazheng Yang, Jingjing Xu, Xu Sun, et al. M3it: A large-scale dataset towards multi-modal multilingual instruction tuning. arXiv preprint arXiv:2306.04387 , 2023a. - Yifei Li, Zeqi Lin, Shizhuo Zhang, Qiang Fu, Bei Chen, Jian-Guang Lou, and Weizhu Chen. Making language models better reasoners with step-aware verifier. In Anna Rogers, Jordan BoydGraber, and Naoaki Okazaki (eds.), Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) , pp. 5315-5333, Toronto, Canada, July 2023b. Association for Computational Linguistics. doi: 10.18653/v1/2023.acl-long.291. URL https://aclanthology.org/2023.acl-long.291 . - Hunter Lightman, Vineet Kosaraju, Yura Burda, Harri Edwards, Bowen Baker, Teddy Lee, Jan Leike, John Schulman, Ilya Sutskever, and Karl Cobbe. Let's verify step by step. arXiv preprint arXiv:2305.20050 , 2023. - Haipeng Luo, Qingfeng Sun, Can Xu, Pu Zhao, Jianguang Lou, Chongyang Tao, Xiubo Geng, Qingwei Lin, Shifeng Chen, and Dongmei Zhang. Wizardmath: Empowering mathematical reasoning for large language models via reinforced evol-instruct. arXiv preprint arXiv:2308.09583 , 2023. - Qianli Ma, Haotian Zhou, Tingkai Liu, Jianbo Yuan, Pengfei Liu, Yang You, and Hongxia Yang. Let's reward step by step: Step-level reward model as the navigators for reasoning. arXiv preprint arXiv:2310.10080 , 2023. - OpenAI. GPT-4 technical report. CoRR , abs/2303.08774, 2023. doi: 10.48550/arXiv.2303.08774. URL https://doi.org/10.48550/arXiv.2303.08774 . - Sarah Pan, Vladislav Lialin, Sherin Muckatira, and Anna Rumshisky. Let's reinforce step by step. arXiv preprint arXiv:2311.05821 , 2023. - Joon Sung Park, Joseph O'Brien, Carrie Jun Cai, Meredith Ringel Morris, Percy Liang, and Michael S Bernstein. Generative agents: Interactive simulacra of human behavior. In Proceedings of the 36th Annual ACM Symposium on User Interface Software and Technology , pp. 1-22, 2023. - David Silver, Aja Huang, Chris J Maddison, Arthur Guez, Laurent Sifre, George Van Den Driessche, Julian Schrittwieser, Ioannis Antonoglou, Veda Panneershelvam, Marc Lanctot, et al. Mastering the game of go with deep neural networks and tree search. nature , 529(7587):484-489, 2016. - Yifan Song, Weimin Xiong, Dawei Zhu, Cheng Li, Ke Wang, Ye Tian, and Sujian Li. Restgpt: Connecting large language models with real-world applications via restful apis. corr, abs/2306.06624, 2023. doi: 10.48550. arXiv preprint arXiv.2306.06624 . - Maciej ´ Swiechowski, Konrad Godlewski, Bartosz Sawicki, and Jacek Ma´ ndziuk. Monte carlo tree search: A review of recent modifications and applications. Artificial Intelligence Review , 56(3): 2497-2562, 2023. - Hugo Touvron, Louis Martin, Kevin Stone, Peter Albert, Amjad Almahairi, Yasmine Babaei, Nikolay Bashlykov, Soumya Batra, Prajjwal Bhargava, Shruti Bhosale, et al. Llama 2: Open foundation and fine-tuned chat models. arXiv preprint arXiv:2307.09288 , 2023. - Jonathan Uesato, Nate Kushman, Ramana Kumar, Francis Song, Noah Siegel, Lisa Wang, Antonia Creswell, Geoffrey Irving, and Irina Higgins. Solving math word problems with process-and outcome-based feedback. arXiv preprint arXiv:2211.14275 , 2022. - Guanzhi Wang, Yuqi Xie, Yunfan Jiang, Ajay Mandlekar, Chaowei Xiao, Yuke Zhu, Linxi Fan, and Anima Anandkumar. Voyager: An open-ended embodied agent with large language models. arXiv preprint arXiv:2305.16291 , 2023a. - Peiyi Wang, Lei Li, Liang Chen, Feifan Song, Binghuai Lin, Yunbo Cao, Tianyu Liu, and Zhifang Sui. Making large language models better reasoners with alignment. arXiv preprint arXiv:2309.02144 , 2023b. - Peiyi Wang, Lei Li, Liang Chen, Dawei Zhu, Binghuai Lin, Yunbo Cao, Qi Liu, Tianyu Liu, and Zhifang Sui. Large language models are not fair evaluators. arXiv preprint arXiv:2305.17926 , 2023c. - Xuezhi Wang, Jason Wei, Dale Schuurmans, Quoc V. Le, Ed H. Chi, Sharan Narang, Aakanksha Chowdhery, and Denny Zhou. Self-consistency improves chain of thought reasoning in language models. In The Eleventh International Conference on Learning Representations, ICLR 2023, Kigali, Rwanda, May 1-5, 2023 . OpenReview.net, 2023d. URL https://openreview.net/ pdf?id=1PL1NIMMrw . - Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Brian Ichter, Fei Xia, Ed H. Chi, Quoc V. Le, and Denny Zhou. Chain-of-thought prompting elicits reasoning in large language models. In NeurIPS , 2022. URL http://papers.nips.cc/paper\_files/paper/2022/hash/ 9d5609613524ecf4f15af0f7b31abca4-Abstract-Conference.html . - Zeqiu Wu, Yushi Hu, Weijia Shi, Nouha Dziri, Alane Suhr, Prithviraj Ammanabrolu, Noah A Smith, Mari Ostendorf, and Hannaneh Hajishirzi. Fine-grained human feedback gives better rewards for language model training. arXiv preprint arXiv:2306.01693 , 2023. - Heming Xia, Tao Ge, Furu Wei, and Zhifang Sui. Lossless speedup of autoregressive translation with generalized aggressive decoding. arXiv preprint arXiv:2203.16487 , 2022. - Fei Yu, Anningzhe Gao, and Benyou Wang. Outcome-supervised verifiers for planning in mathematical reasoning. arXiv preprint arXiv:2311.09724 , 2023a. - Longhui Yu, Weisen Jiang, Han Shi, Jincheng Yu, Zhengying Liu, Yu Zhang, James T Kwok, Zhenguo Li, Adrian Weller, and Weiyang Liu. Metamath: Bootstrap your own mathematical questions for large language models. arXiv preprint arXiv:2309.12284 , 2023b. - Zheng Yuan, Hongyi Yuan, Chengpeng Li, Guanting Dong, Chuanqi Tan, and Chang Zhou. Scaling relationship on learning mathematical reasoning with large language models. arXiv preprint arXiv:2308.01825 , 2023. - Xiang Yue, Xingwei Qu, Ge Zhang, Yao Fu, Wenhao Huang, Huan Sun, Yu Su, and Wenhu Chen. Mammoth: Building math generalist models through hybrid instruction tuning. arXiv preprint arXiv:2309.05653 , 2023. - Yifan Zhang, Jingqin Yang, Yang Yuan, and Andrew Chi-Chih Yao. Cumulative reasoning with large language models. arXiv preprint arXiv:2308.04371 , 2023. - Lianmin Zheng, Wei-Lin Chiang, Ying Sheng, Siyuan Zhuang, Zhanghao Wu, Yonghao Zhuang, Zi Lin, Zhuohan Li, Dacheng Li, Eric Xing, et al. Judging llm-as-a-judge with mt-bench and chatbot arena. arXiv preprint arXiv:2306.05685 , 2023. - Xinyu Zhu, Junjie Wang, Lin Zhang, Yuxiang Zhang, Yongfeng Huang, Ruyi Gan, Jiaxing Zhang, and Yujiu Yang. Solving math word problems via cooperative reasoning induced language models. In Anna Rogers, Jordan Boyd-Graber, and Naoaki Okazaki (eds.), Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) , pp. 4471-4485, Toronto, Canada, July 2023. Association for Computational Linguistics. doi: 10.18653/v1/2023. acl-long.245. URL https://aclanthology.org/2023.acl-long.245 .