MR-Ben: A Meta-Reasoning Benchmark for Evaluating System-2 Thinking in LLMs

Large language models (LLMs) have shown increasing capability in problem-solving and decision-making, largely based on the step-by-step chain-of-thought reasoning processes. However, evaluating these reasoning abilities has become increasingly challenging. Existing outcome-based benchmarks are beginning to saturate, becoming less effective in tracking meaningful progress. To address this, we present a process-based benchmark MR-Ben that demands a meta-reasoning skill, where LMs are asked to locate and analyse potential errors in automatically generated reasoning steps. Our meta-reasoning paradigm is especially suited for system-2 slow thinking, mirroring the human cognitive process of carefully examining assumptions, conditions, calculations, and logic to identify mistakes.MR-Ben comprises 5,975 questions curated by human experts across a wide range of subjects, including physics, chemistry, logic, coding, and more. Through our designed metrics for assessing meta-reasoning on this benchmark, we identify interesting limitations and weaknesses of current LLMs (open-source and closed-source models). For example, with models like the o1 series from OpenAI demonstrating strong performance by effectively scrutinizing the solution space, many other state-of-the-art models fall significantly behind on MR-Ben, exposing potential shortcomings in their training strategies and inference methodologies.

Paper

Similar papers

Peer review

Reviewer JijF5/10 · confidence 3/52024-07-10

Summary

The paper introduces MR.BEAN, a benchmark designed to evaluate the meta-reasoning capabilities of large language models (LLMs). This benchmark focuses on the models' ability to detect and correct errors in reasoning steps, addressing the limitations of existing outcome-based benchmarks.

Strengths

- The paper is well-articulated, providing clear examples that effectively demonstrate the dataset's composition and intended utility. - This work contributes a new benchmark for identifying and correcting errors in reasoning steps rather than just the final outcomes. - The benchmark covers diverse topics and is paired with reasonable metrics. - A comprehensive set of experiments is conducted that can tell performance disparities among various LLMs.

Weaknesses

- The $ACC_{reason}$ metric shows limitations in consistency. Its dependency on the judgments of different LLMs or human evaluators could lead to variability in scoring. - The weights assigned to different metrics might need recalibration when new models are tested or when the validation set is updated. - A detailed report on each metric's individual performance and impact is lacking.

Questions

- Since the MR metric is newly proposed, would it be reasonable to incorporate human evaluations to support its credibility and relevance? By comparing the model outputs against human judgments on at least a sampled subset, it can provide empirical evidence on how well the MR metric aligns with human reasoning and offer insights into the absolute performance of models.

Rating

5

Confidence

3

Soundness

2

Presentation

3

Contribution

2

Limitations

n/a

Reviewer AKte5/10 · confidence 4/52024-07-11

Summary

This paper introduces a new benchmark for evaluating the reasoning capabilities of large language models (LLMs). Current methods primarily focus on final outcomes and not sufficiently capture the intricacies of the reasoning process. To address this issue, this paper propose MR.BEAN, a process-based benchmark that demands meta-reasoning skills from LLMs, requiring models to identify and analyze potential errors in automatically generated reasoning steps. The authors conducted an extensive analysis of various LLMs using MR.BEAN, revealing significant limitations and previously unidentified weaknesses in their reasoning abilities. They found that while many LLMs can generate correct answers, they struggle to pinpoint and correct errors in the reasoning process. The paper also discusses the potential for improving reasoning abilities through techniques like the use of high-quality synthetic data.

Strengths

1) The paper introduces a novel benchmark, MR.BEAN, which focuses on meta reasoning—a higher-order thinking skill. This design pushes beyond traditional outcome-based evaluations to assess the reasoning process itself. And MR.BEAN covers a wide range of subjects, including physics, chemistry, logic, coding, and more. 2) The benchmark's questions and error analyses are curated and annotated by human experts, ensuring a high level of quality and relevance in the evaluation process. 3) The paper's evaluation of various LLMs reveals previously unidentified weaknesses in their reasoning abilities, providing valuable insights for researchers and developers. The benchmark's application to a diverse array of models, from small to large, open-source and closed-source, provides a broad comparative analysis that can inform future development in AI reasoning.

Weaknesses

1) The paper may not provide a thorough comparison between the model's automatic annotations and human annotations. Without such validation, it is challenging to assess the reliability and accuracy of the model-generated annotations. 2) The concept of meta-reasoning typically involves instructing models on how to reason, which may include decision-making during the reasoning process. The paper primarily analyzes the reasoning steps after they have been generated, which might not fully align with the proactive aspect of meta-reasoning. 3) The paper could provide more transparency regarding the annotation process, such as the total number of annotators involved, the total time spent on annotations, and the average number of samples annotated per person per day. This information is crucial for understanding the scalability and efficiency of the annotation process. 3) The paper may not fully address how the dataset can be used to validate an LLM's reasoning capabilities on new, unseen examples. Understanding the generalizability of the findings is essential for assessing the true breadth of an LLM's reasoning skills. 4) The paper could benefit from including specific test cases using GPT-4 to demonstrate the benchmark's effectiveness in identifying strengths and weaknesses in state-of-the-art models. This would provide a clearer picture of how GPT-4 performs against the benchmark and highlight the benchmark's utility.

Questions

See weaknesses.

Rating

5

Confidence

4

Soundness

3

Presentation

3

Contribution

2

Limitations

See weaknesses.

Reviewer TVR66/10 · confidence 4/52024-07-12

Summary

This paper introduces MR.BEAN, a comprehensive benchmark for evaluating meta-reasoning capabilities of large language models (LLMs). Comprising 6,006 questions across various subjects including physics, chemistry, logic, coding, and more, MR.BEAN requires LLMs to analyze and correct errors in automatically generated reasoning steps. The benchmark is meticulously constructed using a three-step annotation process involving answer correctness evaluation, error step identification, and error reason analysis. The main contributions of this work are: 1. A novel, large-scale benchmark for meta-reasoning evaluation covering diverse subjects and reasoning types. 2. A rigorous methodology for creating and annotating meta-reasoning questions, ensuring high-quality data. 3. Comprehensive evaluation of 15 LLMs on the benchmark, revealing limitations and weaknesses in their reasoning abilities.

Strengths

1. Comprehensive and well-organized dataset: MR.BEAN covers many subjects (p.3, Table 1) and offers a broad assessment of LLM meta-reasoning capabilities across diverse domains. The paper is structured clearly, with detailed explanations of the dataset creation process, evaluation metrics, and experimental results. 2. Novel meta-reasoning focus: By requiring LLMs to identify and correct errors in reasoning, MR.BEAN offers a unique perspective on evaluating AI reasoning capabilities, going beyond traditional outcome-based assessments. 3. Extensive empirical study and analysis: The authors evaluate different LLMs on MR.BEAN and have tested different prompting methods to comprehensively analyze current model capabilities and reveal interesting limitations in their reasoning abilities.

Weaknesses

1. Validation of meta-reasoning specificity: While the paper describes the evaluation metrics in detail (p.5, l.185-197), it is still not clear such metrics measures meta-reasoning abilities rather than general language understanding or domain knowledge. 2. Prompt sensitivity: The paper doesn't adequately address the potential impact of prompt design on the generated solutions and errors. Given that LLMs are known to be sensitive to prompt wording, this could significantly affect the nature and distribution of errors in the dataset. 3. Lack of annotation details: While the paper mentions a three-stage annotation process (p.4, l.116-135), it lacks specific details on annotator qualifications, training, and inter-annotator agreement rates. This information is crucial for assessing the reliability and consistency of the annotations.

Questions

1. Can you provide more details on the annotation process, including inter-annotator agreement rates and resolution of disagreements? 2. What steps were taken to ensure that MR.BEAN specifically measures meta-reasoning abilities and not merely language understanding or domain knowledge? 3. How sensitive is the dataset to the specific prompts for generating solutions and errors? Did you experiment with different prompt formulations, and if so, how did this affect the resulting dataset?

Rating

6

Confidence

4

Soundness

3

Presentation

3

Contribution

3

Limitations

Yes

Reviewer SioJ6/10 · confidence 4/52024-07-14

Summary

This paper proposes a meta-reasoning benchmark for evaluating the solutions generated by a large-language model (LLM) to shift the focus more to process-based evaluation of an LLM's reasoning abilities rather than outcome-based evaluation. *Evaluation*: On a variety of question-solution pairs, the model is asked to score its own generated solution for correctness, to identify the first erroneous step, and if these are correctly identified, the reasoning of it solution is evaluated by GPT-4. These three are used to compute three metrics respectively-- the MCC for a binary classification of the solution correctness, the ACC_step to judge the number of correctly predicted first-error steps as compared to the total number of incorrect solutions, and the ACC_reason to judge the number of correctly predicted first-error steps and error-reasons as compared to the total number of incorrect solutions. These three are combined to form an MR-score that is used to compare different models. *Dataset*: The question-solution pairs (~6k in all) themselves are formed such that the questions are sampled from MMLU (arithmetic reasoning), LogiQA(logical reasoning), and MHPP (code-based reasoning) and the solutions are generated via chain-of-thought prompting by GPT-3.5-Turbo-0125, Claude-2, and Mistral-Medium. The annotations for these question-solution pairs are annotated by human annotators-- these annotations of solution correctness, first error-step, and the error-reason are used for evaluation.

Strengths

- The authors highlight the importance of more carefully evaluating the solutions generated by LLMs to complex problems rather than just comparing their accuracy which is a notable contribution to the body of work of improving evaluation of LLMs. - The annotation process followed for experimental evaluation is thorough and multi-step for quality assurance - The authors raise an interesting question about problem solving dynamics, which can be potentially explored further in future work on LLMs

Weaknesses

These questions/remarks might also have arisen due to my lack of proper understanding, so I am willing to increase my score if these can be clarified: - My main issue is with the mixing of reasoning based and accuracy based evaluation into a singular score. What is the reasoning behind combining the three metrics (MCC, ACC_step, and ACC_reason) into a single score (MR-Score)? Does this not affect the motivating insight that models might have a high MCC but a super low ACC_step and an even lower ACC_reason? Is there somewhere the three metrics can be viewed individually also so it's clearer how the models vary along the three-- right now, it is not clear to me by looking at the MR-scores in Table 2 how the 15 models actually fare when it comes to identifying the correct error-reason, especially when the MR-scores for the models are in the same ballpark. Writing: - (pre-existing work): it's not clear to me what the real differentiating aspect of MR. Bean is vis-a-vis MR-GSM8K and MR-Math. MR-GSM8K and MR-Math also consider the first-error reason? So, is it that they are math-reasoning based, and MR. Bean now just combines multiple datasets to essentially follow the same evaluation pipeline? If so, this should be clearly mentioned. They don't just go a step further, that is the furthest step gone to, even considering this work (if that is indeed the case). - The abstract makes it seem like the MR-score is a metric designed by the authors in this work, although it's already been proposed "*Through our designed metrics*" which is somewhat over-selling the work. - (minor) typo: A.2 Negative Societal Impacts - The connection between task difficulty and reasoning capability can be explored further-- I am assuming ACC_reason on more difficult tasks would be comparably worse than the MCC, which would be the actual assessment of reasoning capabilities. Also, from Fig 2, it seems that Logic based tasks are the most difficult, then wouldn't it be fair to include them in the comparison too or perform the comparison of task difficulty not just based on high-school/college, but also take into consideration how the model actually interprets the task difficulty based on MR-score? Experiments: - what is the main takeaway on self-refine prompting since it seems to be in contrast to existing observations? How does the shift from incorrect to correct predictions depend on the model family and task? Is it seen only across MR-Scores or also in the ACC_reason (which should probably be affected more by model size?)? - Is the difficulty of logic-based reasoning also because they contain the longest questions? is there any impact of question length on the model's reasoning ability, if not solution length?

Questions

Some questions for clarification: - I might have missed this, so in Table 2: is the meaning of k the number of shots of prompting the model gets? If so, it would help to mention that in the caption since k is not mentioned in the text anywhere else. what should be the take-away on self-correction? - Is the difference between the Self-Refine prompting (section 6.2) and Solution Correctness Prior (section 6.3) the absence/presence of of the ground truth (resp.)? How do the two compare with each other?

Rating

6

Confidence

4

Soundness

3

Presentation

3

Contribution

2

Limitations

- The proposed benchmark also relies on step-wise solutions to evaluate reasoning which reduces its applicability and novelty as compared to existing work.

Authorsrebuttal2024-08-07

Continual of Rebuttal to Reviewer SioJ

> **W4: Takeaways on self-refine prompting** The self-refine prompting experiment was designed to unveil if LLMs are capable of discovering their own reasoning flaws and effectively rectifying them. The result was indeed intriguing and therefore **we have decomposed the behavior of models across tasks in Appendix D and visualized it in Figure 9**. We summarize our observation below: - **Small Models** like Gemma-2B are too limited to perform effective self-reflection. - **Competent Models** like GPT4-Turbo are confident in their initial decisions, hardly switching the decisions during self-reflection. - **Intermediate Models** like Llama3-70B exhibit substantial changes during self-reflection, indicating a lack of consistency in their decisions. However, its change of decisions from incorrect to correct happens to be significantly higher in locating the first error step than examining solution correctness and explaining the error reason, therefore boosting the overall MR-Score by a large margin. We believe the lack of consistency does not necessarily indicate a more robust or advanced reasoning ability, despite the increase of the evaluation results. - **Conclusion**: Our results support the observation that LLMs generally lack effective self-refinement capabilities [1][2]. Ref: [1] Large Language Models Cannot Self-Correct Reasoning Yet. 2024 [2] LLMs cannot find reasoning errors, but can correct them given the error location. 2024 > **W5: Clarifications** - **Clarification of 'k'**: 'k' as in ‘k-shot’, represents the number of demonstrations in a prompt. - **Typos**: Thank you for pointing out the typo in Appendix A2. We will correct them in the revised manuscript. > **Q: Difference between Self-Refine and Solution Correctness Prior** Yes, the difference is indeed the absence/presence of the ground truth, specifically: - **Self-Refine**: LLMs are asked to generate a three-step reasoning process, where the LLMs first answer the question directly and then self-critic on their own response. Finally, the LLMs generate the final refined response based on its original response and critics. - **Solution-Correctness Prior**: The information that the provided solution is incorrect is included as part of the input prompt. LLMs are only asked to identify the first error step and explain the reason for it. > **Limitation: The proposed benchmark relies on stepwise solutions to evaluate reasoning which reduces its applicability and novelty as compared to existing work.** We would like to clarify that our evaluation mechanism ensures the robustness of the process. LLMs are tasked to determine the solution's correctness, first error step, and error reason. Even if the model made correct predictions on the solution correctness and first error step via a flawed reasoning process, such a process will generally lead to incorrect/incomplete error reasons. **When the proxy scoring language models (e.g. GPT4-Turbo) are presented with the question/solution pair and the detailed error reasons provided by several annotators from different perspectives, the flawed error reasons generated by the evaluating models will be easy to diagnose via contrast.** This is supported by the high author-model agreement rates(92% as written in line 208) in the automatic error reason scoring process.

Reviewer SioJ2024-08-14

reply to authors

- Based on the attached pdf with the individual scores and the authors' explanation, I am convinced the MR score rightly captures process based evaluation as well (for eg, it shows a higher MR score when $ACC_{reason}$ score is higher whereas MCC is much lower). This was my biggest concern. - However, based on the authors' response, now it is clearer to me that MR.Bean is indeed an incremental improvement over MR-GSM8K and MR-Math, employing the **same metric** and the **same evaluation protocol** on a larger dataset. In my opinion, that does not necessitate a new paper. The authors have also not edited the abstract that says *"Through our designed metrics"* to reflect that the MR score is a pre-existing metric which they use. Even so, I have already provided a positive score, which I cannot increase.

Authorsrebuttal2024-08-14

Reply to Reviewer SioJ

Dear Reviewer SioJ: We are happy to see that your biggest concern is addressed. Since we are not allowed to edit the paper submitted during the rebuttal period, we will make sure to include that in our next iteration of our paper. We are truly grateful for your kind review and detailed comments! Wish you a great day ahead : ) Sincerely, Authors

Reviewer TVR62024-08-11

Response to the rebuttal

Thanks for your detailed response; I would like to keep my rating positive for this paper.

Authorsrebuttal2024-08-11

Response to the reviewer TVR6

Dear reviewer TVR6, thanks for your kind review and insightful questions. We are happy that you find the reply helpful. We wish you all the best : )

Authorsrebuttal2024-08-11

Reply to reviewer JiJF

Dear reviewer JiJF: Thanks for your kind encourgement and we will make sure to include the clarification enlighted by your comments into our work. If you believe your concerns are addressed by our explaination, we would be very appreciative if you can kindly consider updating the ratings for us. We wish you a wonderful day ahead. Sincerely Authors

Program Chairsdecision2024-09-25

Decision

Accept (poster)

© 2026 NYSGPT2525 LLC