Research map / Large language models / Agents, code and math
MR-Ben: A Meta-Reasoning Benchmark for Evaluating System-2 Thinking in LLMs
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Introduces MR-Ben, a process-based benchmark of ~6k expert-annotated LLM reasoning solutions across nine subjects where models must judge solution correctness, locate the first error step, and explain the error, scored by the MR-Score metric.
Problem
Outcome-based reasoning benchmarks are saturating and do not assess whether LLMs can detect and explain errors in intermediate reasoning steps; existing process benchmarks are narrow in subject scope and size.
Approach
A meta-reasoning paradigm where LLMs act as teachers: given a question and a CoT solution, the model must judge correctness, identify the first error step, and give the error reason, with human experts annotating each solution across math, physics, biology, chemistry, medicine, logic, and coding. Performance is aggregated into an MR-Score combining MCC, first-error-step accuracy, and error-reason
Key results
- o1-preview achieves the highest average MR-Score of 58.9 (zero-shot), well above GPT-4-Turbo at 43.2 and most other models.
- Phi-3-3.8B (avg 11.9) outperforms much larger models such as Deepseek-Coder-33B (7.0), which the authors attribute to synthetic data.
- Self-refine improves Llama3-70B from 17.7 to 27.5 MR-Score but degrades smaller models; GPT-4-Turbo's agreement with human error-reason judgments is 92%.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use o1-preview instead of other evaluated models | LLM evaluation on MR-Ben · MR-Ben (Avg., k=0) | MR-Score: 58.9 | Helped |
| Add one-shot demonstrations (ablation) | GPT-4-Turbo · MR-Ben (Avg.) | MR-Score: 43.2 -> 44.7 | Helped |
| Add one-shot demonstrations (ablation) | Llama3-8B · MR-Ben (Avg.) | MR-Score: 11.7 -> 9.2 | Hurt |
| Increase few-shot examples from one to two (ablation) | Gemma-2B · MR-Ben (Avg.) | MR-Score: 0.2 -> 0.4 | Helped |
| Add three-round self-refinement (ablation) | Gemma-2B · MR-Ben (Avg.) | MR-Score: 0.1 -> 0.2 | Helped |
| Add three-round self-refinement (ablation) | Llama3-8B · MR-Ben (Avg.) | MR-Score: 11.7 -> 11.3 | Hurt |
| Add three-round self-refinement (ablation) | Llama3-70B · MR-Ben (Avg.) | MR-Score: 17.7 -> 27.5 | Helped |
| Add three-round self-refinement (ablation) | GPT-4-Turbo · MR-Ben (Avg.) | MR-Score: 43.2 -> 45.5 | Helped |
| Provide prior knowledge that the solution is incorrect (ablation) | Llama3-8B · MR-Ben (incorrect solutions) | Detection accuracy: 15.5 -> 26.4 | Helped |
| Provide prior knowledge that the solution is incorrect (ablation) | Llama3-70B · MR-Ben (incorrect solutions) | Reason accuracy: 9.1 -> 25.7 | Helped |
| Provide prior knowledge that the solution is incorrect (ablation) | GPT-4-Turbo · MR-Ben (incorrect solutions) | Detection accuracy: 40.9 -> 41.6 | Helped |
| Provide prior knowledge that the solution is incorrect (ablation) | GPT-4-Turbo · MR-Ben (incorrect solutions) | Reason accuracy: 37.9 -> 38.0 | No clear effect |
Surprising findings
- Providing a hint that the solution is incorrect can drastically change performance, and removing the step-wise analysis drops MR-Score to near zero.
- Strong code-specialized models (Deepseek-Coder) do not outperform general models on algorithmic reasoning, and the small Phi-3-3.8B beats much larger open models.
Limitations
- Limited to English-language questions and to three LLMs' solutions (GPT-3.5, Mistral-Medium, Claude 2), so error patterns may not generalize.
- Not applicable to holistic or creative subjects such as humanities or essay writing.
- Error-reason scoring relies on GPT-4-Turbo as proxy judge with 92% agreement with human annotations on a 100-sample check.
Experimental setup
Datasets: MMLU LogiQA MHPP
Benchmarks: MR-Ben
Metrics: MR-Score MCC accuracy of first error step accuracy of error reason
Baselines: GPT-4-Turbo GPT-4o o1-preview o1-mini Llama3-70B Qwen2-72B DeepSeek-V2-236B Phi-3-3.8B
Builds on: PRM800K MR-GSM8K BIG-Bench Mistake MMLU LogiQA
Closest papers by meaning
- MR-GSM8K: A Meta-Reasoning Benchmark for Large Language Model Evaluation (ICLR 2025 · similarity 0.91)
- Rethinking Thinking Tokens: LLMs as Improvement Operators (ICML 2026 · similarity 0.89)
- CogMath: Assessing LLMs' Authentic Mathematical Ability from a Human Cognitive Perspective (ICML 2025 · similarity 0.89)
- Unveiling the Magic of Code Reasoning through Hypothesis Decomposition and Amendment (ICLR 2025 · similarity 0.89)
- The Illusion of Thinking: Understanding the Strengths and Limitations of Reasoning Models via the Lens of Problem Complexity (NeurIPS 2025 · similarity 0.89)
- Off-Trajectory Reasoning: Can LLMs Collaborate on Reasoning Trajectory? (ICLR 2026 · similarity 0.88)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.