Mr-Ben: A Comprehensive Meta-Reasoning Benchmark for Large Language Models

1Chinese University of Hong Kong, 2University of Cambridge, 3University of Edinburgh, 4City University of Hong Kong, 5Tsinghua University, 6University of Texas at Austin, 7University of Hong Kong, 8Nanyang Technological University, 9Massachusetts Institute of Technology
Corresponding author


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, it has been increasingly challenging to evaluate the reasoning capability of LLMs. Concretely, existing outcome-based benchmarks begin to saturate and become less sufficient to monitor the progress. To this end, 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. Mr.Ben is a comprehensive benchmark comprising 5,975 questions collected from human experts, covering various subjects such as physics, chemistry, logic, coding, and more. By incorporating this approach, Mr.Ben facilitates a multidimensional evaluation of LLM reasoning abilities. We conducted an extensive assessment of open-source and closed-source LLMs using Mr.Ben, which revealed previously unidentified limitations and weaknesses in their meta-reasoning capabilities across different tasks.



Figure 1: This is the illustration of the dataset creation pipeline of Mr-Ben.


Figure 2: Overview of the evaluation paradigm and representative examples in Mr-Ben.

Evaluation Results

Model #params Avg Mr-Score (k=0) Avg Mr-Score (k=1) Cost-Per-Million-Tokens
Closed-source Model
Claude3-Haiku - 4.4 3.1 Input:$0.25/Output:$1.25
GPT-3.5-Turbo - 4.0 5.5 Input:$1.0/Output:$2.0
Doubao-pro-4k - 8.8 11.6 Input:$0.11/Output:$0.28
Mistral-Large - 21.3 23.8 Input:$4.0/Output:$12.0
Yi-Large - 32.2 32.3 Input:$3.0/Output:$3.0
Moonshot-v1-8k - 32.5 33.0 Input:$1.65/Output:$1.65
Claude3.5-Sonnet - 33.5 37.6 Input:$3.0/Output:$15.0
Zhipu-GLM-4 - 38.7 39.4 Input:$13.78/Output:$13.78
GPT-4-Turbo - 43.2 44.7 Input:$10.0/Output:$30.0
Open-source models Small
Qwen1.5 1.8B 0.0 0.0 N/A
Gemma 2B 0.1 0.2 N/A
Qwen2 1.5B 2.1 5.4 N/A
Phi-3-Mini 3.8B 11.9 11.0 N/A
Open-source models medium
GLM-4 9B 6.7 2.1 N/A
Deepseek-llm 7B 3.7 3.6 N/A
Deepseek-Coder 33B 7.0 6.3 N/A
Deepseek-Coder 7B 10.2 10.2 N/A
Llama-3 8B 12.2 9.8 N/A
Yi-1.5 9B 10.2 12.6 N/A
Open-source models large
Qwen-1.5 72B 11.5 13.3 N/A
Deepseek-llm 67B 15.2 16.5 N/A
Llama-3 70B 19.2 20.2 N/A
Deepseek-coder-v2 236B 25.0 31.7 N/A
Deepseek-chat-v2 236B 30.2 32.3 N/A
Qwen-2 72B 33.3 34.2 N/A

Table 1: Evaluation results on Mr.Ben.
Note: All models used in our experiments are instruction-finetuned versions, although this is not indicated in their abbreviated names


        author       = {Zhongshen Zeng and Yinhong Liu and Yingjia Wan and Jingyao Li and Pengguang Chen and Jianbo Dai and Yuxuan Yao and Rongwu Xu and Zehan Qi and Wanru Zhao and Linling Shen and Jianqiao Lu and Haochen Tan and Yukang Chen and Hao Zhang and Zhan Shi and Bailin Wang and Zhijiang Guo and Jiaya Jia},
        title        = {MR-BEN: A Comprehensive Meta-Reasoning Benchmark for Large Language Models},
        journal      = {CoRR},
        volume       = {abs/2406.13975},
        year         = {2024},
        url          = {},
        eprinttype    = {arXiv},
        eprint       = {2406.13975}