ProcessBench

3,400 competition math solutions with step-by-step human annotations evaluating Process Reward Models (PRMs) on identifying the exact first wrong step.26 Oct 2026

One row per model and compatible score measure. Mean = equal-weight arithmetic average of imported reports. Distinct configurations reported for the same model in a source table remain separate. Reports can overlap; the range is not a confidence interval. Unbundle to inspect each observation.
		Evidence	What the result shows
o1-mini →OpenAI	87.9 F1 points 1 report	ProcessBench · Average F1[1]ProcessBench · original comparison · Proprietary language models, prompted as Critic Models	Leads this displayed comparison. Source notes Source cell: 87.9
★ QwQ-32B-Preview →	71.5 F1 points 1 report	ProcessBench · Average F1[1]ProcessBench · original comparison · Open-source language models, prompted as Critic Models	#2 of 25 in this displayed comparison; 16.4 F1 points below the leader. Source notes Source cell: 71.5
GPT-4o-0806 →OpenAI	61.9 F1 points 1 report	ProcessBench · Average F1[1]ProcessBench · original comparison · Proprietary language models, prompted as Critic Models	#3 of 25 in this displayed comparison; 26 F1 points below the leader. Source notes Source cell: 61.9
Qwen2.5-72B-Instruct →Qwen	61.2 F1 points 1 report	ProcessBench · Average F1[1]ProcessBench · original comparison · Open-source language models, prompted as Critic Models	#4 of 25 in this displayed comparison; 26.7 F1 points below the leader. Source notes Source cell: 61.2
Llama-3.3-70B-Instruct →Meta	58 F1 points 1 report	ProcessBench · Average F1[1]ProcessBench · original comparison · Open-source language models, prompted as Critic Models	#5 of 25 in this displayed comparison; 29.9 F1 points below the leader. Source notes Source cell: 58.0
Qwen2.5-Math-7B-PRM800K (our trained) →Qwen	56.5 F1 points 1 report	ProcessBench · Average F1[1]ProcessBench · original comparison · Open-source Process Reward Models (PRMs)	#6 of 25 in this displayed comparison; 31.4 F1 points below the leader. Source notes Source cell: 56.5
Qwen2.5-Coder-32B-Instruct →Qwen	56.1 F1 points 1 report	ProcessBench · Average F1[1]ProcessBench · original comparison · Open-source language models, prompted as Critic Models	#7 of 25 in this displayed comparison; 31.8 F1 points below the leader. Source notes Source cell: 56.1
Llama-3.1-70B-Instruct →Meta	52.7 F1 points 1 report	ProcessBench · Average F1[1]ProcessBench · original comparison · Open-source language models, prompted as Critic Models	#8 of 25 in this displayed comparison; 35.2 F1 points below the leader. Source notes Source cell: 52.7
Qwen2.5-14B-Instruct →Qwen	52.2 F1 points 1 report	ProcessBench · Average F1[1]ProcessBench · original comparison · Open-source language models, prompted as Critic Models	#9 of 25 in this displayed comparison; 35.7 F1 points below the leader. Source notes Source cell: 52.2
Qwen2-72B-Instruct →Qwen	49.8 F1 points 1 report	ProcessBench · Average F1[1]ProcessBench · original comparison · Open-source language models, prompted as Critic Models	#10 of 25 in this displayed comparison; 38.1 F1 points below the leader. Source notes Source cell: 49.8
Qwen2.5-32B-Instruct →Qwen	49.3 F1 points 1 report	ProcessBench · Average F1[1]ProcessBench · original comparison · Open-source language models, prompted as Critic Models	#11 of 25 in this displayed comparison; 38.6 F1 points below the leader. Source notes Source cell: 49.3
Qwen2.5-Math-72B-Instruct →Qwen	45.5 F1 points 1 report	ProcessBench · Average F1[1]ProcessBench · original comparison · Open-source language models, prompted as Critic Models	#12 of 25 in this displayed comparison; 42.4 F1 points below the leader. Source notes Source cell: 45.5
Skywork-PRM-7B →	42.1 F1 points 1 report	ProcessBench · Average F1[1]ProcessBench · original comparison · Open-source Process Reward Models (PRMs)	#13 of 25 in this displayed comparison; 45.8 F1 points below the leader. Source notes Source cell: 42.1
Qwen2.5-Coder-14B-Instruct →Qwen	37.8 F1 points 1 report	ProcessBench · Average F1[1]ProcessBench · original comparison · Open-source language models, prompted as Critic Models	#14 of 25 in this displayed comparison; 50.1 F1 points below the leader. Source notes Source cell: 37.8
Skywork-PRM-1.5B →	36.4 F1 points 1 report	ProcessBench · Average F1[1]ProcessBench · original comparison · Open-source Process Reward Models (PRMs)	#15 of 25 in this displayed comparison; 51.5 F1 points below the leader. Source notes Source cell: 36.4
Qwen2.5-7B-Instruct →Qwen	32.6 F1 points 1 report	ProcessBench · Average F1[1]ProcessBench · original comparison · Open-source language models, prompted as Critic Models	#16 of 25 in this displayed comparison; 55.3 F1 points below the leader. Source notes Source cell: 32.6
Math-Shepherd-PRM-7B →	31.5 F1 points 1 report	ProcessBench · Average F1[1]ProcessBench · original comparison · Open-source Process Reward Models (PRMs)	#17 of 25 in this displayed comparison; 56.4 F1 points below the leader. Source notes Source cell: 31.5
Meta-Llama-3-70B-Instruct →Meta	29.1 F1 points 1 report	ProcessBench · Average F1[1]ProcessBench · original comparison · Open-source language models, prompted as Critic Models	#18 of 25 in this displayed comparison; 58.8 F1 points below the leader. Source notes Source cell: 29.1
RLHFlow-PRM-Mistral-8B →Mistral	28.4 F1 points 1 report	ProcessBench · Average F1[1]ProcessBench · original comparison · Open-source Process Reward Models (PRMs)	#19 of 25 in this displayed comparison; 59.5 F1 points below the leader. Source notes Source cell: 28.4
RLHFlow-PRM-Deepseek-8B →DeepSeek	26.6 F1 points 1 report	ProcessBench · Average F1[1]ProcessBench · original comparison · Open-source Process Reward Models (PRMs)	#20 of 25 in this displayed comparison; 61.3 F1 points below the leader. Source notes Source cell: 26.6
Qwen2.5-Math-7B-Instruct →Qwen	19.9 F1 points 1 report	ProcessBench · Average F1[1]ProcessBench · original comparison · Open-source language models, prompted as Critic Models	#21 of 25 in this displayed comparison; 68 F1 points below the leader. Source notes Source cell: 19.9
Qwen2-7B-Instruct →Qwen	13.6 F1 points 1 report	ProcessBench · Average F1[1]ProcessBench · original comparison · Open-source language models, prompted as Critic Models	#22 of 25 in this displayed comparison; 74.3 F1 points below the leader. Source notes Source cell: 13.6
Meta-Llama-3-8B-Instruct →Meta	11.1 F1 points 1 report	ProcessBench · Average F1[1]ProcessBench · original comparison · Open-source language models, prompted as Critic Models	#23 of 25 in this displayed comparison; 76.8 F1 points below the leader. Source notes Source cell: 11.1
Qwen2.5-Coder-7B-Instruct →Qwen	6.7 F1 points 1 report	ProcessBench · Average F1[1]ProcessBench · original comparison · Open-source language models, prompted as Critic Models	#24 of 25 in this displayed comparison; 81.2 F1 points below the leader. Source notes Source cell: 6.7
Llama-3.1-8B-Instruct →Meta	5.1 F1 points 1 report	ProcessBench · Average F1[1]ProcessBench · original comparison · Open-source language models, prompted as Critic Models	#25 of 25 in this displayed comparison; 82.8 F1 points below the leader. Source notes Source cell: 5.1

Sources

ProcessBench: Identifying Process Errors in Mathematical Reasoning — published result table ↗
Zheng, Chujie, Zhang, Zhenru, Zhang, Beichen, Lin, Runji, Lu, Keming, Yu, Bowen et al. · 2024
All 9 authors
Zheng, Chujie, Zhang, Zhenru, Zhang, Beichen, Lin, Runji, Lu, Keming, Yu, Bowen, Liu, Dayiheng, Zhou, Jingren, Lin, Junyang
Dates & evaluation details
ProcessBench · Average F1 · original comparison · 2026-10-26
Published model comparison, 2024. ProcessBench · original comparison. Mean of subset F1 scores combining correct-solution recognition and first-error identification. This is a process-error assessment score, not problem-solving accuracy. Table 3; source precision and any reported uncertainty retained. These are source-reported observations, not evaluations run by RLTasks.

TRACE THE SCORE · Per-model notes use published breakdowns where available.

What it measures

3,400 competition math solutions with step-by-step human annotations evaluating Process Reward Models (PRMs) on identifying the exact first wrong step.

Reading the measurement

F1 balances precision and recall. Check both: a system can miss relevant evidence or include unsupported material, and those errors have different consequences. Keep the matching rule, averaging policy and denominator fixed.

Published evaluation

Mean of subset F1 scores combining correct-solution recognition and first-error identification. This is a process-error assessment score, not problem-solving accuracy.

Where this fits in an evaluation suite

Format compliance, preference, truthfulness, and safety require separate measurements. Keep desirable writing style from masking factual errors.

Start with a capability hypothesis, then choose a second test with a different failure mode. Keep the checkpoint, reasoning effort, agent tools, sampling policy, and scoring procedure fixed when measuring a training intervention. Record failures and missing outputs in the denominator.

A reproducible comparison

Mean of subset F1 scores combining correct-solution recognition and first-error identification. This is a process-error assessment score, not problem-solving accuracy.

Publish the exact run configuration, item coverage, score units, and available uncertainty. If a source supplies only an aggregate or a single run, preserve that limit. A new model name or a rounded score alone does not establish a capability improvement.

PRIMARY EVIDENCE & FURTHER READING

ProcessBench: Identifying Process Errors in Mathematical Reasoning — published result table ↗
Zheng, Chujie, Zhang, Zhenru, Zhang, Beichen, Lin, Runji, Lu, Keming, Yu, Bowen et al. · 2024
All 9 authors
Zheng, Chujie, Zhang, Zhenru, Zhang, Beichen, Lin, Runji, Lu, Keming, Yu, Bowen, Liu, Dayiheng, Zhou, Jingren, Lin, Junyang

Reviewed 2026-10-26. Explanation and interpretation by RLTasks. Benchmark ownership remains with its original authors.

Continue exploring

All instructions & behavior benchmarks →

Sources

What it measures

Reading the measurement

Published evaluation

Where this fits in an evaluation suite

A reproducible comparison

Continue exploring

AA-PersonalityBench

IFBench

IFEval

WildBench