PRMBench
Models identify errors and redundant steps in mathematical solutions. PRMScore combines positive and negative F1 scores across simplicity, soundness, and sensitivity tests.
The benchmark contains 6,216 problems with 83,456 step labels. The public dataset repository retains the PRMBench_Preview name.
- Made by
- Fudan University, Shanghai AI Laboratory, Soochow University, Stony Brook University, and Chinese University of Hong Kong
- Released
- January 2025
- Kind
- Word problems
- Level
- University
- Scored
- Checked automatically
- Answers
- Short answer
- Tasks
- All public
- Measures
- Capability
- Subject
- Mathematics
- Language
- English
- Interaction
- Single turn
- Refresh
- One fixed set
- Run by
- Anyone
- Strongest model tested
- o1-mini #142 on the capability index
- Last model added
- February 2025
- Scores refreshed
- November 20th, 2026
In the authors’ words
Process-level Reward Models (PRMs) are crucial for complex reasoning and decision-making tasks, where each intermediate step plays an important role in the reasoning process. Since language models are prone to various types of errors during the reasoning process, PRMs are required to possess nuanced capabilities for detecting various implicit error types in real-world scenarios. However, current benchmarks primarily focus on step correctness, failing to evaluate PRMs' performance systematically. To address this gap, we introduce PRMBench, a process-level benchmark specifically designed to assess the fine-grained error detection capabilities of PRMs. PRMBench comprises 6,216 carefully designed problems and 83,456 step-level labels, evaluating models across multiple dimensions, including simplicity, soundness, and sensitivity. In our experiments on 15 models, spanning both open-source PRMs and closed-source large language models prompted as critic models, we uncover significant weaknesses in current PRMs. These findings underscore the challenges inherent in process-level evaluation and highlight key directions for future research. We hope PRMBench can be a robust bench for advancing research on PRM evaluation and development.
o1-mini leads with 68.8 among models released by February 2025 o1-mini leads with 68.8 among models released by February 2025
Best score per entry from the benchmark's paper, read on November 20th, 2026. A paper’s table doesn’t add newer models. Where a model ran with several settings or harnesses, we keep its best. Names in gray aren't in our model list: fine-tunes, agent systems, products and older models. They don't count toward leaders. This view shows only models we track, each at its best.
The newest model with a score here came out in February 2025, and nobody has tested a newer one. Among the 20 highest-ranked models on the capability index, these have no score yet:
- GPT-6.2 Astra
- Claude Fable 5.2
-
Gemini 4 Carbon
- GPT-6 Astra
- and 16 more
- 68.8 Gemini-2.0-thinking-exp-1219
- 68.8 o1-mini
- 68.2 Qwen2.5-Math-PRM-72B
- 67.8 DeepSeek-R1 (snapshot unspecified)
- 66.8 GPT-4o (snapshot unspecified)
- 66.0
Gemini 2.0 Flash
- 65.5 Qwen2.5-Math-PRM-7B
- 65.3 Pure-PRM-7B
- 65.1 Skywork-PRM-7B
- 63.6
QwQ-32B-Preview
- 61.1 Skywork-PRM-1.5B
- 60.5 ReasonEval-34B
- 60.1 ReasonEval-7B
- 57.5 R1-Distill-Llama3.1-70B
- 57.4 Qwen2.5-Math-72B
- 54.4 RLHFlow-PRM-Mistral-8B
- 54.2 MATHMinos-Mistral-7B
- 54.2 RLHFlow-PRM-Deepseek-8B
- 52.6 R1-Distill-Qwen-7B
- 52.0 Llemma-PRM800k-7B
- 50.5 Llemma-MetaMath-7B
- 50.3 Llemma-oprm-7B
- 49.7 MetaMath-7B
- 49.4 MetaMath-13B
- 47.0 MathShepherd-Mistral-7B
- 1 Gemini-2.0-thinking-exp-1219 prompted as critic 68.8
- 2 o1-mini prompted as critic 68.8
- 3 Qwen2.5-Math-PRM-72B 68.2
- 4 DeepSeek-R1 (snapshot unspecified) prompted as critic 67.8
- 5 GPT-4o (snapshot unspecified) prompted as critic 66.8
- 6
Gemini 2.0 Flash Feb 2025, prompted as critic 66.0
- 7 Qwen2.5-Math-PRM-7B 65.5
- 8 Pure-PRM-7B 65.3
- 9 Skywork-PRM-7B 65.1
- 10
QwQ-32B-Preview prompted as critic 63.6
- 11 Skywork-PRM-1.5B 61.1
- 12 ReasonEval-34B 60.5
- 13 ReasonEval-7B 60.1
- 14 R1-Distill-Llama3.1-70B prompted as critic 57.5
- 15 Qwen2.5-Math-72B prompted as critic 57.4
- 16 RLHFlow-PRM-Mistral-8B 54.4
- 17 MATHMinos-Mistral-7B 54.2
- 18 RLHFlow-PRM-Deepseek-8B 54.2
- 19 R1-Distill-Qwen-7B prompted as critic 52.6
- 20 Llemma-PRM800k-7B 52.0
- 21 Llemma-MetaMath-7B 50.5
- 22 Llemma-oprm-7B 50.3
- 23 MetaMath-7B prompted as critic 49.7
- 24 MetaMath-13B prompted as critic 49.4
- 25 MathShepherd-Mistral-7B 47.0
Source: The PRMBench paper
Source: The PRMBench paper
| # | Entry | Score | Setting | Date | Source |
|---|---|---|---|---|---|
| 1 | Gemini-2.0-thinking-exp-1219 closed weights | 68.8 | prompted as critic | Official leaderboard | |
| 2 | o1-mini | 68.8 | prompted as critic | Sep 2024 | Official leaderboard |
| 3 | Qwen2.5-Math-PRM-72B open weights | 68.2 | Official leaderboard | ||
| 4 | DeepSeek-R1 (snapshot unspecified) open weights | 67.8 | prompted as critic | Official leaderboard | |
| 5 | GPT-4o (snapshot unspecified) closed weights | 66.8 | prompted as critic | Official leaderboard | |
| 6 | | 66.0 | prompted as critic | Feb 2025 | Official leaderboard |
| 7 | Qwen2.5-Math-PRM-7B open weights | 65.5 | Official leaderboard | ||
| 8 | Pure-PRM-7B open weights | 65.3 | Official leaderboard | ||
| 9 | Skywork-PRM-7B open weights | 65.1 | Official leaderboard | ||
| 10 | | 63.6 | prompted as critic | Nov 2024 | Official leaderboard |
| 11 | Skywork-PRM-1.5B open weights | 61.1 | Official leaderboard | ||
| 12 | ReasonEval-34B open weights | 60.5 | Official leaderboard | ||
| 13 | ReasonEval-7B open weights | 60.1 | Official leaderboard | ||
| 14 | R1-Distill-Llama3.1-70B open weights | 57.5 | prompted as critic | Official leaderboard | |
| 15 | Qwen2.5-Math-72B open weights | 57.4 | prompted as critic | Official leaderboard | |
| 16 | RLHFlow-PRM-Mistral-8B open weights | 54.4 | Official leaderboard | ||
| 17 | MATHMinos-Mistral-7B open weights | 54.2 | Official leaderboard | ||
| 18 | RLHFlow-PRM-Deepseek-8B open weights | 54.2 | Official leaderboard | ||
| 19 | R1-Distill-Qwen-7B open weights | 52.6 | prompted as critic | Official leaderboard | |
| 20 | Llemma-PRM800k-7B open weights | 52.0 | Official leaderboard | ||
| 21 | Llemma-MetaMath-7B open weights | 50.5 | Official leaderboard | ||
| 22 | Llemma-oprm-7B open weights | 50.3 | Official leaderboard | ||
| 23 | MetaMath-7B open weights | 49.7 | prompted as critic | Official leaderboard | |
| 24 | MetaMath-13B open weights | 49.4 | prompted as critic | Official leaderboard | |
| 25 | MathShepherd-Mistral-7B open weights | 47.0 | Official leaderboard |
| # | Model | Score | Setting | Date | Source |
|---|---|---|---|---|---|
| 1 | o1-mini | 68.8 | prompted as critic | Sep 2024 | Official leaderboard |
| 2 | | 66.0 | prompted as critic | Feb 2025 | Official leaderboard |
| 3 | | 63.6 | prompted as critic | Nov 2024 | Official leaderboard |
Rankings by part
We read 3 parts from the benchmark’s paper: Simplicity, Soundness and Sensitivity. Each ranks models on one part of the test. o1-mini leads every one of them. Arrows show how far a model moved from its overall rank among the models we track.
Simplicity
3 models we track have a score here.
- 1 o1-mini 64.6
- 2
Gemini 2.0 Flash Feb 2025 62.7
- 3
QwQ-32B-Preview 56.4
Soundness
3 models we track have a score here.
- 1 o1-mini 72.1
- 2
QwQ-32B-Preview ↑1 68.2
- 3
Gemini 2.0 Flash Feb 2025 ↓1 67.3
Sensitivity
3 models we track have a score here.
- 1 o1-mini 75.5
- 2
Gemini 2.0 Flash Feb 2025 75.4
- 3
QwQ-32B-Preview 73.5
Number one over time
Lab claims count from the day the card came out, independent runs from the model’s release.
Source: The PRMBench paper
Scores over time
All the scores we have, one dot each, so a model tested at several settings appears more than once.
Source: Official PRMBench leaderboard
No lab model card reports it yet
We read the headline results of 540 launch posts from 57 labs, and none of them list this benchmark.
More in Math
- AIME 2026 In use Models solve the 30 integer-answer problems from the 2026 AIME I and II math competitions. Scored by exact answer match. GPT-5.5 100%
- IMO-AnswerBench In use 400 olympiad problems with short verifiable answers, robustified against memorization. Scored by answer match. GLM-5.2 91%
- HMMT 2026 In use Models solve the 33 final-answer problems from the February 2026 Harvard-MIT Mathematics Tournament. Scored by exact answer match. GPT-5.5 98.5%
- PolyMath In use Models solve 9,000 parallel math problems in 18 languages across four difficulty levels, from school math to frontier math. Final answers are checked.
Qwen3.7-Max 86.5%
- MathArena (Apex, IMO, USAMO) In use Evaluates models on newly released math competitions after release. Final-answer or human/LLM-judged proof scores. Claude Opus 4.8 81.3%
- MathArena Apex Shortlist In use Models solve 47 hard final-answer problems that MathArena selected from 2025 competitions, where the best models scored about 50% at selection. Scored by accuracy over repeated attempts. GPT-5.5 98.4%
Data audit
8 of 9 checks pass, 1 warning. Look first at the Newest score check. The newest came 21 months ago, in February 2025, for Gemini 2.0 Flash. Every benchmark
Catalog record 2 pass
- Pass Task access All tasks are public.
- Pass Confidence Checked against a primary source.
Scores 1 warning 2 pass 2 notes
- Pass Independent runs Scores from the benchmark’s paper (25 entries).
- Pass Models scored 3 models we track, plus 22 other entries on its leaderboard.
- Warning Newest score The newest came 21 months ago, in February 2025, for Gemini 2.0 Flash.
- Note Frontier models None of the top 10 has a score. The strongest tested is o1-mini (#142).
- Note Headroom Scored as a plain number, with no fixed ceiling.
Leaderboard import 3 pass
- Pass Importer We read the board on arxiv.org: 3 rows for models we track and 22 other entries at the last read.
- Pass Last read November 20th, 2026.
- Pass Read worked The last read worked.
Consistency 1 pass 1 note
- Note Lab claims No model has both a lab number and an independent run.
- Pass Known issues No published audit found problems.