FormalRewardBench - Pointwise: leaderboard

Pointwise accuracy (%) on FormalRewardBench (2026), a benchmark of reward models for Lean 4 theorem proving: the model scores a verified proof and a flawed proof of the same theorem independently, and an item counts when the correct one scores higher. Flawed proofs come from five error strategies such as minimal edits, forced mistakes and injected Python; higher is better.

Metric: Pointwise accuracy (%): the model scores the correct and the incorrect Lean proof of each item independently, and the item counts when the correct proof scores higher; temperature 1.0, top-p 0.75; higher is better. Source: arxiv.org. Saturation forecast: Around February 2027. 26 models tracked. Updated 7 Oct 2026.

Top models

#ModelScoreOverall rank
1Claude Opus 4.570.1#160
2Claude Sonnet 4.562#261
3Con-J-Qwen2-7B52.8#329
4Gemini 2.5 Flash50.9#463
5Claude Sonnet 449.8#410
6GPT-5.248.9#185
7Selene-146.8
8GPT-4o44#751
9GPT-4.144#505
10CompassJudger-1-7B-Instruct43.6
11GPT-5.141.8#240
12LMUnit-qwen2.5-72B41.2#524
13CompassJudger-1-14B-Instruct40.3
14Qwen 2.5 72B Instruct39.8#1117
15Skywork-Critic-Llama-3.1-70B39.2

No result here: #1 Claude Fable 5.2, #3 Gemini 4 Neon, #4 GPT-6.2 Astra.

Interactive version: theaggregate.ai/benchmark?slug=formalrewardbench-pointwise · How It Works · Data refreshed daily, snapshot 2026-11-21.