FormalRewardBench: leaderboard
Recent neural theorem provers use reinforcement learning with verifiable rewards (RLVR), where proof assistants provide binary correctness signals. While verifiable rewards are cheap and scalable without reward hacking issues, they suffer from sparse credit assignment: models receive no learning signal from difficult problems where partial progress goes unrewarded. This motivates learned reward mo
Metric: Pairwise Accuracy (self-reported). Source: benchmarklist.com. Saturation forecast: Around February 2027. 26 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Claude Opus 4.5 | 59.8 | #83 |
| 2 | Claude Sonnet 4.5 | 45.7 | #142 |
| 3 | GPT-5.2 | 39.7 | #110 |
| 4 | Qwen 2.5 Coder 32B | 37.6 | #667 |
| 5 | Claude Sonnet 4 | 32.4 | #199 |
| 6 | GPT-4.1 | 32.1 | #248 |
| 7 | GPT-5.1 | 30.2 | #135 |
| 8 | Gemini 2.5 Flash | 25.2 | #245 |
| 9 | GPT-4o | 23.1 | #343 |
| 10 | Qwen2.5-Math-7B | 0 | #1482 |
No result here: #3 Claude Opus 5.5, #6 GPT-6 Astra, #7 GPT-6.1 Sol.
Interactive version: theaggregate.ai/benchmark?slug=formalrewardbench · How It Works · Data refreshed daily, snapshot 2026-10-20.