RepoReason: leaderboard
Repository-level code reasoning benchmark (RepoReason, Li, Su and Lyu; CUHK and Sun Yat-sen University, 2026): 2,492 unit-test assertions from seven Python libraries are mutated by an LLM, re-executed to capture fresh ground truth and masked, and an agent in OpenHands' read-only mode browses the repository to state the value that makes each assertion pass. Answers are checked by rule-based equivalence (SymPy, structural normalization); the score is Pass@1 accuracy in percent at temperature 0, higher is better.
Metric: Pass@1 accuracy (%; 2,492 masked assertions over seven Python repositories, OpenHands read-only agent). Source: arxiv.org. Saturation forecast: Around February 2027. 5 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Claude Sonnet 4.5 | 66.98 | #244 |
| 2 | DeepSeek V3.1 Terminus | 60.96 | #442 |
| 3 | GPT-5.2 | 56.86 | #172 |
| 4 | Kimi K2 | 54.74 | #435 |
| 5 | Qwen 3 Coder 480B A35B | 50.56 | #677 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=reporeason · How It Works · Data refreshed daily, snapshot 2026-11-12.