MCR-Bench - Defect Identification F1: leaderboard
MCR-Bench (arXiv 2608.27442v1, Table 5): multi-round code review, where a model follows pull-request discussions and code revisions across review rounds and must identify the defects raised. Defect identification F1 against annotated defects across Python, Java, JavaScript, TypeScript and C#, percent; plain prompting without a review scaffold. Higher is better.
Metric: F1 (%; matched defects, all languages). Source: arxiv.org. Saturation forecast: Around February 2027. 7 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Claude Haiku 4.5 | 55.1 | #287 |
| 2 | GPT-5.2 | 54.2 | #115 |
| 3 | Gemini 3 Flash | 53.2 | #103 |
| 4 | DeepSeek V3.2 | 49 | #210 |
| 5 | GLM-4.7 | 47 | #202 |
| 6 | Kimi K2 | 37.3 | #252 |
| 7 | Qwen 3 Max | 35.7 | #213 |
No result here: #3 Claude Opus 5.5, #6 GPT-6 Astra, #7 GPT-6.1 Sol.
Interactive version: theaggregate.ai/benchmark?slug=mcr-bench-defect-identification-f1 · How It Works · Data refreshed daily, snapshot 2026-11-05.