ComBench (Compilation Error Repair): leaderboard
Metric: Semantic correctness (%) of patches for the 200 reproducible real CI compilation errors from Bitcoin, OpenSSL, RocksDB and LLVM: the patch compiles without new errors and, by the authors' manual analysis, matches the developer's intended semantics, single-shot repair with perfect fault localization: the prompt gives the full source of the failing file and of every file the developer's fix modified plus the raw compiler diagnostics, and the model answers once with a cross-file search-and-replace patch (one retry when no patch applies), temperature 0.2; higher is better. Source: arxiv.org. Saturation forecast: Around July 2027. 9 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-5 | 41 | #91 |
| 2 | Gemini 2.5 Pro | 36.5 | #145 |
| 3 | Claude Sonnet 4 | 34 | #194 |
| 4 | GPT-4.1 | 32 | #240 |
| 5 | Claude 3.7 Sonnet | 28.5 | #241 |
| 6 | DeepSeek V3 | 27.5 | #312 |
| 7 | Qwen 3 235B A22B | 24.5 | #304 |
| 8 | Qwen 3 32B | 21 | #424 |
| 9 | Qwen 3 30B A3B | 20 | #488 |
No result here: #3 Claude Opus 5.5, #5 GPT-6 Astra, #8 Claude Fable 5.1.
Interactive version: theaggregate.ai/benchmark?slug=combench-compilation-error-repair · How It Works · Data refreshed daily, snapshot 2026-10-11.