LLVM-Bench (SWE-agent): leaderboard
LLVM-Bench (2026) poses 423 real, reproducible LLVM compiler issues; a patch counts as resolved only if it applies, builds and passes the full LLVM test suite with the issue tests in the LLVM-Gym environment. Resolved rate in percent under the SWE-agent scaffold with at most 50 turns, higher is better.
Metric: Resolved rate (%; share of the 423 validated LLVM issues, versions 18-21, whose generated patch applies, builds and passes the full LLVM test suite including the issue tests in LLVM-Gym; SWE-agent scaffold with at most 50 interaction turns, temperature 0, 8,192-token generation limit). Source: arxiv.org. Saturation forecast: Around 2030. 4 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Grok Code Fast 1 | 10.87 | #706 |
| 2 | DeepSeek V3.2 | 7.8 | #406 |
| 3 | Gemini 3 Flash | 4.26 | #164 |
| 4 | Qwen 3 Coder Plus | 3.78 | #717 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=llvm-bench-swe-agent · How It Works · Data refreshed daily, snapshot 2026-11-13.