LintEval - Reduction@P: leaderboard
LintEval (arXiv 2610.20027): the share (%) of the original linter alerts a model's refactor removes, net of the new alerts it introduces, over the files that still compile and pass their tests (corpus-level; 498 LintEval-Lite programs in C++, Python and Java, one turn, greedy decoding). Each model is scored on its own passing files, and the value turns negative when a model adds more alerts than it fixes. Higher is better. Source: https://arxiv.org/abs/2610.20027
Metric: Reduction@P (%; corpus-level net reduction of the original linter alerts, fixed minus newly introduced over original, on the programs that still pass; negative when a model adds more alerts than it removes). Source: arxiv.org. Saturation forecast: Around July 2027. 14 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-5.4 Mini | 90.4 | #208 |
| 2 | GPT-5.2 | 87.8 | #109 |
| 3 | DeepSeek V3.2 | 73 | #201 |
| 4 | Qwen 3 8B | 70.7 | #670 |
| 5 | Qwen 3.6 27B | 62 | #168 |
| 6 | Qwen 3.6 35B A3B | 61.5 | #224 |
| 7 | Qwen 2.5 Coder 32B Instruct | 57.1 | #524 |
| 8 | Qwen 3 4B | 32 | #829 |
| 9 | Qwen 3.5 9B | 31.3 | #377 |
| 10 | Qwen 2.5 Coder 7B Instruct | 24.7 | #939 |
| 11 | Llama 3.2 3B Instruct | 19.8 | #1347 |
| 12 | Qwen 2.5 Coder 3B Instruct | 9.5 | #1021 |
| 13 | Llama 3.2 1B Instruct | 0.5 | #1534 |
No result here: #3 Claude Opus 5.5, #6 GPT-6 Astra, #7 GPT-6.1 Sol.
Interactive version: theaggregate.ai/benchmark?slug=linteval-reduction-p · How It Works · Data refreshed daily, snapshot 2026-10-19.