Source-Test Repair - Hunk-level Executable: leaderboard
Source-test co-evolution repair on 500 Java cases from 15 repositories; Java17/Maven3.9.9, temperature0.6. Native fractions retain denominator500. Executable means repaired tests pass, not semantic equivalence to developer intent; compilable measures build validity. Context and pipeline configurations are separate. Hunk-level; Executable.
Metric: Executable fraction. Source: arxiv.org. Saturation forecast: Rough model projection: around 2026. 3 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-5.4 (low) (TestRepair Run: context=Hunk-level; pipeline=baseline) | 0.52 | |
| 2 | Qwen3-Coder-480B-Instruct (TestRepair Run: context=Hunk-level; pipeline=baseline) | 0.44 | |
| 3 | DeepSeek-V3.2 (Non-reasoning) (TestRepair Run: context=Hunk-level; pipeline=baseline) | 0.32 |
Interactive version: theaggregate.ai/benchmark?slug=source-test-repair-hunk-level-executable · How It Works · Data refreshed daily, snapshot 2026-11-05.