SWE-Explore - File Hit Rate: leaderboard
SWE-Explore (2026) isolates repository exploration for coding agents: for each of 848 issues from 203 repositories in 10 languages the explorer returns its five most relevant code regions, scored against line-level context distilled from successful repair trajectories, here with each LLM driving the Mini-SWE-Agent scaffold. This board is the percent of ground-truth files that the returned regions touch; higher is better.
Metric: File-level hit rate (%; share of ground-truth files that the returned regions touch) on 848 repository-exploration instances (issues from 203 repositories in 10 languages): given the repository and issue, the explorer returns five ranked code regions, scored against line-level core context distilled from independent successful repair trajectories; the Mini-SWE-Agent scaffold with each LLM; higher is better. Source: arxiv.org. Saturation forecast: Around November 2027. 6 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-5.4 | 65.5 | #123 |
| 2 | GPT-5.4 Mini | 64.9 | #389 |
| 3 | Claude Sonnet 4.5 | 53.5 | #250 |
| 4 | Kimi K2.6 | 50.9 | #178 |
| 5 | Gemini 3 Pro | 36.9 | #137 |
| 6 | GLM-4.7 | 34.3 | #366 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=swe-explore-file-hit-rate · How It Works · Data refreshed daily, snapshot 2026-11-13.