an executable, live benchmark for test and code co-evolution
mined from real open-source Java projects
Tests have to change when the code they check changes. TestEvo-Bench turns real commits from open-source Java (Maven) projects into tasks for coding agents: given a code change, the agent must write or edit the tests that go with it. Every task comes with a buildable repository and is validated by running the developer's own test, and every task records its commit date, so results can be restricted to commits made after a model's training cutoff.
The current snapshot has 3,020 tasks from 196 repositories with commits from January 2019 to June 2026, and a 584-task Lite subset for cheaper evaluation. Beyond pass/fail, each test is scored by the coverage it reaches and the share of mutants it kills. On Lite, the six agent configurations we evaluated (mini-SWE-agent and Claude Code with four models) succeed on 98.3–99.1% of tasks and reach an overall score of 79.8–84.4%, on par with the developers' own tests (81.8%). The leaderboard therefore ranks them by this overall score, which also counts coverage and mutation.
Agent results on the 584-task Lite subset, recomputed for the selected commit-date window. Over the full window the numbers match the paper.
Loading…
| # | Agent | Model | Success | CovOnPass | MutOnPass | Overall | Tasks |
|---|---|---|---|---|---|---|---|
| Loading leaderboard… | |||||||
Browse every repository in the full benchmark. Filter by commit date, track or the Lite subset, then click a row to see its tasks and the GitHub diffs.
Loading…
| Repository | Total | update | generation | refinement | Date range | |
|---|---|---|---|---|---|---|
| Loading tasks… | ||||||