SWE-Duel: Towards a Self-Scaling Arena for Coding Agents
SWE-Duel proposes a coding benchmark whose tasks are generated during evaluation by the agents being tested.
The arena has agents create features, hide bugs, and then challenge another agent to review and repair the pull request without being told a bug exists. A submitted task is accepted only if it passes feature and existing tests, fails only the sealed bug tests, and can be repaired blind by a fresh session of the same agent. In a 1,500-turn tournament across 11 participants and 15 repositories, GPT-5.5 with Codex CLI led as a solver, repairing 98.3% of reviewed defects. Claude Opus 4.8 with mini-swe-agent ranked second as a task author, with a 27.3% Red win rate. HF Daily Papers' note
The arena has agents create features, hide bugs, and then challenge another agent to review and repair the pull request without being told a bug exists. A submitted task is accepted only if it passes feature and existing tests, fails only the sealed bug tests, and can be repaired blind by a fresh session of the same agent. In a 1,500-turn tournament across 11 participants and 15 repositories, GPT-5.5 with Codex CLI led as a solver, repairing 98.3% of reviewed defects. Claude Opus 4.8 with mini-swe-agent ranked second as a task author, with a 27.3% Red win rate. HF Daily Papers' note
score 6