HarnessAdapt-RL trains the coding agent's harness too, beating GRPO by 9.8 points on SWE-bench Verified
tencent · hf · 2026-10-22
The paper HarnessAdapt-RL jointly trains a coding policy and adapts its evaluation harness during RL, instead of keeping the harness fixed.
- A frozen teacher generates task-specific mechanism-level skills to fix flawed repair strategies; a separate teacher evolves tools and hooks, keeping only changes that improve pass rate or trainability
- With a 9B base model, pass@1 beats GRPO by 9.8, 12.6, and 6.4 points on SWE-bench Verified, Multilingual, and Pro, and outperforms GRPO with the human-designed OpenHands harness
- Gains from mechanism-level skills persist after removal at evaluation; removing evolved tools/hooks at deployment costs up to 7.7 points — training yields a model-harness pair
More from coding & agent
- Developer connects Claude, Codex and Grok agents across machines to talk — baditaflorin · 2026-10-22
- Copilot CLI v1.0.102 ships subagent permission inheritance and sandbox proxy fixes — copilot-cli-release-app[bot] · 2026-10-22
- Ethan Mollick: AI coding tools help, but product bottlenecks lie elsewhere — emollick · 2026-10-22
- LiteLLM 1.107.0 on PyPI and Docker shipped an infostealer and backdoor — smakosh · 2026-10-22
- Tsinghua paper: letting agents manage their own KV cache cuts cost by 27.7% — dair_ai · 2026-10-22
- Every shows how a shared Slack agent lets teammates borrow each other's AI skills — every · 2026-10-22