HarnessAdapt-RL trains coding-agent policy and harness together
Original title: Training the Whole Agent, Not Just the Model
On a 9B model, pass@1 rises 9.8, 12.6, and 6.4 points over GRPO on three SWE-bench sets.
HarnessAdapt-RL updates a coding-agent policy during reinforcement learning while also adapting its harness. A frozen teacher writes task-specific skills for flawed repair strategies; a second teacher evolves tools and hooks only when pass rate or trainable rate improves without lowering the other. With a 9B base model it reports higher pass@1 than GRPO on SWE-bench Verified, Multilingual, and Pro, and beats GRPO using the human-designed OpenHands harness. Deployment keeps the evolved tools and hooks but removes the skills.
- Pass@1 gains are 9.8, 12.6, and 6.4 percentage points on three SWE-bench sets.
- Skill gains persist after those skills are removed at evaluation.
- Removing evolved tools and hooks at deployment costs up to 7.7 points.
- Re-evolving tools as the policy changes beats freezing early versions.
- Products
- HarnessAdapt-RLOpenHands
From the original article
1 min read at huggingface.co
Training coding agents typically updates the policy while keeping the harness fixed. During training, trajectories reveal flawed repair strategies and interaction problems that vary across tasks and change as the policy learns. We introduce HarnessAdapt-RL, which trains the policy and adapts its harness during reinforcement learning.
By Lingxiao Tang, Wei Tao, Yiwen Guo · huggingface.co
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.