HarnessAdapt-RL lifts 9B agent 12.6 pts on SWE-bench Multilingual

Co-training the agent's policy and its tool-and-hook harness during the same reinforcement-learning run lifts pass@1 over a GRPO baseline by 9.8, 12.6 and 6.4 percentage points on SWE-bench Verified, Multilingual and Pro, using only a 9B base model. The method, called HarnessAdapt-RL, is described in a paper posted to Hugging Face on 21 October by Wei Tao, Yiwen Guo and Lingfeng Bao, who credit the work to an internship at Lightspeed.

The standard recipe freezes the harness. The authors point out that RL trajectories 'reveal flawed repair strategies and interaction problems that vary across tasks and change as the policy learns.' A frozen teacher generates task-specific 'mechanism-level skills' that give the when and why of a repair but not the concrete answer, while a separate teacher evolves tools and hooks, 'keeping each change only if it improves the observed pass rate or the observed trainable rate without lowering either.'

The two pieces behave differently at deployment. Mechanism-level skills can be stripped off at test time and the gains hold. Evolved tools and hooks cannot: removing them costs up to 7.7 percentage points on Multilingual. The authors conclude that training 'produces a model–harness pair: the trained policy and the evolved tools and hooks, with skills removed at deployment.'

On all three SWE-bench variants the same 9B pair also outperformed a GRPO baseline that used OpenHands' human-designed harness.