Harness-Zero: Harness Distillation via Agent-as-Harness
Harness-Zero trains the model to keep a specialized agent harness’s gains after that harness is removed.
The paper frames agent harnesses as useful but hard to deploy broadly because different tasks, models, and instances may need different external systems. Harness-Zero uses an “agent-as-harness” to rewrite or correct student model responses into the target harness’s action space, creating trajectories for fine-tuning. In the reported experiments, the distilled model’s macro-average task success rises from 23.3% to 44.3% under a fixed target harness. The authors say it also recovers 82.3% of measured harness-induced behavior patterns across knowledge work, tool use, and science domains. HF Daily Papers' note
The paper frames agent harnesses as useful but hard to deploy broadly because different tasks, models, and instances may need different external systems. Harness-Zero uses an “agent-as-harness” to rewrite or correct student model responses into the target harness’s action space, creating trajectories for fine-tuning. In the reported experiments, the distilled model’s macro-average task success rises from 23.3% to 44.3% under a fixed target harness. The authors say it also recovers 82.3% of measured harness-induced behavior patterns across knowledge work, tool use, and science domains. HF Daily Papers' note
score 5