Harness-Aware Distillation for Small Language Model Agents
The paper argues small agent models should learn what the harness cannot supply, not copy the teacher wholesale.
Harness-Aware Distillation trains against the gap between a teacher’s actions with and without harness information, then filters out preference pairs that conflict with the harness records. The method is presented as reward-free and label-free, using no future information. Across long-horizon agent benchmarks, the authors report gains over on-policy distillation under the same fixed harness. They also say HAD reduces unproductive loops and improves recovery from errors. HF Daily Papers' note
Harness-Aware Distillation trains against the gap between a teacher’s actions with and without harness information, then filters out preference pairs that conflict with the harness records. The method is presented as reward-free and label-free, using no future information. Across long-horizon agent benchmarks, the authors report gains over on-policy distillation under the same fixed harness. They also say HAD reduces unproductive loops and improves recovery from errors. HF Daily Papers' note
score 4