Harness-Aware Distillation for Small Language Model Agents
Harness-Aware Distillation trains smaller agents on skills the fixed software harness cannot supply.
Harness-Aware Distillation (HAD) trains smaller language-model agents on teacher abilities that a fixed software harness cannot supply. It adds an action preference, scored after the student's own reasoning, and a validity check that discards pairs whose preferred action contradicts harness records. The method needs no task rewards, success labels, or future information. Across long-horizon agent benchmarks, HAD beats on-policy distillation with the same harness, with fewer unproductive loops and more frequent error recovery.
- HAD distills only what the teacher adds beyond the agent harness.
- Action preference contrasts the same actions with and without harness information.
- A validity check drops pairs that contradict harness records.
- No task rewards, success labels, or future information are required.
- Students enter fewer unproductive loops and recover from errors more often.
From the original article
1 min read at huggingface.co
Language model agents are deployed with a harness, the software around the model that manages its context, tools, and feedback. When such an agent is distilled into a smaller one, the harness stays in place, so the student mainly needs the teacher-specific abilities that the harness cannot provide, such as acting correctly on harness information.
By Moonseok Choi, Taehong Moon, Giung Nam · 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.