← News

ResearchSources: arXiv · 19 Nov 2026

COPD distills black-box teacher skills into smaller agents via natural-language requests

Researchers propose Communicative On-Policy Distillation, letting a student model decide when to ask a black-box teacher for help in plain language, then learn from those exchanges without later access to the teacher.

In plain words

A research team describes a training method where a smaller AI agent improves by asking a larger, closed-off model questions in ordinary language during tasks. The agent keeps successful exchanges as teaching material and later performs well even when the larger model is no longer available. The method is described in a paper abstract, so full details sit in the complete paper.

What was announced

On-policy distillation typically improves language-model agents by learning from a teacher model's internal outputs, but text-only black-box models do not expose those signals. COPD instead uses natural-language communication: during task execution, the student autonomously decides when to consult the teacher and writes open-ended requests, with no predefined question templates or intervention schedules. Successful interactions are retained as privileged references to supervise fresh adviser-free student rollouts through on-policy self-distillation, trained jointly with task-reward reinforcement learning, according to the abstract.

The authors report that on the CAMG-II benchmark, COPD achieved 2.46 times the gain of the strongest black-box OPD baseline they compared against. Capabilities learned through communication also generalized to Terminal-Bench 2.1 and SWE-Bench Pro V2, improving task success without adviser access. These figures come from the paper's own abstract and have not been independently verified; the excerpt does not describe the benchmarks, baselines, or model sizes in detail.

What it does not show

  • Evidence is limited to the paper's abstract; benchmark details, baseline definitions, and model configurations are not described in the supplied passages.
  • Reported gains are the authors' own claims and the passages do not indicate independent confirmation.

For security and IT teams

Scope
The supplied material covers only the paper's abstract, so deployment details and method internals are outside the available evidence.

Sources

Source-support review completed 20 Nov 2026. This story was drafted with AI assistance and checked against retained source passages. Automated checking is not independent verification.

Coverage: arXiv: abstract.

Updates and corrections

No corrections so far.

First published 20 Nov 2026

New to the subject? AI from the Ground Up starts from the beginning.