Communicative On-Policy Distillation: Distilling Black-Box Teacher Capabilities through Communication for Long-Horizon Tasks
On-policy distillation (OPD) is a powerful paradigm for improving language-model agents on long-horizon tasks, yet its conventional formulation relies on teacher logits that text-only black- box models do not expose. To address this limitation, we introduce Communicative On-Policy Distillation (COPD), which uses natural-language communication as a medium for distilling a stronger model's capabilities into a smaller student. During task execution, the student au- tonomously decides when to consult a black-box teacher and formulates open-ended natural- language requests without predefined question templates or intervention schedules. Successful student-teacher interactions are retained as privileged references to supervise fresh adviser-free student rollouts through on-policy self-distillation, jointly with task-reward reinforcement learning. On CAMG-II, COPD achieves 2.46 times the gain of the strongest black-box OPD baseline. The capabilities it learns through communication generalize to Terminal-Bench 2.1 and SWE-Bench Pro V2, improving task success without adviser access.
Authors
- Cong Fang (ORCID: https://orcid.org/0000-0002-5076-7897)
- Luyu Chen (ORCID: https://orcid.org/0000-0002-2153-5712)
- Kelong Mao (ORCID: https://orcid.org/0000-0002-5648-568X)
- Gu Simiu
- Rongqing Li
- Kieran Wong
- Jiayin Zhu
- Lirui Luo
- Yudong Guo
- Sulong Xu
- Xinrui Wang
- Xinwei Yang
Publication Details
- Journal
- arXiv (Cornell University)
- Published
- 2026-11-18
- DOI
- https://doi.org/10.48550/arxiv.2611.23673
- Primary Topic
- Reinforcement Learning in Robotics
- Type
- preprint