# New method lets AI agents learn from black‑box teachers via natural language

> Researchers propose Communicative On‑Policy Distillation, letting smaller models ask open‑ended questions to larger black‑box models and improve long‑task performance.

Oossa · 2026-11-19 · https://oossa.com/en/new-method-lets-ai-agents-learn-from-black-box-teachers-via-natural-language

Lirui Luo and co‑authors released a paper on Nov 19, 2026 describing Communicative On‑Policy Distillation (COPD). The technique lets a student model decide when to query a black‑box teacher using free‑form natural language, then uses those exchanges to train itself without further teacher help. On the CAMG‑II benchmark COPD achieved 2.46 × the improvement of the best existing black‑box on‑policy distillation method. The authors also report gains on Terminal‑Bench 2.1 and SWE‑Bench Pro V2.

## The facts

- Paper published Nov 19 2026
- COPD improves performance by 2.46 × over the strongest black‑box baseline on CAMG‑II

## Why it matters

If the method works in practice, developers could boost AI agents for complex tasks without needing full access to large proprietary models.

## Sources & references

1. [Communicative On-Policy Distillation: Distilling Black-Box Teacher Capabilities through Communication for Long-Horizon Tasks](https://arxiv.org/abs/2611.23673) – arXiv cs.LG, 2026-11-19

Last updated: 2026-11-19
