On-policy distillation (OPD) has emerged as a promising approach for continual learning, where a student policy is supervised at the token level. A new paper on arXiv introduces a framework that addresses a key limitation of current OPD methods: the use of a single teacher per prompt.

The authors argue that compatibility between teacher and student can vary across tokens, and propose selecting different teachers for different tokens within a single prompt. This token-level selection is designed to improve the quality of supervision during distillation.

The abstract is truncated, so full implementation details and experimental results are not yet available. The paper is listed as a new announcement on arXiv under reference 2610.21234.