Flux-OPD: On-Policy Distillation with Evolving Contexts

Researchers proposed Flux-OPD, a new on-policy distillation method for training open-ended language models.

It turns evolving contexts into supervision and uses conflict-aware weighting to stabilize learning.

The paper analyzes reverse KL distillation with context-conditioned teachers, identifying a geometric-mean effect and a conflict term.

This helps train models when rewards are hard to verify by reducing clashes between teacher signals.