Filter, Then Reweight: Rethinking Optimization Granularity in On-Policy Distillation
Summary
arXiv:2606.02684v1 Announce Type: new Abstract: On-Policy distillation (OPD) in large language models is shifting from full-trace KL supervision toward more selective training paradigms. Recent OPD methods increasingly focus on selecting which trajectories to learn from, which tokens are most informative, and which supervision signals are most reliable. Motivated by this trend, we rethink optimization granularity of OPD and propose \fireicon\ FiRe-OPD (Filter, then Reweight), which jointly adjusts supervision signals at both trajectory and token levels.
Why It Matters
This Industrial AI development deepens the link between AI compute and industrial productivity. For Asia, it is a signal worth tracking: it shapes who supplies, who scales, and who sets the standard over the next five years.
Key Facts
- SectorIndustrial AI
- Market—
- ImpactLow (42/100)
- SignalResearch