arXiv 2609.35259 · cs.LG · 2026-09-28
On-Policy or Off-Policy Learning? A Systematic Study of Distillation Dynamics
Julianna Piskorz, Antonin Berthon, Mihaela van der Schaar
University of Cambridge
In latest · #77 Added on 2026-10-21
One-line takeaway
A controlled study isolating on- vs off-policy rollouts in distillation, finding a more nuanced picture than 'on-policy is always better'.
在受控实验中分离 on-policy 与 off-policy 蒸馏的作用，发现结论比“on-policy 总更好”复杂。
Read on arXiv PDF Hugging Face discussion Project page
Topics & disciplines
More on Efficient training & inference
- The Teacher Is a Direction, Not a Destination: Extrapolating RL-Induced Representation Residuals in On-Policy Distillation ▲511
- Scaling Properties of Same-Family On-Policy Distillation ▲327
- Beyond Domain-Routed Supervision: Cross-Teacher Guarding for Multi-Teacher On-Policy Distillation ▲251
- Beyond Teacher Assignment: Domain-Normalized Multi-Teacher On-Policy Distillation ▲181
- TokenRouter: Efficient Serving System for Token-Level LLM Routing ▲143
About this data
Metadata from the arXiv API (verified); upvotes from Hugging Face Daily Papers; citations from Semantic Scholar. Last refreshed 2026-10-22.
The takeaway is written by AIMap from the abstract; the abstract itself is not reproduced — read the original.