Microsoft Research’s ReOPD reuses teacher prefixes to distill multi-turn agents offline
MicrosoftResearch · hf · 2026-07-24
ReOPD turns multi-turn distillation into an offline prefix replay problem
Microsoft Research studies on-policy distillation (OPD) for multi-turn agent tasks and proposes Replayed-Prefix On-Policy Distillation (ReOPD).
- Standard online OPD is expensive because each update needs fresh student rollouts and teacher queries.
- ReOPD reuses pre-collected teacher trajectories as replayed prefixes, so the student acts only at selected steps while the teacher provides dense supervision.
- The paper identifies a prefix trap: later prefixes are more on-policy for the student, but can fall into histories where the teacher’s targets become unreliable.
- To handle that tradeoff, ReOPD uses a step-decaying sampling schedule that emphasizes earlier, lower-shift prefixes.
- Across mathematical reasoning with Python and search environments, ReOPD preserves or improves OPD-level accuracy.
- It uses zero tool calls during student training and is reported to be at least 4× faster per rollout than online OPD.
The core idea is to convert expensive agent-environment interaction into a reusable offline resource for scalable distillation.
More from coding & agent
- Kimi K2.8 Preview rolls out: near-K3 coding performance, 1M context for all tiers — teortaxesTex · 2026-09-11
- Looking for a classifier of software engineering task shapes to pick models per task — StewartalsopIII · 2026-09-11
- Steal this idea: prompt-to-hardware where agents assemble custom devices — paraschopra · 2026-09-11
- Model Is the Least Interesting Part: A Guide to Six Core AI Architectures from RAG to Multi-Agent — goyalshaliniuk · 2026-09-11
- Non-coder builds layered memory architecture: 20k tokens tracks a year of agent conversations — matteoianni · 2026-09-11
- Warp's six non-engineering teams all run on Linear and Claude Code — mon__lim · 2026-09-11