About this paper
On-policy distillation can improve or degrade language-model agent interaction across domains, and selectively retaining teacher-preferred positions mitigates negative transfer at a cost to source-domain learning.
On-policy distillation (OPD) transfers specialized capabilities to language-model agents, but its cross-domain effects on interaction remain unclear. We investigate how OPD reshapes ReAct interaction patterns across mathematics, ALFWorld, and τ^2-bench, focusing on reasoning, actions, and observation acquisition. General-reasoning supervision reduces mathematical tool attempts from 70.4% to 14.4% while increasing accuracy from 16.7% to 38.0%; code-interpreter supervision improves household actions and retail queries. In contrast, ALFWorld supervision reduces τ^2-bench accuracy from 41.5% to 6.6%, accompanied by frequent tool calls that violate user-information dependencies. To address…
on-policy distillation (OPD)language-model agentsReAct interaction patternscross-domain effectsgeneral-reasoning supervisioncode-interpreter supervisionmathematical tool attemptsobservation acquisitionnegative transfersource-domain training trajectories
- Upvotes
- 7
- Authors
- 8
- Published
- Nov 14, 2026
Summary and abstract excerpt via Hugging Face Papers ↗ · arXiv.
Trending history
Agentic On-Policy Distillation Implicitly Reshapes LLMs' ReAct Patterns was first seen on Hugging Face · Daily papers on Nov 18, 2026. Since then it has appeared in 11 checks on 2 different days. Its best position was #18 of 30, reached on Nov 18, 21:28 UTC. It dropped off the list after Nov 19, 03:28 UTC.