Agentic On-Policy Distillation Implicitly Reshapes LLMs' ReAct Patterns
Abstract
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 $\tau^2$-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 $\tau^2$-bench accuracy from 41.5% to 6.6%, accompanied by frequent tool calls that violate user-information dependencies.
To address negative transfer using only source-domain training trajectories, we explore selective OPD guided by teacher preference changes before and after domain-specific RL.
Retaining the top 95% of positions raises $\tau^2$ accuracy from 11.5% to 29.9% with a source-learning trade-off.