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 observatio
Agentic On-Policy Distillation Implicitly Reshapes LLMs' ReAct Patterns
On-policy distillation (OPD) implicitly reshapes the ReAct interaction patterns of language-model agents across mathematics, ALFWorld and τ^2-bench, according to a study examining OPD's cross-domain effects on reasoning, actions and observations. The research reports that OPD transfers specialized capabilities to LLM agents but that its effects on interaction patterns beyond the target domain remain unclear.
After on-policy distillation, re-run ReAct interaction evals on out-of-domain benchmarks such as ALFWorld and τ^2-bench to catch shifted reasoning, action and observation patterns before deployment.
more in playbook: Evals and testing for LLM appsRun your AI side-project on zahid.host
EU-based hosting, git-push deploys, automatic HTTPS, no cold starts. Free tier with a custom domain — perfect for shipping the agent you just read about.