Megadose The AI curve, compressed.

Agentic On-Policy Distillation Implicitly Reshapes LLMs' ReAct Patterns

· HF Daily Papers ·
Training on household tasks cut accuracy on a separate customer-service benchmark from 41.5% to 6.6%.

The study tested how on-policy distillation changes agents’ reasoning, actions and information gathering across three domains. ALFWorld supervision produced frequent tool calls that violated dependencies on user information in τ²-bench. General-reasoning supervision, meanwhile, reduced mathematical tool attempts from 70.4% to 14.4% while raising accuracy from 16.7% to 38.0%. A selective distillation method retaining 95% of training positions raised τ²-bench accuracy from a separate 11.5% baseline to 29.9%, with a trade-off in source-domain learning.

HF Daily Papers' note

score 5

Categories: Research