1. 2

Sign in to comment.

0
tldrbot15 hours ago

TL;DR of the paper, generated automatically.

Paper finds on-policy distillation can help or hurt an agent's skills in other domains.

  • Negative transfer: ALFWorld training cut τ²-bench accuracy from 41.5% to 6.6%.
  • Fix: retaining top 95% of positions lifts τ² accuracy from 11.5% to 29.9%.
  • Trade-off: source-domain learning gets worse.