Diao et al.
Offline trajectories can train agents without environment interaction while limiting forgetting
SAPD decomposes recorded multi-turn histories into decision prefixes and uses relative step advantages to stabilize policy distillation, improving results on ALFWorld, WebShop and Search-based QA.
Research Lab
Juncheng Diao · Haibin Wang · Wei Shi · Yifan Ding · Qingbin Li · Rongxiang Weng · +1 more
Meituan · University of Chinese Academy of Sciences
Research Digest··3 min read
The authors propose Step-Advantage Policy Distillation (SAPD), a method that trains LLM agents from recorded interaction histories without revisiting executable environments.
§