Reproduce “Agentic On-Policy Distillation Implicitly Reshapes LLMs' ReAct Patterns” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Deyang Kong, Zhaoyi Li, Sijian Ren, Ying Wei, Defu Lian, Shikun Zhang, Wei Ye
The authors study how on-policy distillation (OPD) transfers ReAct behaviors across math, household tasks, and simulated customer support using Qwen3-4B students. They find that ALFWorld-only training lowers τ²-bench accuracy from 41.5% to 6.6%, with dependency violations in 55.8% of reviewed trajectories. Selecting the top 95% of token positions raises τ² accuracy from 11.5% to 29.9%, but remains below initialization.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.