Reproduce “More Methods, Better Agents? A Reality Check on Self-Distillation in Agentic Reinforcement Learning” 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
Linhao Zhong, Bozhen Fang, Xuehao Wang, Duochao Shi, Zekai Luo, Guangming Yao, Yi Yuan, Hao Chen and 2 more
The authors benchmark ten self-distillation methods against GRPO across four tasks and Qwen3.5 models at 2B, 4B, and 9B, using a shared evaluation protocol. Across twelve task-model settings, PCSD has the highest equal-weight mean, 54.26 versus GRPO’s 49.68, but leads no individual setting, winning six, trailing five, and tying one; results are limited to this benchmark.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.