Reproduce “BREAD: Branched Rollouts from Expert Anchors Bridge SFT & RL for Reasoning” 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

Xuechen Zhang, Zijian Huang, Yingcong Li, Chenshun Ni, Jiasi Chen, Samet Oymak

Researchers at the University of Michigan developed BREAD, a method that enhances small language models' complex reasoning by unifying supervised fine-tuning and reinforcement learning. BREAD provides adaptive partial expert hints during training, leading to higher accuracy and a 75% reduction in computational costs compared to traditional methods.

Key insights

  • Small language models (SLMs) can effectively utilize partial expert reasoning traces as hints, even when direct supervised fine-tuning (SFT) on full traces is ineffective or detrimental.
  • Adaptive branching and guided rollouts within the reinforcement learning (RL) process create a natural learning curriculum and densify reward signals, overcoming sparse reward challenges.
  • Training with partial expert guidance can enable SLMs to generalize reasoning patterns and improve performance even on problems without hints.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.