Reproduce “Making Small Language Models Efficient Reasoners: Intervention, Supervision, Reinforcement” 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, Chenshun Ni, Ziyang Xiong, Jiasi Chen, Samet Oymak

Researchers from the University of Michigan developed methods to enhance token-efficient reasoning in small language models (SLMs) by addressing their tendency for verbose and repetitive outputs. They introduced a post-hoc temperature scaling technique and a reinforcement learning framework with length penalties, achieving up to a 50% reduction in response length and a 50% increase in accuracy on challenging reasoning benchmarks.

Key insights

  • Supervised fine-tuning alone can inadvertently introduce significant redundancy and repetitive outputs in SLMs, particularly for incorrect answers, hindering token efficiency.
  • Explicitly incorporating length penalties during reinforcement learning can enable SLMs to achieve a superior 'sweet spot' in the accuracy-efficiency trade-off, often improving accuracy while drastically reducing response length.
  • Controlling the probability of the End-of-Sequence (EOS) token, either through post-hoc adjustments or direct reward signals, is a powerful mechanism for effectively managing reasoning trace length and enhancing model efficiency.

Summary by alphaXiv.

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