Reproduce “DRP: Distilled Reasoning Pruning with Skill-aware Step Decomposition for Efficient Large Reasoning Models” 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

Yuxuan Jiang, Dawei Li, Francis Ferraro

The research introduces Distilled Reasoning Pruning (DRP), a framework that refines verbose reasoning traces from small language models using teacher guidance and skill-aware step decomposition. This approach reduced token usage by up to 64% on GSM8K and 28-44% on out-of-domain tasks, while maintaining or improving Pass@1 accuracy, particularly for 1.5B models with a 12.7% gain on GSM8K.

Key insights

  • Effective training Chain-of-Thoughts (CoTs) must be both informative and structurally consistent with the student's reasoning process to facilitate knowledge transfer and narrow capability gaps.
  • Skill-based decomposition of reasoning steps provides finer-grained, more stable boundaries and semantic coherence, leading to more robust pruning and supervision signals.
  • Direct distillation of overly concise teacher CoTs to verbose student models can lead to a loss of critical intermediate steps and poor generalization to out-of-domain tasks.

Summary by alphaXiv.

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