Reproduce “AutoL2S: Auto Long-Short Reasoning for Efficient Large Language 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
Feng Luo, Yu-Neng Chuang, Guanchu Wang, Hoang Anh Duy Le, Shaochen Zhong, Hongyi Liu, Jiayi Yuan, Yang Sui and 3 more
Researchers from Rice University and collaborators introduced AutoL2S, a distillation framework that enables large language models to adaptively choose between long and short reasoning paths for increased efficiency. This method achieved up to a 71.7% reduction in average reasoning length while maintaining task accuracy across various benchmarks.
Key insights
- Explicitly providing both long and correctness-verified short reasoning paths as supervisory signals, along with a lightweight switching token, effectively teaches LLMs instance-wise reasoning adaptation.
- A GRPO-style reinforcement learning objective, when applied to these adaptively generated long-short rollouts, successfully optimizes for both accuracy and reasoning efficiency.
- Long CoT reasoning paths can initially serve as valuable auxiliary information to guide the learning of concise short reasoning, eventually allowing these mechanisms to operate distinctly for flexible mode switching.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.