Reproduce “Towards Concise and Adaptive Thinking in Large Reasoning Models: A Survey” 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

Jason Zhu, Hongyu Li

This survey provides a comprehensive overview of methods for achieving concise and adaptive thinking in Large Reasoning Models (LRMs). It categorizes existing research into training-free and training-based approaches, detailing how models can shorten verbose reasoning chains and dynamically adjust their 'thinking' depth based on input difficulty, thereby improving efficiency and practical usability.

Key insights

  • Large Reasoning Models (LRMs) widely exhibit an 'overthinking phenomenon,' generating unnecessarily lengthy reasoning chains for many tasks.
  • Optimal reasoning lengths are non-uniform and vary with input difficulty, necessitating adaptive thinking to balance performance and efficiency.
  • Current research on concise and adaptive LRM thinking falls into two main paradigms: training-free methods (e.g., prompting, pipeline-based, decoding manipulation) and training-based methods (e.g., fine-tuning, reinforcement learning).

Summary by alphaXiv.

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