Source: Do NOT Think That Much for 2+3=? (Chen et al., 2024)

Paper characterising 'overthinking' in o1-style models (many tokens on trivial problems) and a self-training fix that shortens reasoning without losing accuracy.

Contents
  • What: research paper (arXiv 2412.21187)
  • Authors: Xingyu Chen, Jiahao Xu, Tian Liang, … Dong Yu (Tencent AI Lab)
  • Date: 2024-12-30 (v2 2025-02-01)
  • Original: arXiv:2412.21187

Summary

The first systematic study of overthinking: o1-like models spend large amounts of compute on simple problems for little or no benefit. It proposes outcome- and process-level efficiency metrics, then shows that self-training can cut that waste.

Key claims

  • Long-CoT models allocate excessive reasoning to easy problems, with minimal accuracy gain.
  • New efficiency metrics judge whether each extra reasoning round adds a correct answer or a genuinely new approach.
  • Self-training toward shorter, sufficient solutions reduces compute while keeping accuracy on GSM8K, MATH500, GPQA and AIME.

Feeds

Test-time compute · Reasoning models

Suggest changes →