Reproduce “Measuring Reasoning Utility in LLMs via Conditional Entropy Reduction” 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

Xu Guo

This study introduces conditional entropy reduction as an internal, training-free metric to quantify the utility of reasoning steps in Large Language Models (LLMs). It demonstrates a strong correlation between the decrease in this entropy and the correctness of the final answer, revealing that incorrect reasoning paths typically show no such reduction and are often longer than correct ones.

Key insights

  • A consistent decrease in conditional entropy on the answer span reliably indicates effective reasoning and correlates strongly with correct final answers.
  • Longer LLM reasoning chains, especially those leading to incorrect answers, are often inefficient and do not translate to better outcomes, challenging the 'more steps equal better reasoning' assumption.
  • Human reasoning strategies differ from LLMs, characterized by a rapid, early reduction in uncertainty compared to the LLMs' more gradual approach.
  • Commonly used metrics like semantic similarity and cross-entropy are unreliable for gauging reasoning utility or predicting answer correctness in LLMs.

Summary by alphaXiv.

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