Verifying Large Language Models' Reasoning Paths via Correlation Matrix Rank

OVERVIEW

This work studies whether a model's internal behavior can indicate the credibility of its own reasoning. It introduces Self-Indicator, a plug-and-play method that uses the rank of a correlation matrix between a problem and a generated reasoning path to reweight candidate solutions.

CITATION

@article{liu2025verifying,
  title = {Verifying Large Language Models' Reasoning Paths via Correlation Matrix Rank},
  author = {Liu, Jiayu and Dai, Wei and Huang, Zhenya and Miao, Ning and Chen, Enhong},
  journal = {arXiv preprint arXiv:2510.24299},
  year = {2025},
  url = {https://arxiv.org/abs/2510.24299}
}