Logic LLM: Logic-Regularized Verifier Elicits Reasoning from LLMs

This project is the implementation of Logic-Regularized Verifier Elicits Reasoning from LLMs.

NOTE: The project is still a work in progress—for instance, the feature generation code is not yet included. We are actively refining the codebase and enhancing the project.

feature (hidden states from LLMs)

The feature has a shape of (num_data, N, 2, d, num_layers), where:

• num_data: Number of data samples
• N: Sampling count
• 2: Two contrastive assertions
• d: Feature dimension
• num_layers: Number of LLM layers

We provide some feature data (Google Drive) for running the code.

The feature data is based on GSM8K and Mistral-7B-Instruct-v0.3. We randomly selected 1,500 samples from the GSM8K training set and the full 1,319 samples from the test set. Only the last-layer features were extracted.

train & test

sh train.sh

Acknowledgement

CCS: https://github.com/collin-burns/discovering_latent_knowledge

CoT-Decoding: https://github.com/shirley-wu/cot_decoding

MATH: https://github.com/hendrycks/math

Cite

@inproceedings{wang2025logic,
  title={Logic-Regularized Verifier Elicits Reasoning from LLMs},
  author={Wang, Xinyu and Sun, Changzhi and Cheng, Lian and Wu, Yuanbin and Zhang, Dell and Wang, Xiaoling and Li, Xuelong},
  booktitle={Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers)},
  pages={32617--32630},
  year={2025}
}