TSV

Source code for ICML 2025 paper Steer LLM Latents for Hallucination Detection by Seongheon Park, Xuefeng Du, Min-Hsuan Yeh, Haobo Wang, and Yixuan Li

---

Requirements

conda env create -f tsv.yml

---

LLM response generation

Generate responses for each question to construct an unlabeled QA dataset in the wild.

bash gen.sh

---

GT generation

Generate BLEURT score for each QA pair

bash gt.sh

---

Train TSV

Train TSV for hallucination detection.

bash train.sh

---

Citation

@inproceedings{
park2025steer,
title={Steer {LLM} Latents for Hallucination Detection},
author={Seongheon Park and Xuefeng Du and Min-Hsuan Yeh and Haobo Wang and Yixuan Li},
booktitle={Forty-second International Conference on Machine Learning},
year={2025}
}

---

Acknowledgement

We gratefully acknowledge HaloScope, ITI, and ICV for their inspiring ideas and open-source contributions, which served as valuable foundations for this work.