How Does DPO Reduce Toxicity? A Mechanistic Neuron-Level Analysis

Yushi Yang, Filip Sondej, Harry Mayne, Andrew Lee, Adam Mahdi. How Does DPO Reduce Toxicity? A Mechanistic Neuron-Level Analysis. In Christos Christodoulopoulos 0001, Tanmoy Chakraborty 0002, Carolyn Rose, Violet Peng, editors, Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing, EMNLP 2025, Suzhou, China, November 4-9, 2025. pages 29524-29543, Association for Computational Linguistics, 2025. [doi]

@inproceedings{YangSMLM25,
  title = {How Does DPO Reduce Toxicity? A Mechanistic Neuron-Level Analysis},
  author = {Yushi Yang and Filip Sondej and Harry Mayne and Andrew Lee and Adam Mahdi},
  year = {2025},
  doi = {10.18653/v1/2025.emnlp-main.1501},
  url = {https://doi.org/10.18653/v1/2025.emnlp-main.1501},
  researchr = {https://researchr.org/publication/YangSMLM25},
  cites = {0},
  citedby = {0},
  pages = {29524-29543},
  booktitle = {Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing, EMNLP 2025, Suzhou, China, November 4-9, 2025},
  editor = {Christos Christodoulopoulos 0001 and Tanmoy Chakraborty 0002 and Carolyn Rose and Violet Peng},
  publisher = {Association for Computational Linguistics},
  isbn = {979-8-89176-332-6},
}