@inproceedings{zhang-etal-2025-understanding,
    title = "Understanding the Dark Side of {LLM}s' Intrinsic Self-Correction",
    author = "Zhang, Qingjie  and
      Wang, Di  and
      Qian, Haoting  and
      Li, Yiming  and
      Zhang, Tianwei  and
      Huang, Minlie  and
      Xu, Ke  and
      Li, Hewu  and
      Yan, Liu  and
      Qiu, Han",
    editor = "Che, Wanxiang  and
      Nabende, Joyce  and
      Shutova, Ekaterina  and
      Pilehvar, Mohammad Taher",
    booktitle = "Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers)",
    month = jul,
    year = "2025",
    address = "Vienna, Austria",
    publisher = "Association for Computational Linguistics",
    url = "https://preview.aclanthology.org/ingestion-acl-25/2025.acl-long.1314/",
    pages = "27066--27101",
    ISBN = "979-8-89176-251-0",
    abstract = "Intrinsic self-correction was initially proposed to improve LLMs' responses via feedback solely based on their inherent capability. However, recent works show that LLMs' intrinsic self-correction fails without oracle labels as feedback. In this paper, our research goal is to *interpret LLMs' intrinsic self-correction for different tasks, especially for those failure cases.* By including one simple task and three complex tasks with state-of-the-art (SOTA) LLMs like ChatGPT, Llama, and DeepSeek, we design three interpretation methods to reveal the dark side of LLMs' intrinsic self-correction. We identify intrinsic self-correction can (1) cause LLMs to waver both intermedia and final answers and lead to prompt bias on simple factual questions; (2) introduce human-like cognitive bias on complex tasks. In light of our findings, we also provide two simple yet effective strategies for alleviation: question repeating and supervised fine-tuning with a few samples. We open-source our work at https://x-isc.info/."
}