跳转至

文章背景与核心概要

大语言模型（LLM）的监督微调（SFT）虽然是模型适配的标准方法，但它往往会带来一个严重的缺陷：加剧模型对预训练阶段已习得知识的事实性幻觉。本文深入探讨了这一现象的根本原因，指出重叠语义表征之间的干扰是导致幻觉的主要驱动因素，并提出了切实可行的缓解策略——包括基于自蒸馏的 SFT 方法和选择性参数冻结，从而在吸收新知识与保留原有知识之间取得平衡。

Why Fine-Tuning Encourages Hallucinations and How to Fix It

arXiv: 2604.15574 [cs.CL]
Published in: CoLM 2026 Conference
Authors: Guy Kaplan, Zorik Gekhman, Zhen Zhu, Lotem Rozner, Yuval Reif, Swabha Swayamdipta, Derek Hoiem, Roy Schwartz

arXiv: 2604.15574 [cs.CL]
Published in: CoLM 2026 Conference
Authors: Guy Kaplan, Zorik Gekhman, Zhen Zhu, Lotem Rozner, Yuval Reif, Swabha Swayamdipta, Derek Hoiem, Roy Schwartz

📌 Executive Summary

Supervised fine-tuning (SFT) is a standard method for adapting large language models (LLMs), but it often introduces a critical flaw: an increase in factual hallucinations regarding knowledge acquired during pre-training.

This research investigates the root causes of these hallucinations, identifies interference among overlapping semantic representations as a primary driver, and proposes actionable mitigation strategies—including a self-distillation-based SFT method and selective parameter freezing—to balance new knowledge acquisition with the preservation of pre-existing knowledge.

📌 Executive Summary

Supervised fine-tuning (SFT) is a standard method for adapting large language models (LLMs), but it often introduces a critical flaw: an increase in factual hallucinations regarding knowledge acquired during pre-training.

This research investigates the root causes of these hallucinations, identifies interference among overlapping semantic representations as a primary driver, and proposes actionable mitigation strategies—including a self-distillation-based SFT method and selective parameter freezing—to balance new knowledge acquisition with the preservation of pre-existing knowledge.

📄 Abstract

Large language models are prone to hallucinating factually incorrect statements. A key source of these errors is exposure to new factual information through supervised fine-tuning (SFT), which can increase hallucinations with respect to knowledge acquired during pre-training.

Since these errors arise as a by-product of knowledge degradation, this paper explores whether established continual learning tools can mitigate them: * Self-Distillation SFT: We propose a self-distillation-based supervised fine-tuning method that facilitates effective factual learning while minimizing hallucinations w.r.t. pre-existing knowledge by regularizing output-distribution drift. * Suppressing Factual Plasticity: We show that when new knowledge acquisition is unnecessary, freezing parameter groups preserves task performance while reducing hallucinations. * Mechanistic Insights: We investigate the underlying mechanism—contrasting capacity limitations, behavior cloning, and localized interference. Our experiments reveal that a primary driver is interference among overlapping semantic representations, which self-distillation mitigates and an associative-memory model explains (i.e., forgetting grows with the overlap between new and stored facts).

📄 Abstract

Large language models are prone to hallucinating factually incorrect statements. A key source of these errors is exposure to new factual information through supervised fine-tuning (SFT), which can increase hallucinations with respect to knowledge acquired during pre-training.

Since these errors arise as a by-product of knowledge degradation, this paper explores whether established continual learning tools can mitigate them: * Self-Distillation SFT: We propose a self-distillation-based supervised fine-tuning method that facilitates effective factual learning while minimizing hallucinations w.r.t. pre-existing knowledge by regularizing output-distribution drift. * Suppressing Factual Plasticity: We show that when new knowledge acquisition is unnecessary, freezing parameter groups preserves task performance while reducing hallucinations. * Mechanistic Insights: We investigate the underlying mechanism—contrasting capacity limitations, behavior cloning, and localized interference. Our experiments reveal that a primary driver is interference among overlapping semantic representations, which self-distillation mitigates and an associative-memory model explains (i.e., forgetting grows with the overlap between new and stored facts).

📚 Metadata & Article Links

  • Subjects: Computation and Language (cs.CL); Artificial Intelligence (cs.AI); Machine Learning (cs.LG); Neural and Evolutionary Computing (cs.NE)
  • Cite As: arXiv:2604.15574 [cs.CL] (or arXiv:2604.15574v2 [cs.CL] for the latest revision)
  • DOI: 10.48550/arXiv.2604.15574
  • Submission History:
  • [v1] Thu, 16 Apr 2026
  • [v2] Mon, 31 Aug 2026 (Current Version)

🔗 Full-Text & Tools

📚 Metadata & Article Links

  • Subjects: Computation and Language (cs.CL); Artificial Intelligence (cs.AI); Machine Learning (cs.LG); Neural and Evolutionary Computing (cs.NE)
  • Cite As: arXiv:2604.15574 [cs.CL] (or arXiv:2604.15574v2 [cs.CL] for the latest revision)
  • DOI: 10.48550/arXiv.2604.15574
  • Submission History:
  • [v1] Thu, 16 Apr 2026
  • [v2] Mon, 31 Aug 2026 (Current Version)

🔗 Full-Text & Tools