arXiv 2505.161462025-09-15cs.CVcs.AIcs.CLcs.LG

通过稀疏自编码器引导大型视觉语言模型以缓解幻觉

Steering LVLMs via Sparse Autoencoder for Hallucination Mitigation

School of Computer Science and Engineering, Southeast University（东南大学计算机科学与工程学院）；Key Laboratory of New Generation Artificial Intelligence Technology and Its Interdisciplinary Applications (Southeast University)（东南大学新一代人工智能技术及其交叉应用关键实验室）；Foundation Model Research Center, Institute of Automation, Chinese Academy of Sciences（中国科学院自动化研究所基础模型研究中心）；School of Artificial Intelligence, University of Chinese Academy of Sciences（中国科学院大学人工智能学院）；Department of Computer Science and Technology, Tsinghua University（清华大学计算机科学与技术系）；Wuhan University of Technology（武汉理工大学）；National University of Singapore（新加坡国立大学）

Zhenglin Hua, Jinghan He, Zijun Yao, Tianxu Han, Haiyun Guo, Yuheng Jia, Junfeng Fang

备注：: Accepted to Findings of EMNLP 2025

阅读 PDF

arXiv 页面 ↗HTML 原文 ↗

AI总结：

本文提出SSL方法，利用稀疏自编码器提取解耦的幻觉相关表示，通过引导忠实方向缓解大型视觉语言模型的幻觉，显著优于现有方法且跨架构可迁移。

摘要：

大型视觉语言模型（LVLMs）在多模态任务上取得了显著的性能。然而，它们仍然存在幻觉问题，生成与视觉输入不一致的文本，这在实际应用中带来了重大风险。解决该问题的现有方法侧重于引入外部知识库、对齐训练或解码策略，这些都需要大量的计算成本和时间。近期的研究尝试通过调整LVLMs的内部表示来探索更高效的替代方案。尽管前景广阔，但这些方法可能导致幻觉抑制不足，或引发过度干预从而对正常语义产生负面影响。在本研究中，我们利用稀疏自编码器（SAEs）来识别与忠实度或幻觉密切相关的语义方向，从而提取更精确且解耦的幻觉相关表示。我们的分析表明，沿着识别出的忠实方向进行干预可以缓解幻觉，而沿着幻觉方向进行干预则会加剧幻觉。基于这些见解，我们提出了Steering LVLMs via SAE Latent Directions（SSL），这是一种基于SAE导出的潜在方向的即插即用方法，用于缓解LVLMs中的幻觉。大量实验表明，SSL在缓解幻觉方面显著优于现有的解码方法，同时在不同模型架构间保持可迁移性，且额外时间开销可忽略不计。代码可在 https://github.com/huazhenglin2003/SSL 获取。

英文摘要：

Large vision-language models (LVLMs) have achieved remarkable performance on multimodal tasks. However, they still suffer from hallucinations, generating text inconsistent with visual input, posing significant risks in real-world applications. Existing approaches to address this issue focus on incorporating external knowledge bases, alignment training, or decoding strategies, all of which require substantial computational cost and time. Recent works try to explore more efficient alternatives by adjusting LVLMs' internal representations. Although promising, these methods may cause hallucinations to be insufficiently suppressed or lead to excessive interventions that negatively affect normal semantics. In this work, we leverage sparse autoencoders (SAEs) to identify semantic directions closely associated with faithfulness or hallucination, extracting more precise and disentangled hallucination-related representations. Our analysis demonstrates that interventions along the identified faithful direction can mitigate hallucinations, while those along the hallucinatory direction can exacerbate them. Building on these insights, we propose Steering LVLMs via SAE Latent Directions (SSL), a plug-and-play method based on SAE-derived latent directions to mitigate hallucinations in LVLMs. Extensive experiments demonstrate that SSL significantly outperforms existing decoding approaches in mitigating hallucinations, while maintaining transferability across different model architectures with negligible additional time overhead. The code is available at https://github.com/huazhenglin2003/SSL.

机构、总结、中文摘要均由AI分析而来，可能存在错误，请以论文正文为准。