arXiv 2602.006212026-01-31cs.CV

通过对比神经元引导实现大型视觉语言模型中幻觉分析与缓解

Towards Interpretable Hallucination Analysis and Mitigation in LVLMs via Contrastive Neuron Steering

School of Electronic Engineering, Xidian University, Xi'an, China（西安电子科技大学电子工程学院）；School of Computer Science and Technology, Xidian University, Xi'an, China（西安电子科技大学计算机科学与技术学院）；College of Computer and Information, Hohai University, Nanjing, China（河海大学计算机与信息学院）；Institute for Infocomm Research, A*STAR, Singapore（新加坡资讯研究院）

Guangtao Lyu, Xinyi Cheng, Qi Liu, Chenghao Xu, Jiexi Yan, Muli Yang, Fen Fang, Cheng Deng

阅读 PDF

arXiv 页面 ↗HTML 原文 ↗

AI总结：

通过对比神经元引导方法，分析并缓解大型视觉语言模型中的幻觉问题，提升视觉表示的稳健性和语义基础性。

摘要：

大型视觉语言模型（LVLMs）在多模态理解和生成方面取得了显著成就，但仍然容易产生幻觉。现有的缓解方法主要集中在输出层面的调整，忽略了导致这些幻觉的内部机制。为了更深入地理解，我们采用表征层面的视角，通过引入稀疏自编码器（SAEs）将密集的视觉嵌入分解为稀疏、可解释的神经元。通过神经元层面的分析，我们识别出不同的神经元类型，包括始终激活的神经元和图像特定的神经元。我们的发现表明，幻觉往往源于图像特定神经元的破坏或虚假激活，而始终激活的神经元则基本稳定。此外，有选择地增强或抑制图像特定神经元可以实现对LVLM输出的可控干预，提高视觉基础性并减少幻觉。基于这些见解，我们提出了对比神经元引导（CNS），通过清洁和噪声输入之间的对比分析来识别图像特定神经元。CNS有选择地放大信息神经元，同时抑制扰动引起的激活，产生更加稳健和语义基础的视觉表示。这不仅增强了视觉理解，也有效缓解了幻觉。通过在预填充阶段操作，CNS与现有的解码阶段方法完全兼容。在针对幻觉和一般多模态基准的广泛实验中，CNS在减少幻觉的同时保持了整体多模态理解。

英文摘要：

LVLMs achieve remarkable multimodal understanding and generation but remain susceptible to hallucinations. Existing mitigation methods predominantly focus on output-level adjustments, leaving the internal mechanisms that give rise to these hallucinations largely unexplored. To gain a deeper understanding, we adopt a representation-level perspective by introducing sparse autoencoders (SAEs) to decompose dense visual embeddings into sparse, interpretable neurons. Through neuron-level analysis, we identify distinct neuron types, including always-on neurons and image-specific neurons. Our findings reveal that hallucinations often result from disruptions or spurious activations of image-specific neurons, while always-on neurons remain largely stable. Moreover, selectively enhancing or suppressing image-specific neurons enables controllable intervention in LVLM outputs, improving visual grounding and reducing hallucinations. Building on these insights, we propose Contrastive Neuron Steering (CNS), which identifies image-specific neurons via contrastive analysis between clean and noisy inputs. CNS selectively amplifies informative neurons while suppressing perturbation-induced activations, producing more robust and semantically grounded visual representations. This not only enhances visual understanding but also effectively mitigates hallucinations. By operating at the prefilling stage, CNS is fully compatible with existing decoding-stage methods. Extensive experiments on both hallucination-focused and general multimodal benchmarks demonstrate that CNS consistently reduces hallucinations while preserving overall multimodal understanding.

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。