解读与编辑视觉-语言表征以缓解幻觉问题
Interpreting and Editing Vision-Language Representations to Mitigate Hallucinations
AI总结：
针对视觉-语言模型（VLMs）的幻觉问题，通过分析其内部表征发现真实与幻觉物体的概率差异，提出知识擦除算法正交化相关特征，在COCO2014上减少25.7%幻觉并保留性能，还实现零样本分割等新能力。
摘要：
尽管模型规模和训练方法取得进展，幻觉仍是视觉-语言模型（VLMs）的持续挑战。我们研究VLMs的内部表征以解决该问题：将VLMs的内部图像表征投射到其语言词汇表，观察到真实物体的输出概率比幻觉物体更自信；还利用这些输出概率对真实物体进行空间定位。基于此，我们引入知识擦除算法，通过将图像特征与幻觉物体特征线性正交化来消除幻觉。实验显示，在COCO2014数据集上，对模型潜在表征的针对性编辑可将幻觉减少多达25.7%，同时保留性能。研究结果表明，深入理解VLMs的潜在表征能提升可靠性并实现零样本分割等新能力。
英文摘要：
We investigate the internal representations of vision-language models (VLMs) to address hallucinations, a persistent challenge despite advances in model size and training. We project VLMs' internal image representations to their language vocabulary and observe more confident output probabilities on real objects than hallucinated objects. We additionally use these output probabilities to spatially localize real objects. Building on this approach, we introduce a knowledge erasure algorithm that removes hallucinations by linearly orthogonalizing image features with respect to hallucinated object features. We show that targeted edits to a model's latent representations can reduce hallucinations by up to 25.7% on the COCO2014 dataset while preserving performance. Our findings demonstrate how a deeper understanding of VLMs' latent representations can enhance reliability and enable novel capabilities, such as zero-shot segmentation.
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。
