预填充阶段干预用于缓解大视觉-语言模型中的幻觉
Prefill-Time Intervention for Mitigating Hallucination in Large Vision-Language Models
AI总结：
本文提出PTI方法，在预填充阶段干预KV缓存以减少幻觉，通过模态感知方向修正错误表示，提升模型可靠性。
摘要：
大型视觉-语言模型（LVLMs）在视觉-文本理解上取得了显著进展，但其可靠性受到幻觉（生成事实错误或不一致响应）的严重影响。尽管近期使用引导向量的研究在减少幻觉方面表现出色，但其仍存在显著挑战：它们无意中加剧了残余幻觉的严重性。我们归因于其仅专注于解码阶段，其中错误以自回归方式累积并逐步恶化后续的幻觉输出。为了解决这个问题，我们提出了预填充阶段干预（PTI），一种新的引导范式，在预填充阶段仅干预一次，增强错误累积前的初始键值（KV）缓存。具体而言，PTI是模态感知的，为视觉和文本表示分别推导出不同的方向。这种干预被解耦以引导键向视觉基础对象和值以过滤背景噪声，从而在源头上纠正易产生幻觉的表示。广泛实验显示PTI在缓解幻觉方面具有显著表现，并且在多样化的解码策略、LVLMs和基准测试中具有通用性。此外，PTI与现有解码阶段方法正交，能够实现即插即用的集成并进一步提升性能。代码可在：https://github.com/huaiyi66/PTI获取。
英文摘要：
Large Vision-Language Models (LVLMs) have achieved remarkable progress in visual-textual understanding, yet their reliability is critically undermined by hallucinations, i.e., the generation of factually incorrect or inconsistent responses. While recent studies using steering vectors demonstrated promise in reducing hallucinations, a notable challenge remains: they inadvertently amplify the severity of residual hallucinations. We attribute this to their exclusive focus on the decoding stage, where errors accumulate autoregressively and progressively worsen subsequent hallucinatory outputs. To address this, we propose Prefill-Time Intervention (PTI), a novel steering paradigm that intervenes only once during the prefill stage, enhancing the initial Key-Value (KV) cache before error accumulation occurs. Specifically, PTI is modality-aware, deriving distinct directions for visual and textual representations. This intervention is decoupled to steer keys toward visually-grounded objects and values to filter background noise, correcting hallucination-prone representations at their source. Extensive experiments demonstrate PTI's significant performance in mitigating hallucinations and its generalizability across diverse decoding strategies, LVLMs, and benchmarks. Moreover, PTI is orthogonal to existing decoding-stage methods, enabling plug-and-play integration and further boosting performance. Code is available at: https://github.com/huaiyi66/PTI.
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。
