通过视觉信息基础实现多模态幻觉控制
Multi-Modal Hallucination Control by Visual Information Grounding
AI总结：
针对视觉语言模型因过度依赖语言先验而产生幻觉的问题，提出无需额外训练的推理时采样方法 M3ID，并可与 DPO 结合提升视觉依赖，显著降低幻觉并提高 VQA 准确率。
摘要：
生成式视觉语言模型（VLMs）容易生成听起来合理但并未始终基于输入图像的文本答案。我们研究了这种现象（通常被称为“幻觉”），并表明它源于对语言先验的过度依赖。具体而言，我们发现随着生成的 token 增多，对视觉提示的依赖性会降低，且该行为与幻觉的出现高度相关。为减少幻觉，我们引入了多模态互信息解码（M3ID），一种用于提示放大的新采样方法。M3ID 放大了参考图像相对于语言先验的影响，从而倾向于生成与视觉提示具有更高互信息的 token。M3ID 可在推理时应用于任何预训练的自回归 VLM，无需进一步训练且计算开销极低。若允许训练，M3ID 可与直接偏好优化（DPO）结合，在无需任何标签的情况下提升模型对提示图像的依赖。实证表明，我们的算法在保持预训练 VLMs 流畅性与语言能力的同时，通过缓解视觉无依据的回答减少了幻觉。具体而言，对于 LLaVA 13B 模型，M3ID 与 M3ID+DPO 分别将图像描述任务中幻觉对象的比例降低了 25% 和 28%，并在 POPE 等 VQA 基准上将准确率提升了 21% 和 24%。
英文摘要：
Generative Vision-Language Models (VLMs) are prone to generate plausible-sounding textual answers that, however, are not always grounded in the input image. We investigate this phenomenon, usually referred to as "hallucination" and show that it stems from an excessive reliance on the language prior. In particular, we show that as more tokens are generated, the reliance on the visual prompt decreases, and this behavior strongly correlates with the emergence of hallucinations. To reduce hallucinations, we introduce Multi-Modal Mutual-Information Decoding (M3ID), a new sampling method for prompt amplification. M3ID amplifies the influence of the reference image over the language prior, hence favoring the generation of tokens with higher mutual information with the visual prompt. M3ID can be applied to any pre-trained autoregressive VLM at inference time without necessitating further training and with minimal computational overhead. If training is an option, we show that M3ID can be paired with Direct Preference Optimization (DPO) to improve the model's reliance on the prompt image without requiring any labels. Our empirical findings show that our algorithms maintain the fluency and linguistic capabilities of pre-trained VLMs while reducing hallucinations by mitigating visually ungrounded answers. Specifically, for the LLaVA 13B model, M3ID and M3ID+DPO reduce the percentage of hallucinated objects in captioning tasks by 25% and 28%, respectively, and improve the accuracy on VQA benchmarks such as POPE by 21% and 24%.
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。
