逐詞元關注多模態生成

摘要

多模态大语言模型（MLLMs）在演化上下文中整合视觉与语言信息，以自回归方式生成回复。现有可解释性研究多聚焦于单个层与电路（即“何处”），而对生成过程中多模态计算在Token级的时间动态（即“何时”）探索不足。为填补这一空白，我们基于语义角色研究注意力转移，逐Token（OTaT）追踪模型对图像、文本、指令及先前生成Token的注意力。我们设计了需在单次回复中显式切换视觉与文本上下文的多模态任务。在两个主流模型家族及四种不同规模的开源MLLM中，我们发现了统一模式：对图像的注意力在需要图像信息的Token处达到峰值，任务转换时指令Token被重新关注，而随着生成推进，对先前生成Token的注意力逐步增强。因果注意力阻断干预验证了这些趋势的功能性作用。我们分析了模型在注意力受扰时的行为，观察到回复出现语言先验回退、跨模态泄露、否认或恢复等现象。最后，基于这一新颖分析所揭示的注意力动态，我们提出了一种简单的测试时干预方法，在恰当时机增强对相关模态的注意力，显著提升了多模态任务表现。

English

Multimodal large language models (MLLMs) generate responses autoregressively, integrating visual and linguistic information in an evolving context. Prior work on interpretability has focused on individual layers and circuits (where), leaving the token-level dynamics of multimodal computation during generation (when) underexplored. We address this gap and study attention shifts as per semantic role; tracking model attention to image, text, instruction, and previously generated tokens, One Token at a Time (OTaT). We introduce multimodal tasks that require explicit switching between visual and textual context within a single response. Across two mainstream model families and four open-weight MLLMs of varying sizes, we establish consistent patterns: attention to image peaks at tokens requiring image-derived information, instruction tokens are revisited during task transitions, and attention to previously generated tokens increases as the generation progresses. Causal attention blocking interventions validate the functional role of these trends. We profile model behavior under disrupted attention and observe responses falling back to language priors, or exhibiting cross-modal leakage, denial, or recovery. Finally, informed of the attention dynamics through our novel analysis, we propose a simple test-time intervention to boost attention to the relevant modality at the right time, significantly improving multimodal task performance.