マルチモーダル生成における一トークンごとのアテンション

要旨

多模态大语言模型（MLLMs）以自回归方式生成响应，在演化语境中整合视觉与语言信息。先前关于可解释性的研究主要关注单个层与电路（即"何处"层面），却忽略了生成过程中多模态计算的令牌级动态（即"何时"层面）。我们针对这一空白展开研究，考察基于语义角色的注意力转移模式——逐令牌（OTaT）追踪模型对图像、文本、指令及已生成令牌的注意力。我们设计了需要在单次响应中明确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权重大模型中，我们建立了稳定的规律：图像注意力在需要获取图像信息的令牌处达到峰值；指令令牌在任务过渡阶段被重新关注；随着生成进程推进，对已生成令牌的注意力逐渐增强。因果注意力阻断干预实验验证了这些趋势的功能性作用。我们刻画了注意力受扰动下的模型行为，观察到响应会退回到语言先验，或出现跨模态泄漏、否认现象或自我修复。最终，基于我们新分析方法所揭示的注意力动态，我们提出一种简单的测试时干预策略——在恰当时间增强对相关模态的注意力，从而显著提升多模态任务性能。

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.