Search papers, labs, and topics across Lattice.
This paper investigates the token-level dynamics of attention in multimodal large language models (MLLMs) during autoregressive generation, focusing on how models integrate visual and linguistic information. By analyzing attention shifts based on semantic roles, the authors reveal that attention to images peaks when image-derived information is needed, while instruction tokens are revisited during task transitions, and attention to previously generated tokens increases over time. The study not only profiles model behavior under disrupted attention but also introduces a test-time intervention that enhances multimodal task performance by optimizing attention allocation to the relevant modality at critical moments.
Attention dynamics reveal that MLLMs can fall back on language priors when visual context is disrupted, highlighting the fragility of multimodal integration.
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.
