Attending to Multimodal Generation One Token at a Time

• URL: http://arxiv.org/abs/2607.03738v1
• Date: Sat, 04 Jul 2026 06:53:35 GMT
• Title: Attending to Multimodal Generation One Token at a Time
• Abstract summary: Multimodal large language models (MLLMs) generate responses autoregressively, integrating visual and linguistic information in an evolving context.<n>We study attention shifts as per semantic role; tracking model attention to image, text, instruction, and previously generated tokens, One Token at a Time.<n>We propose a simple test-time intervention to boost attention to the relevant modality at the right time, significantly improving multimodal task performance.
• Score: 25.084919740596803
• License: http://creativecommons.org/licenses/by-nc-sa/4.0/
• Abstract: 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.

Related papers

• Multimodal Large Language Models for Multi-Subject In-Context Image Generation [56.20395856287325]
  Music is the first MLLM specifically designed for textbfMUlti-textbfSubject textbfIn-textbfContext image generation.<n>We enhance the model's understanding of multi-subject semantic relationships through a vision chain-of-thought mechanism.<n>By incorporating complex subject images during training, we improve the model's capacity for chained reasoning.
  arXiv  Detail & Related papers  (2026-04-08T15:37:42Z)
• Kelix Technical Report [86.64551727600104]
  We present Kelix, a fully discrete autoregressive unified model that closes the understanding gap between discrete and continuous visual representations.<n>Recent work has explored discrete visual tokenization to enable fully autoregressive multimodal modeling.
  arXiv  Detail & Related papers  (2026-02-10T14:48:26Z)
• Towards Understanding Multimodal Fine-Tuning: Spatial Features [25.349396112139214]
  Vision-Language Models (VLMs) achieve strong performance on a wide range of tasks by pairing a vision encoder with a pre-trained language model.<n>We present the first mechanistic analysis of VLM adaptation using stage-wise model diffing.
  arXiv  Detail & Related papers  (2026-02-06T18:48:18Z)
• ZSPAPrune: Zero-Shot Prompt-Aware Token Pruning for Vision-Language Models [7.7352936204066]
  We propose a novel, zero-shot method to model visual token pruning as a balance between task relevance and information diversity.<n>Our method achieves performance that matches or surpasses the state-of-the-art with only minimal accuracy loss.<n>These gains are accompanied by significant reductions in GPU memory footprint and inference latency.
  arXiv  Detail & Related papers  (2025-10-20T06:18:47Z)
• Explaining multimodal LLMs via intra-modal token interactions [55.27436637894534]
  Multimodal Large Language Models (MLLMs) have achieved remarkable success across diverse vision-language tasks, yet their internal decision-making mechanisms remain insufficiently understood.<n>We propose enhancing interpretability by leveraging intra-modal interaction.
  arXiv  Detail & Related papers  (2025-09-26T14:39:13Z)
• Seeing is Understanding: Unlocking Causal Attention into Modality-Mutual Attention for Multimodal LLMs [7.03771340666549]
  Vision-language misalignment in Multimodal Large Language Models (MLLMs) is a critical challenge.<n>We propose MapleLeaf AKI, a novel MLLM that unlocks causal attention into modality-mutual attention (MMA) to enable image tokens to attend to text tokens.<n>Our MMA design is intended to be generic, allowing for application across various modalities, and scalable to accommodate diverse multimodal scenarios.
  arXiv  Detail & Related papers  (2025-03-04T13:18:33Z)
• Towards Text-Image Interleaved Retrieval [49.96332254241075]
  We introduce the text-image interleaved retrieval (TIIR) task, where the query and document are interleaved text-image sequences.<n>We construct a TIIR benchmark based on naturally interleaved wikiHow tutorials, where a specific pipeline is designed to generate interleaved queries.<n>We propose a novel Matryoshka Multimodal Embedder (MME), which compresses the number of visual tokens at different granularity.
  arXiv  Detail & Related papers  (2025-02-18T12:00:47Z)
• mPLUG: Effective and Efficient Vision-Language Learning by Cross-modal Skip-connections [104.14624185375897]
  mPLUG is a new vision-language foundation model for both cross-modal understanding and generation. It achieves state-of-the-art results on a wide range of vision-language downstream tasks, such as image captioning, image-text retrieval, visual grounding and visual question answering.
  arXiv  Detail & Related papers  (2022-05-24T11:52:06Z)
• Behind the Scene: Revealing the Secrets of Pre-trained Vision-and-Language Models [65.19308052012858]
  Recent Transformer-based large-scale pre-trained models have revolutionized vision-and-language (V+L) research. We present VALUE, a set of meticulously designed probing tasks to decipher the inner workings of multimodal pre-training. Key observations: Pre-trained models exhibit a propensity for attending over text rather than images during inference.
  arXiv  Detail & Related papers  (2020-05-15T01:06:54Z)

This list is automatically generated from the titles and abstracts of the papers in this site.

This site does not guarantee the quality of this site (including all information) and is not responsible for any consequences.