aiqbench analysis
GazeVLM introduces a novel mechanism for active vision in Vision-Language Models by training them to dynamically control their own attention mechanisms via internal "gaze" tags. This approach allows the model to emulate top-down meta-cognitive oversight, effectively balancing localized detail inspection with global scene awareness without relying on external cropping tools or additional visual tokens. The method demonstrates significant performance gains on high-resolution benchmarks, offering a more efficient path toward robust multimodal reasoning compared to current agentic pipelines.
GazeVLM: Active Vision via Internal Attention Control for Multimodal Reasoning
arXiv:2605.07817v2 Announce Type: replace-cross Abstract: Human visual reasoning is governed by active vision, a process where meta-cognitive control drives top-down goal-directed attention, dynamically routing foveal focus toward task-relevant details while maintaining peripheral awareness of the global scene. In contrast, modern Vision-Language Models (VLMs) process visual information passively, relying on the static accumulation of massive token contexts that dilute the visual evidence as the reasoning chain grows. Here we propose GazeVLM, a VLM that learns to exert this oversight over its attention resources by generating gaze actions in its reasoning chain, in the form of tags specifying the coordinates of the region to inspect. During training, each block triggers a continuous suppression bias on the attention logits that dampens features outside the regions gazed at so far, and that lifts at the end of the block, restoring the global view. The model is trained with this bias as a teaching signal, first by Supervised Fine-Tuning on curated gaze-reasoning traces, then through Group Relative Policy Optimization (GRPO) with rewards for correct answers and valid grounding. At deployment, the external suppression is removed and the model, running its unmodified forward pass, steers its own attention to the regions indicated by the blocks, retaining most of the effect of the bias and emulating top-down control of spatial attention. This model thus transitions between global spatial awareness and localized focal reasoning without relying on external agentic tools like cropping, and without adding visual tokens from re-encoded patches. Applied to 4B-parameter backbones, GazeVLM demonstrates strong high-resolution multimodal reasoning on HRBench-4k and HRBench-8k, surpassing its base models by about 4 points, and outperforming agentic multimodal pipelines built around thinking with images by 5 to 12 points.
Abstract:Human visual reasoning is governed by active vision, a process where meta-cognitive control drives top-down goal-directed attention, dynamically routing foveal focus toward task-relevant details while maintaining peripheral awareness of the global scene. In contrast, modern Vision-Language Models (VLMs) process visual information passively, relying on the static accumulation of massive token contexts that dilute the visual evidence as the reasoning chain grows. Here we propose GazeVLM, a VLM that learns to exert this oversight over its attention resources by generating gaze actions in its reasoning chain, in the form of <LOOK> tags specifying the coordinates of the region to inspect. During training, each <LOOK> block triggers a continuous suppression bias on the attention logits that dampens features outside the regions gazed at so far, and that lifts at the end of the block, restoring the global view. The model is trained with this bias as a teaching signal, first by Supervised Fine-Tuning on curated gaze-reasoning traces, then through Group Relative Policy Optimization (GRPO) with rewards for correct answers and valid grounding. At deployment, the external suppression is removed and the model, running its unmodified forward pass, steers its own attention to the regions indicated by the <LOOK> blocks, retaining most of the effect of the bias and emulating top-down control of spatial attention. This model thus transitions between global spatial awareness and localized focal reasoning without relying on external agentic tools like cropping, and without adding visual tokens from re-encoded patches. Applied to 4B-parameter backbones, GazeVLM demonstrates strong high-resolution multimodal reasoning on HRBench-4k and HRBench-8k, surpassing its base models by about 4 points, and outperforming agentic multimodal pipelines built around thinking with images by 5 to 12 points.
Computer Vision and Pattern Recognition (cs.CV); Artificial Intelligence (cs.AI); Computation and Language (cs.CL)
arXiv:2605.07817 [cs.CV]
(or arXiv:2605.07817v2 [cs.CV] for this version)
https://doi.org/10.48550/arXiv.2605.07817 Focus to learn more arXiv-issued DOI via DataCite
arXivLabs is a framework that allows collaborators to develop and share new arXiv features directly on our website.
Both individuals and organizations that work with arXivLabs have embraced and accepted our values of openness, community, excellence, and user data privacy. arXiv is committed to these values and only works with partners that adhere to them.
Have an idea for a project that will add value for arXiv's community? Learn more about arXivLabs.
Full text under the publisher's open licence · 379 words
