[2603.23495] VISion On Request: Enhanced VLLM efficiency with sparse, dynamically selected, vision-language interactions

Computer Science > Computer Vision and Pattern Recognition arXiv:2603.23495 (cs) [Submitted on 24 Mar 2026] Title:VISion On Request: Enhanced VLLM efficiency with sparse, dynamically selected, vision-language interactions Authors:Adrian Bulat, Alberto Baldrati, Ioannis Maniadis Metaxas, Yassine Ouali, Georgios Tzimiropoulos View a PDF of the paper titled VISion On Request: Enhanced VLLM efficiency with sparse, dynamically selected, vision-language interactions, by Adrian Bulat and Alberto Baldrati and Ioannis Maniadis Metaxas and Yassine Ouali and Georgios Tzimiropoulos View PDF HTML (experimental) Abstract:Existing approaches for improving the efficiency of Large Vision-Language Models (LVLMs) are largely based on the concept of visual token reduction. This approach, however, creates an information bottleneck that impairs performance, especially on challenging tasks that require fine-grained understanding and reasoning. In this work, we challenge this paradigm by introducing VISion On Request (VISOR), a method that reduces inference cost without discarding visual information. Instead of compressing the image, VISOR improves efficiency by sparsifying the interaction between image and text tokens. Specifically, the language model attends to the full set of high-resolution visual tokens through a small, strategically placed set of attention layers: general visual context is provided by efficient cross-attention between text-image, while a few well-placed and dynamically sele...