Active Vision-Language Models via Sequential Experimental Design

A new arXiv paper proposes a framework to overcome the perceptual bandwidth bottleneck in Vision-Language Models (VLMs) by framing visual perception as a sequential decision-making process. The authors draw on active vision and information foraging paradigms, formalizing the problem as a sequential Bayesian optimal experimental design (S-BOED) challenge. They develop tractable approximations for continuous gigapixel spaces, balancing spatial coverage and resolution. A training-free inference strategy is presented as a practical instantiation of the S-BOED objective, designed as a flexible template that accommodates arbitrary optimization algorithms. The paper is available at arXiv:2605.01345.