Reproduce “The Perceptual Bandwidth Bottleneck in Vision-Language Models: Active Visual Reasoning via Sequential Experimental Design” on OpenResearch

About this paper

Anjie Liu, Ziqin Gong, Yan Song, Yuxiang Chen, Xiaolong Liu, Hengtong Lu, Kaike Zhang, Chen Wei and 1 more

Vision-Language Models often face a "perceptual bandwidth bottleneck" in fine-grained tasks. This work introduces FOVEA, an active visual reasoning framework based on sequential Bayesian optimal experimental design, which consistently improves VLM performance across various high-resolution benchmarks, achieving a 2.6% mean score increase over ReAct baselines with a 30B Qwen3-VL model.

Key insights

• The "perceptual bandwidth bottleneck" is a fundamental limitation in VLMs, where fixed token budgets prevent simultaneous processing of global context and fine-grained details.
• Active visual information acquisition, analogous to biological foveation, can be framed as an S-BOED problem, providing a principled approach beyond passive or heuristic methods.
• The "information cliff" phenomenon, where visual information gain is super-additive, necessitates non-myopic planning for effective visual evidence acquisition, as greedy approaches are often insufficient.

Summary by alphaXiv.