Foveated Reasoning: Stateful, Action-based Visual Focusing for Vision-Language Models

TL;DR

This paper introduces Foveated Reasoner, a vision-language model that mimics human foveation by selectively focusing high-resolution analysis on important regions, improving accuracy with fewer visual tokens.

Contribution

It unifies foveation and reasoning in a single autoregressive framework, trained with reinforcement learning to optimize selective high-resolution evidence acquisition.

Findings

Achieves stronger accuracy under limited visual tokens across benchmarks.

Learns effective foveation policies that focus on important regions.

Reduces compute overhead by selectively acquiring high-resolution evidence.

Abstract

Vision-language models benefit from high-resolution images, but the increase in visual-token count incurs high compute overhead. Humans resolve this tension via foveation: a coarse view guides "where to look", while selectively acquired high-acuity evidence refines "what to think". We introduce Foveated Reasoner, an autoregressive vision-language framework that unifies foveation and reasoning within a single decoding trajectory. Starting from a low-resolution view, the model triggers foveation only when needed, retrieves high-resolution evidence from selected regions, and injects it back into the same decoding trajectory. We train the method with a two-stage pipeline: coldstart supervision to bootstrap foveation behavior, followed by reinforcement learning to jointly improve evidence acquisition and task accuracy while discouraging trivial "see-everything" solutions. Experiments show that the method learns effective foveation policies and achieves stronger accuracy under tight visual-token budgets across multiple vision-language benchmarks.