Search papers, labs, and topics across Lattice.
This paper introduces ActiveVision, a benchmark designed to assess the capacity of multimodal large language models (MLLMs) to engage in active observation, a critical aspect of human vision. The evaluation reveals that leading MLLMs, including GPT-5.5 and Claude Fable 5, perform poorly on this benchmark, solving only 10.6% and 3.5% of tasks respectively, compared to human participants who average 96.1%. These findings underscore a significant gap in the active visual observation capabilities of current MLLMs, highlighting the need for new architectures and training methods that enhance perception-reasoning integration.
Current multimodal large language models struggle with active visual observation, achieving less than 11% accuracy on a benchmark designed to measure this critical capability.
Human vision is a closed loop: gaze is continuously redirected by intermediate hypotheses rather than a single snapshot. Decades of psychophysics and cognitive science have argued that this active observation is essential for a wide range of tasks. Whether today's multimodal large language models (MLLMs) exercise active observation is an empirical question that current vision-language benchmarks do not answer. We introduce ActiveVision, a benchmark that makes active observation measurable for MLLMs, comprising 17 tasks across 3 categories. Tasks are designed to force repeated visual perception rather than a single static description. Frontier MLLMs collapse on ActiveVision: the highest-scoring model we evaluate, GPT-5.5 at the highest exposed reasoning-effort tier, solves only 10.6% of items and scores zero on 11 of the 17 tasks, and even Claude Fable 5, despite topping most reasoning and coding leaderboards, solves just 3.5%, far behind three human participants who average 96.1%. Furthermore, much of the gap persists even when models write and run their own vision code: such code is unreliable on realistic imagery, and catching its failures itself requires the active perception the models lack. Together, these results indicate that current MLLMs lack robust active visual observation, motivating architectures and training objectives that close the perception-reasoning loop.
