An Exam for Active Observers

The paper introduces ActiveVision, a benchmark for testing whether multimodal large language models can repeatedly redirect visual perception based on intermediate hypotheses instead of producing a one-shot image description. It contains 17 tasks across three categories. GPT-5.5 at the highest exposed reasoning-effort tier solves only 10.6% of items and scores zero on 11 tasks, while Claude Fable 5 solves 3.5%. Three human participants average 96.1%. Allowing models to write and execute vision code does not eliminate the gap, because the code is unreliable on realistic imagery and detecting those failures also requires active perception.