AIAny
Icon for item

An Exam for Active Observers

Quantifies active visual observation in multimodal LLMs with ActiveVision, a 17-task benchmark that forces repeated perception rather than one-shot description. Finds frontier MLLMs fail badly (top model 10.6% vs humans 96.1%) and that model-generated vision code does not close the gap.

Introduction

Active, iterative re-inspection is fundamental to human vision, but standard vision–language benchmarks implicitly reward one-shot descriptions. This paper makes that missing capability measurable: ActiveVision forces models to perform repeated visual perception across tasks and shows a large, persistent gap between human active observation and today's MLLMs.

Key Findings
  • Benchmark design: ActiveVision comprises 17 tasks in 3 categories crafted to require sequential observations and information gathering rather than a single static caption or label.
  • Empirical gap: The best evaluated model (GPT-5.5 at its highest reasoning tier) solves 10.6% of items and scores zero on 11 of 17 tasks; Claude Fable 5 solves 3.5%; three human participants average 96.1%.
  • Code-generation limitation: Allowing models to write and run vision-processing code improves some cases but remains unreliable on realistic imagery, and detecting those failures itself requires active perception.
What This Means

Active visual observation is not an emergent capability in current frontier MLLMs at useful reliability levels. Benchmarks and training objectives that assume single-pass perception can overestimate model visual understanding. Closing the perception–reasoning loop likely requires architectural and training changes that explicitly support iterative inspection, verification, and tool-use monitoring.

Who Should Care & Trade-offs

Great fit if you design or evaluate multimodal systems where repeated inspection, verification, or question-driven vision is required—for example, robotic perception pipelines, visual QA with follow-up queries, and safety-critical inspection tasks. Look elsewhere if your use case is genuinely one-shot description (simple captioning or static labeling), since ActiveVision targets sequential perception and will be overly pessimistic for strictly single-pass benchmarks.

Information

  • Websitearxiv.org
  • AuthorsJiarui Zhang, Muzi Tao, Shangshang Wang, Ollie Liu, Xuezhe Ma, Willie Neiswanger
  • Published date2026/07/17

More Items

Proposes Riemannian Isometric Policy Optimization (RIPO) to fix exploration collapse in PPO-style RL for LLMs by aligning policy updates with the policy manifold's Riemannian geometry, improving exploration–exploitation balance and optimization stability across competition benchmarks.

Studies train-time knowledge injection via hypernetworks that generate fixed LoRA adapters from large fact corpora, empirically characterizing power-law scaling across hypernetwork depth, width, and target model size and reporting improved OOD generalization.

Extrapolates long video sequences from very short contexts by restoring memory-writing supervision in autoregressive video diffusion models using a two-pass Self Gradient Forcing (SGF). SGF records a no-gradient rollout at a sampled denoising exit and then recomputes KV context in a second parallel pass so future losses teach earlier latent writes, enabling minutes-long extrapolation from ~5s windows.