Why 'Look Again' Is the One Skill Top Vision Models Can't Patch With Code

Three human participants averaged 96.1% accuracy on ActiveVision, a new 17-task benchmark submitted to arXiv on 17 July 2026 by researchers at the University of Southern California. OpenAI's frontier model GPT-5.5, run at its highest reasoning-effort setting, solved 10.6% of items and scored zero on 11 of the 17 tasks. Anthropic's Claude Fable 5, which tops most reasoning and coding leaderboards, finished at 3.5%.

The gap is not a leaderboard fluke. ActiveVision is built so models cannot bootstrap out of the failure by writing their own code, which is the escape hatch that has rescued frontier models on most vision tests. The benchmark asks an AI to keep looking, not to look harder once. That is a category of capability the current generation of multimodal large language models was not built for.

The 17 tasks fall into three groups: Distributed Scanning (counting and locating objects spread across an image), Sequential Traversal (following a path through a scene in the right order), and Visual Attribute Transfer (matching properties from one part of an image to another). All 85 released items are designed to require repeated visual perception rather than a single static description, the way most vision benchmarks do.

The diagnostic the paper, "An Exam for Active Observers" by Muzi Tao, Shangshang Wang, Ollie Liu, Xuezhe Ma, and Willie Neiswanger at arXiv 2607.16165, surfaces is structural. Current multimodal models work by encoding an image once, into a fixed sequence of visual tokens, and then reasoning over that sequence in text. There is no explicit perception-action loop, no way for the model to decide to look at a different region, change its zoom, or check whether the first read was wrong. The authors call this design pattern "passive perceiving" and position ActiveVision as a way to measure the missing skill: being an active observer.

To make sure the test is hard in the right way, the authors built it with a procedural scaffold. Geometry is specified exactly first, then OpenAI's GPT-image-2 re-renders the scaffold as a photorealistic image while preserving positions, counts, labels, and topology. The result is realistic textures absent from the sparse cartoon-like renderings that AI models can already ace. The benchmark, code, and dataset are open on the project page, on GitHub, and on Hugging Face, so anyone can rerun the numbers.

The strongest falsifier for "this is just an agent scaffolding problem" is the paper's own tool-use data. The authors let frontier models write and run their own vision code. Two Claude Code agents and OpenAI's Codex reached only 24.7% to 50.6% accuracy, a real jump over the no-code baselines but a long way from human performance. The gains clustered on the tasks that admit reliable code-based solutions. Traversal stayed hard, and the agents often missed that their tools had failed on realistic imagery. Giving a vision model a Python sandbox is not the same as giving it a perception loop.

The evaluated frontier set is GPT-4 family, Claude Fable 5, GPT-5.5, and Gemini 3.1 Pro, compared against the saturated MMMU-Pro and CharXiv benchmarks. Increasing reasoning effort barely narrowed the gap. GPT-5.5 at xhigh is the headline number; at lower effort it is worse. Claude Fable 5's 3.5% is the sharper finding. The model that leads on reasoning and coding leaderboards, the place where long chains of thought and tool use are supposed to pay off, lands single digits on a vision test built to require multiple looks. Capability does not transfer from text to vision on a test designed to look again.

The applications the paper flags are the ones where "look once, then reason" runs out of runway first: robotics, design, manufacturing, computer use, spatial understanding, scientific discovery. A factory robot that misses the missing bolt on the first pass, a CAD assistant that counts the holes in a bracket wrong, an autonomous agent that clicks the wrong button because the icon was partially occluded, all share the same architectural limitation. The tokenized image is a one-shot summarization.

The 96-vs-10 number is a single benchmark on a single day, and the paper is an arXiv preprint without peer review or independent replication. The framing is the durable part. A high score on a test that only asks the model to look once is not the same skill as a high score on a test that asks the model to look again, and ActiveVision now gives readers a way to tell which is which.