Humanity's Sixth Sense: Benchmarking Intuitive Visual Reasoning in Multimodal Models

Humanity's Sixth Sense. Benchmarking Intuitive Visual Reasoning in Multimodal Models. By Xingang Guo and twenty-eight others, from Scale AI, Elorian, and one other institution. Published on the eighth of October. No code repository is linked to it.

People take in far more from a scene than the scene literally shows. One glance suggests what caused it and where things are heading. A quick look tells you whether a vehicle will fit between two parked cars. A few seconds of video reveal who holds authority in a room. A brief clip can surface subtle abstract patterns, like unwritten rules or hidden labels. The authors call this humanity's sixth sense: an intuitive kind of reasoning that recovers implicit information beyond what the senses directly record.

That intuition is fast and needs no examples, and everyday navigation and social interaction rest on it. For Multimodal Large Language Models that are deployed alongside people, it is a vital capability. Yet existing visual benchmarks aim elsewhere. They test either deliberate, expert-level analysis in academic and mathematical domains, or low-level perception. The intuitive reasoning in between, the kind people do without effort, has gone largely untested.

Humanity's Sixth Sense, or HSS, is a benchmark built to measure that intuitive visual reasoning. Its prompts probe four kinds of implicit structure that people infer at a glance: temporal, spatial, social and abstract. Each maps onto the everyday examples. Causes and trajectories are temporal. Whether a car fits is spatial. Who holds authority is social. And unwritten rules or hidden labels are abstract.

The benchmark spans diverse inputs, both images and video. Items are organised under a structured taxonomy, and each one is paired with prompts written by humans. The questions are not about naming what is visible in the frame. They are about what a person would read into it without being told. That is the difference from the two kinds of benchmark already in use. There is no long chain of expert analysis to carry out, and no low-level detail to pick out. There is an inference a person makes almost immediately.

Frontier MLLMs fall short of people. Human participants reach ninety-three point one percent accuracy. The strongest model, GPT-6-astra, reaches only fifty-three point six percent, and that is at its maximum reasoning effort. The gap stands out against how well these same models do elsewhere: they excel at many complex tasks that demand advanced perception and knowledge, yet struggle significantly with visual tasks that humans find intuitive.

The authors also try an agentic setup that applies dynamic visual manipulation to HSS. That narrows the gap. It does not close it. Giving the model more ways to work on the input helps, but the distance between ninety-three point one percent and what models reach remains. Neither more reasoning effort nor an agentic setup, as tested here, brings models up to what people do at a glance.

The benchmark's claim is that intuition can be measured as its own axis, separate from expertise and separate from raw perception. Measured that way, it is a capability that models have not picked up along the way, and the authors frame HSS as directing attention to something that scaling on current benchmarks has so far left behind. A model that can solve hard problems still cannot reliably tell, at a glance, what a person would see without thinking.