Humanity's Sixth Sense: Benchmarking Intuitive Visual Reasoning in Multimodal Models
The benchmark puts top multimodal models far below humans on quick visual inference.
HSS tests what people infer from images and videos without explicit clues: causes, trajectories, spatial fit, social authority, and abstract patterns. Human participants reached 93.1% accuracy; the strongest reported model, GPT-6-astra, reached 53.6% even at maximum reasoning effort. Agentic visual manipulation helped but did not close the gap. Source: HF Daily Papers' note.
HSS tests what people infer from images and videos without explicit clues: causes, trajectories, spatial fit, social authority, and abstract patterns. Human participants reached 93.1% accuracy; the strongest reported model, GPT-6-astra, reached 53.6% even at maximum reasoning effort. Agentic visual manipulation helped but did not close the gap. Source: HF Daily Papers' note.
score 5