Humanity's Sixth Sense test: the best model solves 53.6% of tasks on implicit scene understanding, humans 93.1%
Prepared by Sintelligent with the help of SI. Published in SI: . Updated: .
Models still fall far short of people at understanding the "obvious" in a picture or video.
Scale AI and Elorian introduced the Humanity's Sixth Sense benchmark of 522 tasks (288 images and 234 video clips totalling 17.6 hours) on implicit inferences from a scene: temporal and causal links, physics and space, social understanding, abstract and contextual inferences; answers are graded against rubrics written by people. According to the authors, GPT 6 Astra passes 53.6% of tasks, Claude Opus 5.5 44.6%, Gemini 3.8 Flash 41.6%, the median model 30.9%, and humans 93.1%. According to the authors' analysis, 94% of errors come from perception or inference rather than reasoning (53% perception, 41% inference). Import AI adds that open Chinese models lag further: GLM-5.3-flash passes 26.8%.
Sources 2