Pretraining on everyday photos got a model to 70% on a reasoning test.
Kaiming He's MIT team built a vision-only solver for ARC, a grid puzzle test of abstract reasoning. Starting from a model pretrained on natural images, an ensemble scored 70.2% where a specialised language system scored 71.6%. Frontier models still win easily, with Gemini 3 Pro at 98% on the same benchmark.
Why it mattersAbstract reasoning may be partly a seeing problem, not only a language one. That points to much smaller models for tasks we assumed needed giant ones.
Read in the app — free, in 9 languages
Related stories
An AI trained for $8,000 beat Stratego's greatest player.2026-10-03arXiv will take only two papers a month from each researcher.
2026-10-03AI agents tried to break into Canada's national archive.
2026-10-01Google hid a watermark inside proteins that an AI designed.
2026-10-01Today's robots are cheap enough for 0.3% of the work they can do.
2026-10-01