DominoBench Tests AI’s Physical Reasoning from Still Images
Emilio Ferrara’s paper introduces DominoBench, a benchmark for testing whether vision-language models can predict outcomes in 2D rigid-body scenes from still images. It contains 854 items in 427 minimal pairs, with outcomes computed by a physics simulator and a four-level hint ladder ending in the decisive calculation. Tests of 30 models from 10 families found that unaided models did not predict outcomes reliably. Frontier models reached at most 72% accuracy for the person whose fate changes and 34% when both members of a pair had to be correct.
The results show that fluent scene descriptions do not guarantee reliable physical predictions. When the decisive calculation and required quantities were supplied, the same models reached 82–92% accuracy for the deciding person, indicating that the main difficulty was identifying and computing the decisive quantity.
Computer Science > Computer Vision and Pattern Recognition
arXiv:2611.20385 (cs)
[Submitted on 16 Nov 2026]
Title:DominoBench: Evaluating Multi-Step Physical Reasoning from Still Images
Authors:Emilio Ferrara
View a PDF of the paper titled DominoBench: Evaluating Multi-Step Physical Reasoning from Still Images, by Emilio Ferrara