DominoBench: VLMs can't reliably predict physical outcomes from still images

UniversityofSouthernCalifornia · hf · 2026-11-17

USC researchers release DominoBench, which tests vision-language and multimodal models on multi-step physical reasoning from still images using Rube Goldberg-style setups—e.g. minimal pairs differing only by a counterweight 0.25 grid squares wider, yet yielding opposite outcomes to "who dies if A pushes the ball?" Finding: current models can only reliably predict physical outcomes when the critical calculation and necessary quantities are explicitly provided; without them, they cannot infer chain-reaction dynamics from images alone.

Original post →

More from Research

Research channel →