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.
More from Research
- InclusionMed launches open medical AI benchmark atlas after auditing 388 candidates — The InclusionMed Collaboration · 2026-11-17
- Tencent Hunyuan's Accelerate Bench tests 16 coding agents on 32 CPU optimization tasks — Tencent-Hunyuan · 2026-11-17
- SWE-CoWork: coding agents pass 63% of tests but complete fewer than 8% of tasks — Tokenwave-AI · 2026-11-17
- Centered Ablation: a closed-form fix that boosts refusal-direction ablation by up to 46 points — knoveleng · 2026-11-17
- RAPO: reflection anchors curb visual information decay in long-chain multimodal reasoning — Xuan Gong · 2026-11-17
- RL agent beats Pokémon Red with a <10M-parameter policy, 60,500x smaller than DeepSeek-V3 — Bam4d · 2026-11-17