Researchers at Hugging Face have released ReactHuman, described as the first physics-grounded benchmark for human-like reactive decision-making in embodied multimodal large language models (MLLMs). The work targets a gap in existing evaluations, which typically probe intuitive physics passively through video question answering or focus on deliberate, long-horizon tasks such as navigation and rearrangement.
In ReactHuman, the evaluated MLLM acts as the brain of a simulated humanoid robot facing sudden household hazards. The benchmark spans 17 event families and more than 1,000 bit-for-bit reproducible scenes, with exact, annotation-free ground truth derived from 240 Hz rigid-body simulation. It also includes adversarial objects whose appearance contradicts their physics, such as a foam anvil or a steel apple.
The authors designed a five-metric suite that scores each reaction along three axes: reasonable, safe, and physically grounded. Every committed plan is physically executed so that decisions have observable consequences.
Evaluating seven representative MLLMs, the team reports that reactive safety is far from solved. Models mishandle roughly one hazard in three, act from fixed dispositions rather than the observed scene, trust appearance over motion, and miss interception points at meter scale even when the chosen action is correct. Notably, none of these failures shrink with model scale.
ReactHuman is positioned as both a fine-grained diagnostic and a scalable training signal toward physically grounded, safety-aware embodied agents. The benchmark is available on Hugging Face at this dataset link.