A team from NVIDIA, MIT and Oxford reports higher scores on four physics benchmark suites after training video models with captions that name the objects, interactions and consequences in each scene. Its tuned Cosmos3-Nano scored above the base model on all four, but the comparison with Google’s Veo 3.1 was mixed.
Physis-Lang makes those relationships explicit. A caption records the objects involved, what triggers an interaction, the physical rules relevant to it, and the event’s sequence and outcome. The team built PhysCapBench, a caption test set with 246 videos and 3,794 human-verified statements about causes, physical laws and effects. A critic checks whether captions cover important events without asserting details that the video does not support.
The team uses critic feedback to refine captioning instructions over repeated rounds while keeping the captioner’s model weights fixed. It then uses an agent to identify each video model’s weak physics categories and retrieve additional footage showing those processes. The resulting set contains 183,000 video-caption pairs, including 112,000 videos selected to target model weaknesses. The researchers used the data to train existing Cosmos and Wan models further; physical descriptions can also guide a model as it generates a clip.
On PhyGenBench’s 100-point scale, tuned Cosmos3-Nano scored 71.04, compared with 61.67 for the base model. In the paper’s comparison with Veo 3.1, it scored higher on three of the four suites, but slightly lower on the full VideoPhy-2 test: 68.02 to Veo’s 68.87.
The comparison depends partly on the researchers’ choice of evaluator. For PhyGenBench and VideoPhy-2, they replaced each benchmark’s original offline evaluator with GPT-5.5, saying it distinguished model results more clearly. The authors’ public repository currently links to the paper and project materials but contains no model implementation or evaluation code. These tests measure benchmark performance, not whether a model reliably predicts unseen physical events or is ready to guide robots in the real world.
Image credit: Sheila Sund via Wikimedia Commons, under CC BY 2.0. The Newton’s cradle is illustrative and was not part of the study.
.jpg){kind=link}