WorldOracle is a benchmark that evaluates action‑conditioned world models (ACWMs) by comparing their predicted rollouts to paired real‑world executions. It assesses models across four dimensions—physical state correctness, spatial alignment, key interaction fidelity, and multi‑view coherence—using curated training and evaluation suites that include long, multi‑segment episodes with both successful and failed actions. Experiments on seven ACWMs highlight distinct strengths and failure modes, providing a systematic foundation for advancing manipulation modeling.
By Haoyi Duan, Liliang Chen, Nan Wang, Yang Zhang, Yuhang Lu, Xiaotong Chen, Wei-Cheng Tseng, Shenyuan Gao, Zhaoshuo Li, Jingyi Jin, Pooya Jannaty, Dhruv Shah, Yuke Zhu, Ming-Yu Liu, Maoqing Yao, Guanghui Ren
arXiv:2610.08791v1 Announce Type: new
Abstract: Video world models can produce visually convincing yet physically inconsistent sequences, raising concerns about their reliability for prediction and p...
By Mingju Gao, Qingle Liu, Yuzhao Peng, Xinjie Lin, Ziming Qin, Zheng Jiang, Wenyi Li, Calvin Xiao, Youjie Zheng, Kaisen Yang, Qinhuai Na
arXiv:2608. 05948v1 Announce Type: new Abstract: Physics engines facilitate large-scale training and evaluation for embodied intelligence, while generative video world models are emerging as implicit simulators of future states and interactions.
By Shuai Wang, Yaxin Feng, Xuekun Jiang, Shihan Tian, Ningyu Yan, Xing Shen, Chaoyang Lyu, Hui Wang, Yunsong Zhou, Hanqing Wang, Jiangmiao Pang, Yang Xiang, Xing Gao, Chunhua Shen, Weinan Zhang
InfiniteWorld is a new benchmark that pairs physical‑reasoning questions with simulator videos whose physical parameters are systematically altered, allowing researchers to test whether models rely on visual evidence or on familiar physics. The benchmark’s generation pipeline connects simulation, scene screening, paired video translation, question construction, and input checks, and it includes five tasks—law identification, parameter estimation, temporal reasoning, analytic counterfactual comparison, and law induction. Evaluation across five matched input settings (text‑only, ordered, shuffled observations) reveals that video‑language models often default to Earth‑normal physics, showing strong accuracy on normal questions but poor recognition of altered dynamics.
"whyItMatters":"The benchmark provides a measurable way to separate visual evidence from familiar physics, exposing persistent biases in current video‑language models toward Earth‑normal rules."
By Tong Zhang, Yuxuan Cheng, Yun Peng
JEPA-x is a cross‑predictive physics grounding method that aligns visual latent dynamics with privileged physical trajectories. By treating visual observations and physical states as two views of the same action‑conditioned trajectory and sharing a predictor, it forces the model to learn a common transition rule for both modalities. The physical branch is only used during training, so deployment incurs no extra cost, and the approach significantly reduces rollout drift and boosts control success across a multi‑task suite.
By Kehan Wen, Ziming Li, Siyuan Luo, Fan Shi
Physics engines facilitate large-scale training and evaluation for embodied intelligence, while generative video world models are emerging as implicit simulators of future states and interactions. However, existing evaluations of physical fidelity are often conducted in isolation and rely heavily on perceptual similarity or human judgments, providing limited insight into which physical principles or parameters are violated.