Mechanisms of Misgeneralization in Physical Sequence Modeling

Recent physics foundation models claim general spatiotemporal forecasting ability, yet their evaluations often collapse performance into a single average score under a fixed training distribution. This makes it difficult to determine whether a model has learned generalizable physical dynamics or only performs well under particular settings. We construct a benchmark with 8 physical dynamics, 3 training-data mixtures, and 25 test regimes induced by dynamic-scale and initial-condition complexity shifts, covering in-distribution, distribution-shift, and out-of-distribution settings. We evaluate five physics foundation model architectures and four model variants per architecture (scratch and three pretrained sizes), resulting in 60,000 measurements. Our results show that current physics foundation models behave as conditional rather than universal generalists: their generality depends on the physical regime, temporal scale, initial-condition setting, pretraining, model size, and architecture. Improving the training data distribution only partially mitigates this limitation. Pretraining and scaling are also unable to reliably remove their ability biases. We argue that improving physics foundation models requires moving beyond scaling models or expanding data, toward learning mechanisms that better capture transferable physical knowledge across regimes, temporal scales, and distribution shifts.

Physical interactions follow a long-tailed distribution: a set of common and regular interactions dominates human experience and visual data, while a broad spectrum of rare and irregular interactions remains underrepresented. Although recent visual world models, including image and video generation models, achieve impressive realism on existing benchmarks, they primarily focus on simulating common physical interactions. This raises a central question: Do current visual world models internalize and generalize physical principles? In this work, we introduce Tailor-Bench, a benchmark that challenges world models to simulate irregular physical interactions. To enable systematic evaluation, we design three scenario modes that progressively challenge model reasoning: Regular scenarios reflect common tool-task pairs, Unconventional scenarios replace conventional tools with attribute-compatible substitutes to test affordance generalization, and Impossible scenarios introduce attribute-violating tools to probe constraint awareness. Additionally, we design two complementary settings under a unified evaluation protocol: predictive generation requires inferring outcomes without guidance, while descriptive generation specifies the target outcome for faithful realization. Our experimental results reveal a clear long-tail gap in physical world modeling: performance degrades from Regular to Unconventional and Impossible scenarios, indicating limited generalization beyond common interactions. Failure analysis further shows that models rely on superficial visual patterns: image models fail to realize correct state changes, while video models further suffer from temporal inconsistencies.

Monolithic world models predict the entire next state at every step, spending capacity re-predicting the static majority of a scene and injecting error into it. We ask whether explicitly modeling change (a per-object change gate plus a residual delta head that perturbs only the objects the gate flags) is a more effective and interpretable bias for physical prediction and control. On a MuJoCo tabletop pushing benchmark scaling from 3 to 8 objects, the sparse/residual model predicts next-state poses 2.5 to 4.6 times more accurately than a dense multilayer perceptron at 8.6 to 11.1 times fewer parameters, sustains change-detection F1 of 0.80 to 0.87 where the dense baseline is degenerate, transfers across object counts with zero retraining (99.4 percent F1 retention), and reaches about 90 percent of its full-data accuracy with a quarter of the data. In autoregressive rollout it compounds far less error, hugging the no-motion floor while the dense model drifts. Finally, inside a sampling-based planner, prediction-only models fail (though a true-simulator oracle solves the task with the identical planner, confirming the planner is sound), but once featurized and trained for the states a planner visits, the sparse model begins to plan (0.23 plus or minus 0.06 success over three seeds) while the dense monolith stays at zero at every seed. Modeling what changes, rather than re-predicting the whole world, is a simple, effective bias for object-centric physical AI; code, data generators, and all checkpoints will be released upon publication.