Mechanisms of Misgeneralization in Physical Sequence Modeling
The paper discusses the issue of physical misgeneralization in generative sequence models used for planning motion in physical domains. It highlights how local errors in deep learning models can lead to incorrect aggregate distributions of physical quantities. The authors propose a kernel-informed intervention to mitigate these errors based on their findings from controlled synthetic tasks.
- ▪Generative sequence models are trained to plan motion in various physical domains, including robotics.
- ▪Physical misgeneralization occurs when local errors in the model propagate, resulting in incorrect distributions of physical quantities.
- ▪The authors develop a data deviation kernel to estimate errors and predict changes in physical quantities.
Opening excerpt (first ~120 words) tap to expand
Computer Science > Machine Learning arXiv:2605.20299 (cs) [Submitted on 19 May 2026] Title:Mechanisms of Misgeneralization in Physical Sequence Modeling Authors:Kento Nishi, Raphael Tang, Karun Kumar, Core Francisco Park, Hidenori Tanaka View a PDF of the paper titled Mechanisms of Misgeneralization in Physical Sequence Modeling, by Kento Nishi and 4 other authors View PDF HTML (experimental) Abstract:Generative sequence models are often trained to plan motion in physical domains, from robotics to mechanical simulations. When constructing a dataset to train such a model, engineers may curate demonstrations to specify how trajectories should be distributed over a physical quantity like travel distance or mechanical energy.
…
Excerpt limited to ~120 words for fair-use compliance. The full article is at arXiv cs.AI.
Discussion
0 commentsMore from arXiv cs.AI
