ICML · 2026
World Models in Pieces: Structural Certification for General Agents
Proceedings of the International Conference on Machine Learning
Which parts of an agent's world model can support reliable planning?
Structural certification links performance on compositional goals to local guarantees on world model transitions. The results identify reliable pieces of a model without requiring universal competence, and establish limits on the precision of such guarantees.
Research themes
Connections
From language context to constrained actions
INSTRUCT MPC translates human instructions into disturbance predictions. World model certification and VigilMPC study complementary ways of deciding which learned models can support reliable planning and safe updates.
Open questions
Certified contextual control
Can contextual and world model updates be certified before they affect a physical system?
Explore open directionsCite this paper
Yikai Lu, Yifei Wu, Xinyu Lu, Tongxin Li. World Models in Pieces: Structural Certification for General Agents. Proceedings of the International Conference on Machine Learning, 2026. https://arxiv.org/abs/2606.24842
BibTeX
@inproceedings{tongxin-world-models,
title = {{World Models in Pieces: Structural Certification for General Agents}},
author = {Yikai Lu and Yifei Wu and Xinyu Lu and Tongxin Li},
year = {2026},
booktitle = {Proceedings of the International Conference on Machine Learning},
url = {https://arxiv.org/abs/2606.24842}
}
Related papers
VigilMPC: Certified Online Tube Model Predictive Control for Safe Real-Time Neuromodulation
VigilMPC accepts a predictor update only when a backup plan remains feasible under tightened constraints. Recursive feasibility and constraint satisfaction follow under the tube MPC assumptions. Evaluation uses simulated neural dynamics and a switching benchmark, rather than clinical deployment.
Can a predictive controller safely update its learned dynamics online? model predictive control MPC stability recursive feasibility distributed control system level synthesis learning augmented algorithms algorithms with predictions competitive analysis robustness consistency online optimizationINSTRUCT MPC: A Human-LLM-in-the-Loop Framework for Context-Aware Control
A language to distribution module converts contextual instructions into disturbance predictions for MPC. The framework closes the loop between instructions, predictions, and control, with a regret analysis for linear dynamics under the stated training assumptions.
How can human instructions improve predictive control? large language models LLM agents contextual control world models reinforcement learning dueling bandits model predictive control MPC stability recursive feasibility distributed control system level synthesis AI for energy electric vehicle charging demand response renewable energy load forecasting decarbonization