A controlled study of physics priors in latent world models, imposing conservation structure on learned dynamics to improve long-horizon rollout stability and sample efficiency on rigid-body and PDE-governed benchmarks.
[Working] [Code]
A Bayesian treatment of latent world-model predictors, decomposing epistemic and aleatoric uncertainty in representation space to enable risk-aware model-predictive control for safety-critical planning.
[Working] [Code]
A hybrid framework combining semantic role labeling, rule-based systems, and structured prompts to improve systematic reasoning and compositional generalization in deep learning models.
This paper develops a comprehensive framework to measure the efficiency, scalability, and domain adaptability of RAG systems, bridging academic evaluation with enterprise deployment.