Ivan Ruchkin traveled to Germany for his first Dagstuhl Seminar, #26421 Advancing Testability and Verifiability of CPS with Neurosymbolic and Large Language Models. He gave an overview of world models and presented open research directions in physically interpretable world models.
The open problems highlighted by Ivan’s talk:
- Comprehensive guarantees:
- Scalability in state dims & WM size
- Physical parameter identification (loss function interference)
- Causal identifiability in latent space
- Consistency between:
- Physical states and image details
- Multimodal sensor signals
- Handling distribution shifts:
- Discriminating hallucinations in WMs and anomalies in reality
- Robustness to covariate shift in sensor data
- Resilience/adaptation to changing real system
- Open world: learning interpretable entities (not handcrafting physical dimensions)