GuardAD: Markovian Logic for Safer Autonomous Driving MLLMs
Researchers propose GuardAD, a model-agnostic safeguard for multimodal large language models (MLLMs) in autonomous driving. It uses Neuro-Symbolic Logic Formalization and n-th order Markovian Logic Induction to represent safety as an evolving logical state, enabling detection of emerging hazards beyond single-step observations. This addresses limitations of static safeguards in dynamic traffic environments.
Key facts
- GuardAD is a model-agnostic safeguard for MLLMs in autonomous driving.
- It uses Neuro-Symbolic Logic Formalization to represent safety predicates.
- Safety is formulated as an evolving Markovian logical state.
- n-th order Markovian Logic Induction enables inference of latent hazards.
- The approach addresses limitations of static safeguard mechanisms.
- The paper is published on arXiv with ID 2605.10386.
- The method is designed for dynamic driving environments.
- It does not simply veto unsafe actions but reasons over evolving interactions.
Entities
Institutions
- arXiv