GuardAD: Guarding autonomous-driving MLLMs with Markov safety logic

GuardAD is a model-agnostic safety guardrail method for autonomous-driving MLLMs that models driving safety as evolving Markov logic states; in experiments it reduced the accident rate by 32.07% and slightly improved task performance by 6.85%. It continuously derives safety predicates through neuro-symbolic logic formalization and n-th order Markov logic induction, and can infer latent risks beyond single-step observations. Logic-driven action revision corrects actions without modifying the underlying MLLM, and it was validated through closed-loop simulation and real-vehicle testing.