GuardAD: Safeguarding Autonomous Driving MLLMs via Markovian Safety Logic

Researchers introduce GuardAD, a safety framework that enhances autonomous driving systems using multimodal large language models (MLLMs) by incorporating Markovian logic to detect and prevent accidents. The model-agnostic safeguard reduces accident rates by 32% while improving task performance, combining neuro-symbolic logic with dynamic action revision rather than simple action veto mechanisms.

GuardAD addresses a critical vulnerability in autonomous driving systems that increasingly rely on multimodal language models for decision-making. Current safeguard mechanisms employ static logical constraints that fail to account for temporal dynamics in traffic environments where conditions continuously evolve. This research introduces a more sophisticated approach by modeling safety as an evolving Markovian state, enabling the system to reason about emerging hazards beyond what single-step observations can reveal. The neuro-symbolic logic formalization represents safety predicates across different traffic participants and induces them through higher-order Markovian logic, creating a framework that captures complex interactions between vehicles, pedestrians, and environmental factors.

The breakthrough lies in GuardAD's action revision mechanism, which actively refines model outputs rather than simply blocking unsafe actions. This preserves the MLLM's capabilities while layering protection on top, avoiding the performance degradation typical of restrictive safety systems. The empirical results are substantial: 32% accident rate reduction coupled with 6.85% performance improvement demonstrates the framework's ability to enhance both safety and utility simultaneously.

For the autonomous driving and AI safety sectors, this represents meaningful progress toward deployable safeguards for complex AI systems. The validation across multiple benchmarks, closed-loop simulations, and physical vehicle tests suggests practical applicability rather than theoretical elegance alone. As autonomous vehicles move toward real-world deployment, robust safety mechanisms become regulatory and commercial necessities. GuardAD's model-agnostic design means it could apply across different MLLM architectures, increasing its potential adoption value in an industry increasingly standardizing on large language models for perception and decision-making.

• →GuardAD reduces autonomous driving accident rates by 32% while improving task performance by 6.85%, solving the typical tradeoff between safety and capability.
• →The framework uses Markovian logic to detect latent and emerging hazards in dynamic traffic scenarios rather than relying on static safety constraints.
• →Logic-driven action revision actively refines unsafe outputs without modifying the underlying MLLM, enabling compatibility across different model architectures.
• →Validation includes closed-loop simulations and real-world vehicle testing, suggesting practical deployment readiness beyond academic benchmarks.
• →The model-agnostic design allows application across multiple MLLM architectures, increasing potential industry adoption for autonomous systems.