GuardianBench: A Same-Scene Instruction-Contrastive Benchmark for Latent Contextual Risk in Embodied AI

Hidden safety risks emerge from instruction-scene combinations: GuardianBench reveals that a safe instruction in a safe environment can become dangerous when paired together—a type of risk that previous safety evaluations largely overlooked by focusing only on visual context or execution dynamics.

Current AI models struggle with instruction-specific safety reasoning: State-of-the-art vision-language models achieved only 24.1% accuracy on average when asked to distinguish between safe and unsafe instruction-scene pairs, often approving both or rejecting both instructions in the same scene regardless of the actual instruction content.

Models fail to connect instructions to relevant safety details: The research identified that models' primary failure is not understanding which visual cues in a scene matter for a specific instruction—for example, failing to recognize that "move the object quickly" creates different hazards than "move the object carefully" in the same kitchen environment.

Targeted training can improve safety reasoning: A lightweight training approach called Verdict Log-Odds Supervision (VLOS) substantially improved model performance, suggesting that focused interventions on verdict-level reasoning can enhance embodied AI safety without requiring complete model retraining.

Safety standards can guide AI evaluation: GuardianBench grounds its 3,024 test examples in international safety standards, providing a principled framework for evaluating embodied AI systems against established real-world safety practices rather than ad-hoc criteria.