SafetyALFRED Benchmark Reveals Multimodal LLM Agents Cannot Proactively Address Safety Hazards in Interactive Environments

arXiv·medium signal

Torres-Fonseca, Deng et al. introduce SafetyALFRED, an embodied agent benchmark built on ALFRED that evaluates whether multimodal LLM agents can proactively identify and address safety hazards while completing household tasks. Results show current MLLMs are insufficient at safety-conscious planning, frequently executing dangerous actions (like leaving stoves on or handling sharp objects unsafely) even when they can complete the task itself.