OOD-MMSafe: Advancing Multimodal Large Model Safety from Harmful Intent to Hidden Consequences

The researchers propose the OOD-MMSafe benchmark, using 455 query-image pairs to evaluate the ability of multimodal large models to identify potential harm in context-dependent causal chains, extending the safety focus from harmful intent to hidden consequences. The analysis shows that frontier models generally have causal blind spots, with high-capacity closed-source models having a failure rate of up to 67.5%, and that a preference ceiling exists, i.e., as model capacity grows, static alignment brings format-centric failures rather than improvements in safety reasoning. To this end, the authors propose the CASPO framework, which uses the model's own reasoning as a dynamic reference for token-level self-distillation rewards; in experiments, Qwen2.5-VL-7B's risk identification failure rate drops to 7.3% and Qwen3-VL-4B's to 5.7%, while overall effectiveness is maintained.