According to arxiv.org, researchers have developed ReFrame, a training-free framework for improving safety alignment in multimodal large language models (MLLMs) at test time. The research, accepted to EMNLP 2026, addresses the challenge that while MLLMs extend capabilities beyond text, they also make safety alignment more difficult.
According to the paper, existing multimodal safety methods often rely on retraining or internal-state inspection, limiting their applicability to deployed closed-source models. ReFrame takes a different approach by using two agents that share a lightweight locally deployed MLLM: an evidence-generation agent that constructs complementary risk and utility evidence, and a rewrite-and-routing agent that converts this into a safe proxy prompt and image-routing decision.
The researchers identified two key obstacles in test-time safety alignment: “utility dominance” and “reasoning inertia,” which cause models to overlook latent risks or follow malicious reasoning trajectories, according to arxiv.org.
Experiments across multiple MLLMs and benchmarks demonstrated that ReFrame improves jailbreak defense, safety awareness, and reduces over-sensitive refusals while preserving multimodal utility, according to the paper. Notably, the framework operates without modifying the downstream MLLM or accessing its internal information, making it applicable to closed-source models.