SafeGRPO: Self-Rewarded Multimodal Safety Alignment via Rule-Governed Policy Optimization
This addresses safety alignment for multimodal AI systems, which is crucial for real-world deployment, though it is incremental as it builds on existing methods like GRPO.
The paper tackles the problem of multimodal safety risks in large language models by proposing SafeGRPO, a framework that integrates rule-governed rewards into policy optimization, resulting in substantial improvements in safety awareness, robustness, and stability across benchmarks without compromising general capabilities.
Multimodal large language models (MLLMs) have demonstrated impressive reasoning and instruction-following capabilities, yet their expanded modality space introduces new compositional safety risks that emerge from complex text-image interactions. Such cross-modal couplings can produce unsafe semantics even when individual inputs are benign, exposing the fragile safety awareness of current MLLMs. While recent works enhance safety by guiding models to reason about potential risks, unregulated reasoning traces may compromise alignment; although Group Relative Policy Optimization (GRPO) offers self-rewarded refinement without human supervision, it lacks verifiable signals for reasoning safety. To address this, we propose SafeGRPO a self-rewarded multimodal safety alignment framework that integrates rule-governed reward construction into GRPO, enabling interpretable and verifiable optimization of reasoning safety. Built upon the constructed SafeTag-VL-3K dataset with explicit visual, textual, and combined safety tags, SafeGRPO performs step-guided safety thinking to enforce structured reasoning and behavior alignment, substantially improving multimodal safety awareness, compositional robustness, and reasoning stability across diverse benchmarks without sacrificing general capabilities.