Pragma-VL: Towards a Pragmatic Arbitration of Safety and Helpfulness in MLLMs

Researchers introduce Pragma-VL, a new alignment algorithm for Multimodal Large Language Models that balances safety and helpfulness by improving visual risk perception and using contextual arbitration. The method outperforms existing baselines by 5-20% on multimodal safety benchmarks while maintaining general AI capabilities in mathematics and reasoning.

• →Pragma-VL addresses the safety-utility trade-off in MLLMs where current methods either refuse benign queries or miss cross-modal risks.
• →The algorithm uses risk-aware clustering on visual encoders and interleaved datasets to enhance visual risk perception.
• →A theoretically-guaranteed reward model with dynamic weighting enables contextual arbitration between safety and helpfulness.
• →Performance improvements of 5-20% on multimodal safety benchmarks demonstrate significant advancement in MLLM safety.
• →The method preserves general AI capabilities in areas like mathematics and knowledge reasoning while improving safety.