MAVEN: A Macro Social Value Evaluation Framework and Compact Alignment Evaluator for Multimodal Content

The researchers propose the MAVEN framework, which, based on international human rights instruments and cultural value theory, conducts multi-level quantitative evaluation of whether multimodal content aligns with macro social values such as peace, justice, and freedom, organizing values into 6 first-level dimensions and 72 second-level indicators. Based on this, the team constructs the manually verified multimodal benchmark MacroValue-Bench and a soft matching metric, and proposes the span-adaptive multi-level preference optimization method SA-MDPO for evaluator distillation, as well as a training-free multi-role consensus strategy for the inference stage. Evaluation of open and closed-source VLMs shows that models share common tendencies as well as clear differences in social value judgments; a 2B compact evaluator from the same family can match an 8B evaluator and approach closed-source frontier VLMs. The SA-MDPO implementation and MacroValue-Bench have been released at https://github.com/zzzzzzzzjj/MAVEN。