OmniSapiens: A Foundation Model for Social Behavior Processing via Heterogeneity-Aware Relative Policy Optimization

Researchers have introduced OmniSapiens-7B 2.0, a new foundation model for social behavior processing trained with a novel RL method called Heterogeneity-Aware Relative Policy Optimization (HARPO). The model achieves state-of-the-art performance, with gains of up to +16.85% on multitask settings and +9.37% on held-out tasks compared to existing behavioral models. It produces more explicit reasoning traces and balances learning across diverse, heterogeneous social data like affect, cognition, and social attributes.