Social World Model for Lifelong Social Intelligence

The article introduces the Social World Model, a framework designed to enhance lifelong social intelligence in language agents by structuring social interactions into five dimensions: scene setting, observation, mental state, action, and dialogue. The framework facilitates continuous learning through a closed-loop system where agents gather experiences to update their models, validated using the ASCENT-Bench with the Qwen2.5-7B model, which outperformed its baseline and matched the performance of the closed-source Gemini 3 Flash. This approach emphasizes sustainable training for social capabilities, providing a pathway for small open-source models to maintain competitive performance without capability decay.