GroupToM-Bench: Benchmarking Group Theory of Mind and Nonlinear Social Emergence in MLLMs

Researchers have introduced GroupToM-Bench, a new benchmark for assessing group-level Theory of Mind in multimodal large language models, and proposed a closed-loop dual-agent framework for creating lifelike digital humans with social intelligence.

The GroupToM-Bench benchmark, presented in a paper submitted on June 2, 2026^[1], aims to evaluate the ability of multimodal large language models to understand how individual mental states interact and lead to group-level outcomes. Existing models fail at this task, according to the researchers. The benchmark is built around a causal chain that spans micro-level BDI states, meso-level group tension, and macro-level outcome prediction. A seven-level cognitive audit framework is also developed to assess the full range of group-level Theory of Mind. In a separate but related development, researchers proposed a closed-loop dual-agent framework for creating digital humans with genuine social intelligence in a paper submitted in 2026^[2]. This framework integrates perception, social reasoning, and expression into a continuous interaction cycle, analyzing partners' multimodal behaviors and inferring hidden mental states. The expression module generates emotion-controllable dual-agent videos. The method has shown superior performance on key dialogue quality dimensions, even surpassing the full-information Script mode.