Researchers have developed ZenGen, an integrated framework for measuring and improving social intelligence in large language models (LLMs). The framework includes a benchmark called SoMBench that evaluates LLMs' ability to infer mental states, track social relations, and adapt behavior. The results show that current LLMs fall short of human-like performance, with the best model achieving only 72% accuracy. To address this, ZenGen also provides a training recipe for internalizing social intelligence, as well as an inference architecture called Actio that routes support from various sources to improve reasoning. --- Why it matters: This matters because LLMs are increasingly being used in human environments, and their lack of social intelligence can lead to misunderstandings or miscommunications. Improving social intelligence is crucial for developing more effective and human-like language models. Source: https://arxiv.org/abs/2607.23740