AI最尴尬的短板，中国科学院出手了

TL;DR - The Chinese Academy of Sciences’ Zhijing team unveiled an engineering stack for social intelligence in AI: the SoMBench benchmark, Zing models and training system, and Actio deployment architecture. It aims to make social reasoning measurable, trainable, and reliable in real-world interactions.

• SoMBench covers 71 tasks across belief inference, emotion understanding, social norms, and relationship modeling; the best of 20 tested models scored 72.08%.
• Zing uses weakness-targeted synthetic data, two-stage fine-tuning and reinforcement learning, plus on-policy distillation to reduce forgetting.
• Reported five-benchmark averages were 79.80 for multimodal Zing-27B and 76.32 for text-only Zing-32B, narrowly exceeding the cited GPT-5.5 and DeepSeek-V4-Pro results.
• Actio dynamically routes social skills, explicit belief-state memory, reusable reasoning experience, and culturally grounded retrieval during inference.