Chinese Academy of Sciences Fixes AI's Most Awkward Weakness: Social Intelligence
Институт вычислительных технологий Китайской академии наук (Institute of Computing Technology, Chine
The Chinese Academy of Sciences' 'ZhiJing' team has released a complete technical system for social mind in large models, including the SoMBench benchmark, the Zing training framework, and the Actio deployment architecture. SoMBench evaluates social intelligence across 71 fine-grained tasks, while Zing uses adaptive training to improve performance, and Actio enables explicit mental state deployment. In tests, Zing-27B outperformed GPT-5.5 in social mind benchmarks.
On July 24, 2026, the 'ZhiJing' team from the Key Laboratory of Intelligent Algorithm Security at the Institute of Computing Technology, Chinese Academy of Sciences, officially released the ZhiJing Social Mind large model technical system. This includes SoMBench, a benchmark with 3481 instances across 3 primary dimensions, 17 secondary dimensions, and 71 fine-grained tasks, designed to pinpoint social intelligence deficiencies. Zing is a training system featuring the FLARE data flywheel, two-stage training (general then specialized), and OPD online distillation to adaptively improve social reasoning. Actio is a deployment architecture with explicit mental state management using components like PRISM, Starling Memory, SAGE, and Gated RAG. On international benchmarks, Zing-27B achieved an average score of 79.80, surpassing GPT-5.5's 78.44, while Zing-32B scored 76.32, slightly above DeepSeek-V4-Pro's 75.90. Smaller models like Zing-8B and Zing-14B also showed significant gains, especially in higher-order belief reasoning. The system aims to make social intelligence a measurable and trainable engineering capability.
- Сокращения
- SoMBench = Social Mind Benchmark — бенчмарк социального интеллекта
- FLARE —
- GRPO = Group Relative Policy Optimization — групповая относительная оптимизация политики
- OPD = On-Policy Distillation — дистилляция на политике
- PRISM —
- SAGE —
- RAG = Retrieval-Augmented Generation — генерация с дополнением извлечением
- HiToM = Hierarchical Theory of Mind — иерархическая теория разума
- EmoBench = Emotion Benchmark — бенчмарк эмоций
Source: QbitAI 量子位 —
original