Chinese Academy of Sciences Releases Social Mind AI Model to Address AI's Emotional Intelligence Gap

Current AI excels in language and tool intelligence, but often appears awkward and inappropriate in social scenarios such as family dinners or team meetings. The Knowledge Environment team points out that this is not a lack of knowledge, but a lag in the ability to 'read people's minds.' The long-term absence of social mind is a lesson that must be learned to move from 'task execution' to 'social collaboration.'

To this end, the Knowledge Environment team constructed the SoMBench benchmark, decomposing social mind into 3 first-level dimensions, 17 second-level dimensions, and 71 fine-grained tasks, covering a complete spectrum from basic belief inference to high-level emotion understanding. The benchmark includes 3,481 instances reviewed by over ten human experts, using multi-question-type cross-validation and shortcut detection to precisely locate model error patterns.

In tests on 20 top large models, the strongest model achieved only 72.08% accuracy, with none reaching the 90% ceiling, indicating that social mind remains a 'no-man's land' for AI. The difficulty lies in the need to simultaneously handle multiple layers of reasoning such as role relationships, implicit intentions, and emotional changes in the same scenario, which ordinary large models struggle to disentangle.

The Knowledge Environment team designed a training system that can 'self-evolve,' with core components including the FLARE data flywheel, two-stage training, and OPD online distillation. FLARE identifies cognitive weaknesses through evaluation, synthesizes new training samples targeted at those weaknesses, and retains only medium-to-high difficulty samples that the model cannot easily answer but can learn through reasoning, achieving 'practice where weak, practice new problems each time.'

Two-stage training first builds a general social mind foundation through multi-teacher distillation and Mixed-Reward GRPO reinforcement learning, then conducts specialized reinforcement for weak areas such as emotion understanding and intention inference. OPD introduces teacher model token-level distribution constraints on online trajectories to prevent forgetting what was learned during reinforcement learning; GRPO handles exploration, while OPD guards the boundaries.

Zing-27B achieves an average of 79.80 across 5 international benchmarks, surpassing GPT-5.5's 78.44, with clear advantages in HiToM (+4.08pp) and EmoBench (+5.62pp). Zing-32B averages 76.32, slightly exceeding DeepSeek-V4-Pro's 75.90, with a 6.25 percentage point lead in EmoBench. Zing-8B improves by 21.67 and 22.08 percentage points on HiToM third-order and fourth-order belief reasoning respectively, surpassing larger parameter models.

Actio is an explicit mental state deployment architecture that 'calls on demand, not stuffs randomly.' It uses Harness as a unified orchestration core, selectively activating four types of support based on the task's mental requirements: PRISM (76 hierarchical psychological and social skills), Starling Memory (explicitly records mental states), SAGE (cross-task reusable reasoning experience), and Gated RAG (social and cultural knowledge retrieval). A single reasoning process goes through four steps: understand the task and identify mental needs → parallel retrieval → sort, trim, combine, and guide reasoning → offline precipitation of effective experience.

Application prospects include embodied intelligence (home care, child companionship, etc., requiring real-time social reasoning), complex decision-making simulation (policy communication, crisis management, etc.), and human-machine symbiosis (unified mental protocol for multi-agent systems). Small parameter models (8B/14B) enable edge deployment.

The Knowledge Environment team emphasizes that social intelligence is becoming the next core track for general artificial intelligence, and their technical route transforms social intelligence from vague 'emotional intelligence' into an engineering object with evaluation tools, training methods, and runtime interfaces. However, whether training gains can transfer across tasks and cultures, and whether Actio can remain stable in long-term interactions, still require broader validation.