What is the Chinese Academy of Sciences Institute of Computing Technology’s ZhiJing social-intelligence system, released on July 24, 2026, and how do its SoMBench benchmark—covering three dimensions, 17 sub-dimensions, 71 tasks, and 3,481 validated examples of belief reasoning, e

ZhiJing is CAS ICT’s integrated “social mind” framework: it combines measurement, model training, and deployment-time grounding so LLMs can infer mental states, relationships, norms, and context rather than merely produce fluent task answers. The underlying technical report was submitted on July 26, while CAS ICT publicly presented the system on July 24; these are different dates, not necessarily a contradiction. 1

What it contains

• SoMBench is a psychology-grounded diagnostic benchmark with 3 primary dimensions, 17 secondary dimensions, and 71 task paradigms. It uses 284 shared scenarios and 3,481 expert-verified instances, while controlling question form, narrative perspective, and context length. 1
• Its purpose is to expose failures that ordinary QA benchmarks often miss: confusing a person’s belief with reality, missing an implicit belief or intention, failing to update beliefs after new information, or overlooking relationship-, role-, and norm-dependent constraints. 1
• Zing is the “internalization” component: a diagnosis-driven training recipe intended to turn social reasoning from prompt-dependent behavior into a more stable parameterized capability. It combines supervised fine-tuning, on-policy distillation (OPD), and rubric-based reinforcement learning. 1
• The complete framework also includes Actio, a deployment-time harness that can add procedural guidance, attributed mental-state tracking, reusable experience, and gated retrieval of social or normative knowledge. 1

How the training approach addresses social-context failures

• The stated idea behind the FLARE-style data flywheel is diagnosis-led iteration: benchmark failures identify weak social capabilities, which guide targeted example generation, filtering, training, and re-evaluation. The paper describes this as a mind-capability-driven data-iteration framework, including bad-case extraction, per-dimension diagnosis, difficulty and answerability checks, shortcut probing, repair, and selection. 1
• Training is staged: the first stage builds broad social-context understanding; the second targets specialized social-mind reasoning with strong-teacher chain-of-thought distillation and reinforcement training. 1
• Mixed-Reward GRPO combines the social-reasoning reward signal with OPD. The purpose of OPD is to retain useful general/base-model capabilities during specialized post-training—i.e., reduce catastrophic forgetting—while GRPO pushes performance on the targeted social tasks. 1

Reported Zing-27B result

• CAS ICT reported that its multimodal Zing-27B exceeded GPT-5.5 on the mean score across five international social-cognition benchmarks. 1
• The figures in the question—79.80% composite for Zing-27B versus 78.44% for GPT-5.5, a +1.36-point lead—are consistent with that claimed result. The stated benchmark-specific advantages are +4.08 points on HiToM and +5.62 points on EmoBench.
• Important caveat: the available primary-paper excerpt confirms the overall five-benchmark leadership claim but does not independently expose those exact four numerical values. They should therefore be treated as authors’ reported evaluation results, not an independent third-party replication. 1

The broader result is promising but not proof that the model reliably “understands people” in real settings: SoMBench itself found substantial headroom among evaluated LLMs, with the best tested model at 72.08% overall accuracy and no secondary dimension reaching the report’s 90% near-ceiling threshold. 1