Zhijing argues social intelligence is AGI’s missing third pillar, with a benchmark and model stack
机器之心 · wechat · 2026-07-29
The paper frames social intelligence as a “third pole” of AGI
The article argues that beyond symbolic intelligence and physical intelligence, AGI still lacks a third capability: social intelligence — understanding beliefs, intentions, emotions, norms, and recursive mental states like “I know that you know that I know.”
It introduces the Zhijing social-mind model stack:
- SoMBench: a benchmark spanning 17 sub-dimensions and 71 tasks, with 3,481 expert-reviewed examples, designed to diagnose where models fail.
- Zing: a model family trained with a self-evolving pipeline that generates new data for weak cognitive dimensions, using mixed-reward GRPO and online distillation.
- Actio: a deployment framework that combines skills, memory, experience, and social knowledge at runtime.
Reported results include:
- Zing-27B multimodal reaches 79.80 average on five social-mind benchmarks, beating GPT-5.5’s 78.44.
- Zing-32B text scores 76.14, roughly on par with DeepSeek-V4-Pro’s 75.90.
- Zing-8B/14B outperform much larger baselines on higher-order belief reasoning, especially on HiToM.
The authors’ broader claim is that social intelligence can be treated as an engineering target — measurable, trainable, and deployable — rather than a vague notion of “EQ.”
More from AGI Musings
- Zuckerberg Advocates AI Acceleration, Warns Regulation Entrenches Incumbents — ns123abc · 2026-07-29
- Anti-AI-pause post says slowing frontier progress is “anti-human” — basedjensen · 2026-07-29
- Open AI models and basement biology raise a bigger question than any single bio-threat — teortaxesTex · 2026-07-29
- Kevin Bryan Says Minimal RSI Could Arrive Around 2027 as AI R&D Productivity Rises 9% — imjustnewatai · 2026-07-29
- Andrew Chen says AI apps now follow a new smile curve as models improve — andrewchen · 2026-07-29
- Durable context across models is emerging as a path to corporate and personal sovereignty — prestonpesek · 2026-07-29