New AI survey charts path to human-centric foundation models with six-level framework

A team of 17 researchers from institutions including Alibaba Group, The Hong Kong Polytechnic University, Peking University, and the University of Sydney released a comprehensive survey titled "Human-Centric Intelligence in the Era of Foundation Models" on arXiv (2608.18184). The paper argues that human-centric AI — tasks like 3D human reconstruction, motion generation, and social interaction modeling — has not yet fully leveraged the scale and transferability of foundation models, despite rapid progress in the field. To address this fragmentation, the authors introduce a "full-spectrum human context taxonomy" that treats humans along three dual perspectives: as observable subjects (visual appearance and spatial geometry), as dynamic actors (kinematic dynamics and interaction modeling), and as situated agents (world simulation and embodied agency). This six-level hierarchy is designed to connect previously disjoint research efforts.

The survey then dives into methodological foundations, covering human-centric data families (e.g., images, videos, 3D scans, motion capture), computational architecture paradigms (from CNNs to vision-language models and diffusion backbones), and training/inference optimization strategies like 3D Gaussian splatting and large-scale pretraining. It systematically reviews representative methods at each of the six levels, and organizes associated datasets, benchmarks, and evaluation metrics — giving practitioners a practical reference. The authors also outline open challenges, including scalability, trustworthiness, physical grounding, and deployability of human-centric models, and they maintain an actively updated project page and GitHub repository with literature resources. The survey aims to serve as a coherent framework for both researchers and engineers looking to build foundation models that truly understand and predict human behavior at multiple scales.