Next-Generation AI Agents Evolving with 2D-3D Spatial Learning, Recursive Code (ReCode), Latent Space Reasoning
Exploring Ambiguous Query Interaction, Active Robot Control, Infinite 3D World Generation, and Data Agent Autonomy

MetaX Weekly AI Paper Review -- Week 44 of 2025. Key paper: "Concerto: Joint 2D-3D Self-Supervised Learning Emerges Spatial Representations" -- inspired by how humans learn spatial concepts through multiple senses, proposes a new self-supervised learning method utilizing both 2D images and 3D point cloud data. The model combines learning within 3D data itself with learning relationships between 2D and 3D data -- achieving SOTA on major benchmarks like ScanNet for 3D scene recognition tasks, outperforming models using either 2D or 3D data alone. Can be extended to open-world recognition in conjunction with video or language (CLIP). Additional papers covered: recursive code generation architectures (ReCode), latent space reasoning for improved inference efficiency, active robot control with uncertainty-aware exploration, infinite 3D world generation from sparse observations, and data agent frameworks for autonomous data science workflows.