New architectures, training tricks, and safety methods for stronger ML systems

This cluster centers on foundational AI/ML advances: novel architectures (diffusion transformers for social gestures, thalamus-inspired continual learning, tiny arithmetic-capable transformers), refined training and optimization strategies (diagnostic-driven iterative training, hybrid parallelism for faster diffusion, insights from large VAE experiments), and probabilistic perspectives on generative models. Several works tackle richer perception and reasoning, from 3D geometry and open-vocabulary segmentation to safer vision-language models and risk-aware world models for autonomous driving. Others ground generative AI in the physical and built world, such as converting construction drawings into 3D digital twins and co-designing real-world objects with physics, underscoring a shift from mere generation to controllable, reliable, and deployable intelligence.