Core ML and Applied AI Advances: New Architectures, Training Methods, and Benchmarks

Steady stream of new methods: TurboVLA (0.2B param VLA at 32Hz), Weightless neural networks (1000x smaller), Memory Decoder at Scale, PhiZero, and TabPFN quantization. Stanford paper challenges offline-to-online RL assumptions. New benchmarks: CLBench-V (multimodal context learning, best 0.2847), Data-Dependent Regret for OCO. Applied AI sees NVIDIA Cosmos 3, Qwen-UI-Agent, VideoCoCo, and financial AI papers emphasizing decision quality over prediction accuracy. A framework for hardware-aware autonomous intelligence links hardware degradation to AI reasoning.