Core ML, Efficient Hardware, Scientific ML, and Evaluation Integrity

Hierarchical continuous diffusion language models, curriculum alignment with AI feedback, density-aware multi-reward RL, adaptive reward routing, and synthetic multimodal data add promising but unevenly evidenced directions. Reposted leads on looped-MoE scaling laws, tokenization, and SETA's reproducible RL resources remain unvalidated; KaliBench and ROWBench continue to show the value of executable ground truth, replayable records, and independently validated evaluators.