Core ML & Applied AI Advances

New methods: SAF-OPD for stable on-policy distillation, CAPA for personalized coding assistant adaptation, Mental World Modeling, CriPO, QQWorld, RLSVR. Large Korean MoE models (A.X K2, K-EXAONE 2.0) released. Practical benchmarks (ExtractBench) and hardware-software co-design (Persistent State Machines). Classic papers revisited: Lottery Ticket Hypothesis, ColBERT. New applied: adaptive decision support RL to reduce overreliance (Harvard), Meshy T2 fast 3D mesh generation.