Self-Improving AI Models and Architectures

Multiple breakthroughs in self-improving AI: Ornith-1.5 (self-improvement training, SOTA on reasoning/coding), Meta^n (recursive self-improvement via emergent depth, only method above zero on ARC-AGI-2), and CAFE (co-evolving feedback for search agents). These signal a shift toward models that improve autonomously, with practical implications for agentic systems and training loops.