The State Of LLMs 2025: Progress, Problems, and Predictions

A comprehensive 2025 LLM year-in-review by Sebastian Raschka covering the dominant trend of RLVR+GRPO reasoning models (sparked by DeepSeek R1), architectural convergence on MoE + efficient attention, and open problems like continual learning and catastrophic forgetting. The piece traces the yearly evolution of post-training techniques (RLHF→LoRA→mid-training→RLVR) and makes predictions for 2026–2027 including expanded RLVR domains and inference-time scaling. Dense technical substance with practitioner-relevant takeaways on GRPO improvements that materially impact training stability.