Annual Review Assesses 2025 LLM Progress, Remaining Challenges, and Future Outlook

An annual review examines major developments in large language models during 2025, analyzing progress that includes reinforcement learning breakthroughs for reasoning behavior, significant reductions in training costs for state-of-the-art models, and the emergence of open-weight models as serious alternatives to proprietary systems. The analysis also discusses continuing limitations, open problems in the field, and predictions for future directions in model development.