Self-Improvement of Large Language Models: A Comprehensive Technical Overview and Future Outlook
arXiv·medium signal
Yang, Xerri, Park et al. survey the emerging field of LLM self-improvement — techniques that reduce dependence on human supervision as models approach human-level capabilities. The paper covers self-play, constitutional AI refinement, automated reward modeling, and self-distillation methods. Key insight: as LLMs exceed human capability in specific domains, human feedback becomes a bottleneck rather than a guide. This survey maps the landscape for practitioners deciding which self-improvement techniques to apply.