GEPA

Researchers from Imperial College London, Duke and other institutions have released GEAR, an open-source modular toolkit for recursive self-improvement (RSI) of LLM agents, styled after Scikit-learn. It decomposes parent selection, mutation planning, task sampling, scoring and candidate ranking into swappable modules while sharing execution, evaluation and versioning infrastructure. Experiments built on the GEPA algorithm show notable gains on marketing, customer service and coding tasks with frozen model weights.