Show HN: A Lego-like composable Framework for RSI (rsi-gear.github.io)

The recent launch of GEAR, an open-source framework for Recursive Self-Improvement (RSI) of Large Language Model (LLM) agents, marks a significant advancement in the AI/ML community. GEAR implements a two-level meta-learning process, where a rollout agent refines its skills through task feedback in an inner loop, while a Meta Agent learns from multiple task experiences in an outer loop. This structured approach enables the identification of recurring patterns, fostering reusable improvements that enhance overall task performance across different applications. Technically, GEAR's evolutionary search is modular, allowing researchers to test various algorithm components for optimal performance. One of the instantiations, Staged GEPA, utilizes a frozen archive to select parents for improvement, emphasizing flexibility in the experimentation process. Early benchmarks show GEAR's candidates outperforming existing models like GPT 5.6 and GPT 6 across several tasks, indicating its potential for greater efficiency in coding, marketing, and support automation. This innovative framework opens new pathways for research and development in AI, inviting further exploration and experimentation in the field.