benchmark-disjoint

ModularRSI introduces a modular and generalizable framework for recursive self-improvement in AI agent harnesses, using contrastive learning across tasks to evolve modules independently and enhance performance on unseen tasks.