📚 Dream-RSI: Recursive Self-Improvement through Evolving Worlds

Dream-RSI introduces a scalable framework for recursive self-improvement in autonomous AI exploration by treating completed discovery histories as reusable replay simulators. This approach replaces expensive online trial-and-error meta-search with rapid offline dreaming over historical discovery trees. The authors demonstrate that fixed exploration strategies fail to adapt as search spaces scale, while online policy optimization incurs prohibitive delays and computational costs. By making exploration explicit and programmable through a lightweight orchestration layer, Dream-RSI decouples meta-policy evaluation from the underlying coding agent. Through extensive empirical evaluation spanning algorithm engineering, mathematical optimization, and GPU kernel engineering, Dream-RSI achieves competitive or superior discovery quality while reducing discovery agent calls by multiple orders of magnitude compared to existing baselines. Practitioners can adopt this framework to optimize complex, long-horizon AI agent workflows without retraining underlying models, significantly cutting computation budgets and accelerating automated scientific discoveries.