Researchers presented the Dream-RSI framework for recursive self-improvement

Researchers have published Dream-RSI — a new framework for recursive self-improvement of AI agents in dynamic environments. The system uses accumulated discovery history as a replay simulator, allowing models to evaluate and adjust their strategies without the need for costly online testing. The main goal of Dream-RSI is to minimize computational costs while iteratively improving agents. Instead of directly interacting with the external environment at each stage, the algorithm creates an internal model of the world based on past data. This allows the system to conduct thousands of simulations autonomously while maintaining high decision-making accuracy. During testing, the framework demonstrated the ability to autonomously accumulate experience, where the agent itself identifies gaps in its knowledge and fills them through evolving worlds. The method avoids strategy degradation characteristic of standard reinforcement learning approaches by constantly validating through historical data. The technology is aimed at solving problems under limited resources, where traditional training requires significant power. Developers note that using historical replay reduces the load on infrastructure when scaling agents to complex multi-level tasks. Detailed experiment results and mathematical algorithm description are available in the scientific publication.