zhengkid/Dream-RSI

What it solves

Dream-RSI addresses the inefficiency of recursive self-improvement in AI agents. As agents attempt to solve increasingly complex tasks, they often waste significant computing power on ineffective search directions. Traditional systems struggle because they either use fixed exploration strategies that cannot adapt to larger search spaces, or they attempt to optimize their exploration policy online, which is slow and expensive due to delayed feedback.

How it works

Dream-RSI treats the history of an agent's past discovery attempts as a "replay simulator." Because every past exploration decision and its outcome is already recorded, the system can "dream" by re-traversing these past records in different ways—testing new branching, parallel groupings, or stopping decisions—without needing to rerun the original, costly discovery process. This creates a loop: the agent performs online discovery, logs the results, uses those logs to refine its exploration strategy through "dreaming," and then deploys the improved strategy back into the real world.

Who it’s for

This project is for researchers and developers working on AI-driven automated discovery, such as algorithm engineering, mathematical optimization, and GPU kernel optimization, who need to improve the efficiency and effectiveness of their agents' search processes.

Highlights

• History as replay simulator: Uses past discovery logs to provide immediate, low-cost feedback for evaluating new exploration policies.
• Meta-layer RSI loop: Continuously upgrades the agent's exploration strategy by learning from its own history.
• Empirical validation: Demonstrated improvements in efficiency and effectiveness across algorithm engineering, math optimization, and GPU kernel engineering.