Google Dream-RSI Slashes AI Agent Calls by 162x

Researchers from Google, Google DeepMind, the University of Maryland, College Park, and the University of Virginia have unveiled Dream-RSI. This new meta-exploration framework optimizes how AI discovery agents search for solutions by reusing past discovery trees as exact replay simulators. By treating historical attempts, evaluations, and workspace states as a deterministic replay environment, the system can test thousands of alternative exploration policies at zero execution cost.

The core innovation of Dream-RSI is that the underlying coding agent remains fixed. Instead of retraining the model, the framework updates a small executable policy that controls branching, batching, and stopping. During testing on Lasso path discovery, Dream-RSI matched the performance of SimpleTES while requiring roughly 162 times fewer agent calls. Furthermore, the framework outperformed both sklearn and glmnet across all six held-out datasets.

On the KernelBench benchmark, Dream-RSI demonstrated significant efficiency gains. It reached target speeds using 1.79 times to 2.43 times fewer generations. Alternatively, when operating under an equal computational budget, the framework delivered 1.44 times to 2.09 times higher performance.

For AI practitioners and developers, this framework addresses the high computational costs associated with recursive self-improvement loops. Traditionally, hard discovery tasks require thousands of cycles, wasting expensive model calls and evaluator runs on ineffective branches. By shifting policy testing into a simulated replay environment, Dream-RSI allows developers to refine search strategies without re-running costly candidate evaluations or coding-agent calls, dramatically accelerating algorithm design and GPU kernel engineering.