Google Unveils Dream-RSI: AI Improves Search Strategies by 'Dreaming' Without Changing Model Parameters

Google, together with Google DeepMind, the University of Maryland, and the University of Virginia, introduced Dream-RSI, a framework that turns past search records into a replay simulator to test and rewrite exploration strategies without retraining models. Experiments show up to 162x fewer calls on Lasso algorithm discovery and 2.43x fewer generations on GPU kernel optimization, or 2.09x better kernel performance under similar budgets. The work is recursive self-improvement at the agent harness layer and does not modify Gemini's model weights.

+5