Google DeepMind's Dream-RSI cuts AI search costs

Google DeepMind researchers have developed Dream-RSI, a system that improves how AI agents navigate complex search spaces. Instead of running expensive live trials to test new exploration strategies, the method allows agents to "dream" by replaying stored data from previous search histories. By evaluating alternative decisions within an already mapped search tree, the agent can test thousands of strategies without triggering costly model calls or external evaluators. This process refines the search strategy itself while leaving the underlying model unchanged.

The research team evaluated Dream-RSI using Gemini 3.1 Pro and Gemini 3.7 Flash across eight tasks in three distinct domains. In a statistical programming task, the system generated code that outperformed established libraries like sklearn and glmnet across all six test datasets. When paired with Gemini 3.1 Pro, the average runtime of the generated program decreased from 3,587 to 2,931 milliseconds, while the required attempts fell from 550 to just 317. In comparison, a competing system called SimpleTES required 51,200 runs to achieve its results.

The efficiency gains extended to math optimization and GPU kernel writing. On two GPU tasks, Dream-RSI matched the baseline performance while reducing the number of generation runs by a factor of up to 2.43. On two other GPU tasks, the system delivered up to 2.09 times the performance of the baseline within the same computational budget. The researchers also noted that the learned strategy dynamically adjusted its effort, reducing attempts when making progress and increasing them when progress stalled.

For AI practitioners, Dream-RSI offers a highly efficient alternative to traditional recursive self-improvement methods like AlphaEvolve, AutoTTS, or WikiSkill. While some systems rely on explicit instructions that can restrict exploration, Dream-RSI maintains open-ended flexibility. By dramatically lowering the computational barrier to strategy optimization, developers can deploy self-improving agents that find better solutions faster and at a fraction of the traditional compute cost. The researchers have made the code available on GitHub.