Researchers at Google and Deepmind have introduced Dream-RSI, a method designed to optimize how AI agents navigate complex search tasks. Instead of modifying the underlying AI model, the system reuses data from completed search runs to test alternative strategies. This process, which the team calls dreaming, allows agents to evaluate thousands of variations without repeating expensive computations.
During the testing phase, researchers utilized Gemini 3.1 Pro and Gemini 3.7 Flash across eight tasks in three distinct areas. In a statistical calculation task, the agent using Dream-RSI reduced the average runtime from 3,587 to 2,931 milliseconds, while the number of attempts decreased from 550 to 317. The system also outperformed the SimpleTES approach, which required 51,200 runs.
The method demonstrated efficiency in math optimization and GPU kernel development. On two GPU-related tasks, Dream-RSI matched previous performance levels while reducing the number of runs by a factor of up to 2.43. For other tasks, it achieved up to 2.09 times the performance within the same computational budget.
A follow-up analysis indicated that condensing search histories into explicit instructions can limit exploration, as overly specific directions may narrow the search space. The researchers have made the code and further technical details available on GitHub.
