Google and DeepMind Accelerate Search: Dream-RSI Halves Attempts

Researchers from Google and DeepMind presented the innovative method Dream-RSI on September 19, which allows AI agents to significantly speed up their search processes by reusing previous attempts. Instead of making costly repetitions, this approach leverages the results of prior search actions to test new, more efficient strategies. This is particularly relevant for complex tasks such as mathematics and programming, where computational costs can quickly escalate and thousands of attempts are required.

The self-improving AI agents operate within a cycle of proposing, assessing, learning, and retrying. During the experimental phase, where the search space can grow exponentially, the agent must make strategic choices about which paths will yield the best results and which can be discarded. Dream-RSI offers a solution to this challenge by storing data from previous attempts and generating 'dreams', significantly reducing the number of required computations for future search actions. The techniques for playing alternative decisions within existing search structures enable the agent to select the most efficient strategies.

The effectiveness of Dream-RSI was tested with the models Gemini 3.1 Pro and Gemini 3.7 Flash, achieving substantially better results in programming and optimization tasks than traditional approaches. For instance, it was found that the agent using Dream-RSI was able to generate the fastest program in a programming task at lower costs and with fewer attempts compared to competitors like SimpleTES. This illustrates the potential of strategic optimization through the reuse of search data, which reduces computing time and costs without sacrificing performance.

Follow-up research has also shown that overly strict instructions can limit the search space and miss potentially useful paths. Dream-RSI demonstrates that optimizing search strategies through revision within stored search data can improve exploration in open search problems. The available code and additional details can be found on GitHub, making it possible to reproduce and apply the breakthrough methodology to other search problems.

Read the full article from AI Insider.