Google Deepmind's Dream-RSI helps AI agents improve by “dreaming” about past attempts
decoder ·
Researchers at Google and DeepMind developed Dream-RSI, a method that helps AI agents choose better search strategies for difficult optimization tasks without changing the underlying model. It records the agents’ attempts and results in a search tree, then replays that history to test alternative decisions without generating or evaluating solutions again. This “dreaming” process alternates with live searches, allowing thousands of strategies to be compared cheaply before the most promising one is used. Tested with Gemini 3.1 Pro and Gemini 3.7 Flash across eight tasks in three areas, Dream-RSI produced faster or better results with fewer attempts. On a program-optimization task, Gemini 3.1 Pro reduced average runtime from 3,587 to 2,931 milliseconds and cut attempts from 550 to 317, while outperforming SimpleTES, which required 51,200 runs. The method also matched GPU-kernel performance with up to 2.43 times fewer runs or achieved up to 2.09 times higher performance within the same budget. A follow-up analysis found that turning search histories into explicit instructions could sometimes restrict exploration, and the researchers have shared code and further details on GitHub.
decoder