THE CRUNCH
Google Deepmind researchers have developed Dream-RSI, a method that lets AI agents improve their search strategies by replaying past attempts rather than running new computations. The system records the decisions made during a search and then tests different strategies against that recorded history, effectively 'dreaming' about alternative paths. This approach allows the agent to find better solutions with fewer live
The researchers tested Dream-RSI with Gemini 3.1 Pro and Gemini 3.7 Flash on tasks including writing efficient programs, math optimisation, and GPU kernel generation. The system outperformed established libraries and a competing system called SimpleTES, which required significantly more attempts. On a statistical calculation task, Dream-RSI cut the number of attempts from 550 to 317 while improving runtime, and on GPU tasks it matched performance with up to 2.43 times fewer generations or achieved up to 2.09 times higher performance within the same budget.
The method works by changing the search strategy while leaving the underlying AI model untouched. After each live search, the agent replays thousands of variations against stored results to select the best strategy before the next live run. This alternation between live searches and replays allows the agent to refine its approach without the high computational cost of generating and evaluating new solutions from scratch.
Why It Matters: Dream-RSI addresses the computational cost of self-improving AI agents. By reusing data from past searches, the method reduces the number of expensive live runs needed to find optimal solutions, making it easier for AI agents to tackle complex problems like code generation and mathematical optimisation.