Google AI Updates: September 20, 2026

1. Dream-RSI Replays Recorded Searches to Tune Strategy Without Re-Calling the Model

Google and DeepMind published Dream-RSI, a method that improves how an agent searches for a solution rather than improving the model doing the searching. The system logs every search attempt and its result, then replays that history offline to evaluate alternative strategies, which the researchers say lets thousands of candidate strategies be tested without calling the model or the evaluator again. Execution alternates between live exploration and strategy refinement derived from the recorded runs, so the expensive part of the loop, actually generating and scoring candidates, stops being repeated for every strategy variation you want to try.

On code optimization tasks with Gemini 3.1 Pro, Dream-RSI cut attempts from 550 to 317 and runtime from 3,587 to 2,931 milliseconds, and reached its result in 317 runs where the competing SimpleTES needed 51,200. On GPU kernel tasks it matched baseline performance with up to 2.43x fewer runs, or delivered 2.09x better performance inside the same compute budget. The team published code and further detail in the Dream-RSI GitHub repository. Source