Researchers at Google and Google DeepMind posted Dream-RSI on Sept. 14, 2026. It is built for AI agents, AI that carries out tasks on its own, here ones that search for better solutions to coding and math problems. Dream-RSI repeatedly improves where those agents look, while the AI models themselves stay exactly as they were. Talk of recursive self-improvement, the RSI in the name, usually means AI models rewriting themselves; here only the search rules change, as ordinary code edits people can inspect, and no new training run is needed.
What changes is the agent's search strategy: the rules that decide which path a coding agent tries next, how many paths it runs at once, and when it stops. After each live run, the record of every path the agent tried becomes an exact replay of the ground it has already covered. The system then tests other strategies against that replay, which the authors call dreaming, without running the agent or the scoring program again. The best strategy goes into the next live round.
Tong Zheng and coauthors at Google, Google DeepMind, the University of Maryland and the University of Virginia built a thin control layer around the agent that lets those search choices be set in code. The agent, the AI models, the scoring program and the way code is run all stay fixed. The project site stresses that the replay is exact, not an estimate from a world model, an AI that learns how the world behaves so it can predict what happens next. The cycle repeats in three steps: explore live, build the replay, then use dreaming to find a better strategy and put it to work.
The authors report large savings. On one task, Dream-RSI needed up to 162 times fewer agent runs than an earlier method called SimpleTES, and about 1.7 times fewer than the same agent with a fixed search strategy, a closer like-for-like comparison. On math tasks, they report savings of more than 50 times against SimpleTES within about 1,000 attempts. On a set of tasks for writing fast low-level code, they report needing fewer attempts, or producing faster code on the same budget.
The method does not win everywhere: on one task in the paper, SimpleTES still comes out ahead on some measures. The team's GitHub page says the full code is still being prepared for release, and for now the project website offers an interactive walkthrough.
The replay also has a built-in limit. It covers only paths the agent has already tried, so dreaming cannot find a better answer down a path nobody explored. It is not yet known how many rounds of live runs and dreaming it takes before the gains level off, and the savings rest on the authors' own tests, which no one outside the team has repeated.