A paper signed by Google, Google DeepMind, the University of Maryland and the University of Virginia proposes a different route to recursive self-improvement (RSI): let an agent dream inside its own past exploration history, iterate a better exploration policy in that dream, then redeploy it into the real world. Titled "Dream-RSI: Recursive Self-Improvement through Evolving Worlds," it was posted to arXiv on September 14 (2609.14858) and has not been peer reviewed.
RSI is the feedback loop in which AI systems help build better AI, and it is precisely the capability that Anthropic CEO Dario Amodei named when he called last week for the industry to pace frontier progress. Dream-RSI's path leaves model weights untouched. A lightweight orchestration layer makes the exploration policy explicitly programmable code, while the underlying coding agent, the evaluator, and the execution interfaces stay fixed.
The paper's key insight is that accumulated discovery history can serve as a replay simulator. Exploration is recorded as a discovery tree: each node is one generation-evaluation attempt that inherits its parent's context. Offline, the agent "dreams" inside the world built from that history, using recorded outcomes to evaluate candidate policies at negligible cost, and the improved policy is deployed online to drive further discovery, continuously expanding the simulator pool. The paper's closing line: "An agent must dream to recursively self-improve. History is the world it dreams in."
Results, all self-reported, cover three domains. In algorithm engineering, Dream-RSI with Gemini 3.1 Pro optimized a Lasso regularization path solver in 317 calls, cutting mean runtime from 3587.1 ms to 2931.0 ms; the paper contrasts this with roughly 550 calls for a fixed exploration strategy and 51,200 generations for the baseline evolutionary search system SimpleTES. With the cheaper Gemini 3.7 Flash, 1,879 calls reached 2350.6 ms. In mathematical optimization, across three tasks — the sum-difference problem, circle packing, and an autocorrelation inequality — Dream-RSI matched or beat strong baselines within 1,000 generations, and it matched Google's own AlphaEvolve V2 record on circle packing. In GPU kernel engineering (VGG16 and others), it reached equivalent performance with roughly 1/2.43 of the generations of fixed exploration.
The paper also reports a counterintuitive finding: on the math tasks, human-provided directional hints made agents perform worse than no hints at all, which the authors attribute to the enormous meta-search space of exploration policies, where human intuition does not reliably cover the branches worth improving.
Caveats, stated plainly: this is an arXiv preprint, all numbers are author-reported, and none of it has been independently reproduced. Because only the exploration policy changes and the underlying model does not, the gains decouple from model scale; smaller models benefit too. The architecture also suggests a boundary on what this kind of RSI can do: it improves the search for solutions inside a fixed interface, but it does not yet redesign the models, the evaluator, or the task definitions themselves, so the loop is bounded by what the discovery agent can express and what the evaluator can score. That is a narrower claim than improving AI in the broad sense, and the paper is honest about that limit. The timing matters as much as the method: at the very moment Silicon Valley is debating whether to slow frontier capability gains, this paper turns recursive self-improvement from a thing to be feared into a reproducible engineering result.
[1][2]