Google researchers have built a system that lets AI discovery agents become dramatically more capable without doing more underlying work. Dream-RSI, a framework developed by teams at Google DeepMind, the University of Maryland, and the University of Virginia, cuts the number of required discovery-agent calls from 51,200 down to just 317 on a benchmark optimization task — a roughly 162x reduction that achieves results using a small fraction of the calls the baseline demanded.

The framework is described in a paper published on arXiv under the identifier arXiv:2609.14858. At its core, Dream-RSI teaches AI agents to replay and learn from their own prior searches rather than running new ones from scratch every time.

How Dream-RSI Works

The framework operates on a three-stage loop, with each stage building on the last.

In the first stage, the system runs an initial phase of online exploration, generating what the researchers call “discovery trees.” These are structured records of every search path the agent took, including the dead ends as well as the breakthroughs.

In the second stage, those historical trails are transformed into replay simulators. Instead of calling the underlying coding agent again, the system reconstructs the decision landscape from data it already has.

The third stage is where the “dreaming” happens. A separate large language model evaluates and refines the exploration policies — covering branching strategies, batching rules, and when to stop searching — entirely offline, with no new expensive agent calls. The orchestration layer tests different approaches against the replay simulator, determines what works best, and locks in improved policies for future runs.

The result is recursive self-improvement without updating the core model's weights. The agent does not need to be retrained.

The Numbers

On a Lasso path solver task — Lasso being a standard regularized-regression technique — the SimpleTES baseline required 51,200 discovery-agent calls, each one an invocation of the underlying coding agent. Dream-RSI, running on Gemini-3.1-Pro, brought that figure down to 317. Average runtime dropped from 3,587.1 milliseconds to 2,931.0 milliseconds.

The solutions Dream-RSI produced outperformed established libraries such as sklearn, a standard Python machine-learning library, and glmnet, a widely used regularized-regression package, across six held-out datasets. Even compared with a recursive fixed exploration approach that used 550 calls, Dream-RSI's 317 calls represented a meaningful improvement.

On KernelBench, a benchmark for GPU-kernel generation, Dream-RSI achieved performance parity with existing methods while generating 2.43x fewer generations on VGG16, a well-known image-classification network, and 1.79x fewer on LayerNorm, a standard neural-network layer.

Why It Matters

Gemini-3.1-Pro and Gemini-3.7-Flash were used in the experiments without any fine-tuning or weight updates. Dream-RSI sits on top of existing models as an orchestration layer, meaning the gains come from how agents are orchestrated rather than from changes to the models themselves.

The research team, led by Tong Zheng and including contributors such as Xidong Wu and Zheng Zhang from both Google DeepMind and the partnering universities, has set up a project site and a GitHub repository. The code is reportedly being prepared for external release — a step that would let independent developers inspect the implementation and attempt to reproduce the reported reductions in agent calls.