Dream RSI: Google Researchers Taught AI to Practice Getting Smarter

By Omar aka Crypt0 · Sep 17, 2026

Benchmarks measure how smart a model is. A new paper measures something wilder: whether an AI agent can get better at searching for answers by dreaming about its own past searches. The technique is called Dream RSI, and it is the most concrete version of recursive self improvement the field has produced in a working system.

The paper, Dream RSI: Recursive Self Improvement through Evolving Worlds, was submitted September 14 to arXiv by researchers including Tong Zheng, Xidong Wu, and Zheng Zhang, with contributions from Google and Google DeepMind alongside the University of Maryland and the University of Virginia. Its core insight is disarmingly simple: the history of an agent's exploration is itself a resource.

Here is the loop. An agent explores a hard problem while an orchestration layer records the resulting discovery tree: attempted directions, outcomes, relationships between paths. Instead of discarding that history as dead telemetry, Dream RSI converts it into a replay simulator. The agent then dreams: thousands of candidate exploration policies get tested cheaply against the historical search space, skipping the repetitive expensive real world evaluations each alternative would otherwise require. The winning policy deploys into live exploration, discovers new territory, and produces a richer tree, which becomes a richer simulator for the next cycle.

The results carry real weight. Across eight scientific discovery tasks spanning algorithm engineering, mathematical optimization, and GPU kernel engineering, Dream RSI matched or improved discovery quality while cutting discovery costs substantially. In algorithm engineering, a Lasso path solver beat standard libraries while using up to 162 times fewer agent calls than the SimpleTES baseline. In mathematical optimization, it delivered more than 50 times the budget savings against the same baseline. In GPU kernel engineering on KernelBench, it reached target speeds with roughly half the generations, or improved kernel performance by up to 2.09 times under identical budgets.

The most striking finding is behavioral. The learned policy grows adaptive: as performance improves, it conserves compute and reduces evaluated attempts, then expands the search again when progress plateaus. The system learns when to think harder and when to think cheaper. Intelligence, the paper suggests, has two dimensions: quality of cognition and allocation of cognition. Human experts distinguish themselves through the second. Now machines are learning it too.

This relocates improvement inside the operating loop. For years the industry's equation was more training compute equals a better model. Dream RSI adds new terms: model times search times memory times tools times orchestration. A system can grow substantially more capable with its weights untouched, which means capability can advance between major model releases, quietly, through software.

The singularity has always been described as the moment machines start improving themselves. Dream RSI is an early, bounded, rigorously measured taste of exactly that dynamic, aimed squarely at scientific discovery. Readers should watch recursive search the way they once watched parameter counts. It may become the metric that matters.

Sources

New to crypto? Read the crypto glossary, browse frequent questions, read our story, or explore the story archive.