Google's Dream-RSI has agents "dream" over past trajectories: 317 calls beat 51,200-generation search

新智元 · wechat · 2026-09-17

Google, Google DeepMind, and university collaborators published Dream-RSI, a recursive self-improvement framework that evolves an agent's exploration strategy without touching model weights. The agent stores every exploration in a "discovery tree," and new strategies are scored by replaying history offline — no model calls needed, like a chess player studying game records.

How it works: a two-layer setup with a Gemini 3.1 Pro/3.7 Flash discovery agent executing code, and an evolvable Python strategy layer controlling branching, parallelism, and pruning. A closed loop of real exploration → replay-simulator construction → strategy dreaming guarantees deployed strategies never regress.

Results: matched baselines requiring 51,200 generations in just 317 calls on Lasso solvers; matched AlphaEvolve V2 records on circle packing with >50x compute savings; GPU kernel optimization with 2.43x fewer generations or 2.09x better performance at equal budget.

Counterintuitive finding: injecting summarized high-level "lessons" into prompts performed worse than no guidance — abstract principles become cognitive bias that kills exploration diversity.

Related event: Google Open-Sources Dream-RSI: Recursive Self-Improvement via Dream-Style Replay(14 posts)→

Original post →

More from coding & agent

coding & agent channel →