Google's Dream-RSI lets agents dream over past search trees, cutting agent calls up to 162x

量子位 · wechat · 2026-09-17

Google researchers' Dream-RSI treats an agent's past exploration logs as a simulator: an LLM devises new exploration strategies and evaluates them by replaying the discovery tree instead of re-running experiments. Across 8 tasks in algorithm engineering, math optimization and GPU kernels it matches or beats prior discovery systems with up to 162x fewer agent calls; explicitly summarizing lessons into prompts surprisingly hurts performance. Code and paper are public.

Related event: DeepMind's Dream RSI Lets Agents Self-Improve by Replaying Past Explorations(2 posts)→

Original post →

More from coding & agent

coding & agent channel →