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.
More from coding & agent
- Check your fallback config: DeepSeek in Claude Code quietly burns tokens on V4-pro — AlchainHust · 2026-09-17
- funes: open-source local memory for coding agents, indexing Claude Code and Codex traces on Lance — ariG23498 · 2026-09-17
- Tested Cursor, Claude Code, Codex and Antigravity on the Same App Build, Bugs Included — workflowverdict · 2026-09-17
- Dev builds on-device EPUB-to-audiobook converter for Apple Silicon in one night — ojasvi_yadav · 2026-09-17
- NewMax browser extension v1.2.2 adds Browser Use and local model support — yangyi · 2026-09-17
- Google details how 3 companies build with agentic video understanding — MarioLucic_ · 2026-09-17