- Agents record their search attempts and results during live runs.
- Dream‑RSI replays alternative strategies against stored search trees without calling the base model again.
- The approach tests thousands of variations cheaply before committing to a live strategy.
- With Gemini 3.1 Pro, average runtime fell from 3,587 ms to 2,931 ms and attempts dropped from 550 to 317 on reported tasks.
- Dream‑RSI outperformed SimpleTES, which required tens of thousands of runs versus Dream‑RSI's hundreds.
- Gains observed across domains: optimization, GPU kernel writing, and code tasks.
- Operates one level above solution generation by optimizing the search strategy rather than model outputs.
- Authors note explicit instruction approaches can restrict exploration on open-ended tasks.
- Method could reduce compute costs for recursive self‑improvement or agentic search workflows.
Useful?