Dream-RSI Lets Agents 'Dream' in Past Trajectories, Cutting Experiment Calls Up to 162x
teortaxesTex · x · 2026-09-16
teortaxesTex quotes the Dream-RSI paper's claim that "recursive self-improvement is becoming increasingly vital for autonomous AI agents," noting how quickly the Overton window has moved — RSI is now just another research topic like long context or multimodality.
The paper's idea: when real experiments are too expensive, let the agent "dream" in its own history first.
- After each real exploration, branches tried, failed approaches, evaluation results and compute costs are saved into a Discovery Tree.
- This history is reorganized into a Replay Simulator where the agent can cheaply test search strategies (which path first, when to parallelize, when to stop) without re-running real experiments.
- Better strategies go back into the real environment; new trajectories feed the simulator, forming a self-improvement loop.
- The underlying coding agent stays fixed; what evolves is the exploration-strategy layer.
Experiments span algorithm design, math optimization and GPU kernels: on Lasso tasks, agent calls drop up to 162x vs SimpleTES; on GPU kernel tasks, up to 2.09x performance gain at equal budget. History, saved completely enough, becomes a simulator for training the next generation of exploration strategies.
More from AGI Musings
- OpenAI's roon predicts everyone will have Astra-level capabilities within a month or two — Tolopono · 2026-09-16
- Training Data Is Regulation by Another Name: Easy to Add, Nearly Impossible to Unwind — yunta_tsai · 2026-09-16
- OpenAI's Mark Chen: There Are No Race Dynamics Off the Frontier — deanwball · 2026-09-16
- Gary Marcus on BBC: Altman, Huang and Sanders posture at extremes while honest AI safety talk is missing — GaryMarcus · 2026-09-16
- New Paper Maps Roadmap Toward AI Recursive Self-Improvement Across Domains — Yi Duan · 2026-09-16
- Frontier coding agents now write code no human can read, warns tszzl — tszzl · 2026-09-16