# Dream-RSI replays history instead of rerunning
_Dream-RSI treats discovery history as an exact replay simulator, cutting Lasso discovery calls 162x vs SimpleTES in new tests._
**Published:** 2026-09-16
**Source:** https://www.startuphub.ai/ai-news/ai-research/2026/dream-rsi-replays-history-instead-of-rerunning
---
Researchers from the University of Maryland, College Park and Google DeepMind introduced [Dream-RSI](https://arxiv.org/html/2609.14858v1) to make long-horizon AI discovery cheaper by dreaming instead of doing. The framework was posted Sept. 14 on [arxiv.org](https://arxiv.org/html/2609.14858v1) by Tong Zheng and collaborators including Zhankui He, Benjamin Coleman and Wang-Cheng Kang.

The problem is not the coding agent. It is the search around it.

Current discovery loops brute-force thousands of proposal-evaluation cycles with fixed exploration strategies, or they try to tune the exploration policy online and pay delayed, expensive feedback for each candidate. Dream-RSI adds a lightweight orchestration layer that makes branching, parallelism and stopping explicit and programmable, while leaving the underlying coding agent unchanged. Then it reuses what already happened.

The trick is treating history as a world model. After an online rollout, the team converts the recorded discovery tree into a replay simulator. Every node already stores its code, execution outcome and cost. A new exploration policy can then walk that frozen tree, pick different branches in a different order, group them differently, stop earlier, and get instant off-policy feedback without calling the agent or evaluator again. That is the dreaming stage. Thousands of candidate policies can be scored on one expensive trace before the best is redeployed to expand the simulator pool again.

It follows a familiar lineage. The authors cite the Dreamer family of [world models](https://www.startuphub.ai/ai-news/ai-research/2026/today-in-ai-world-models-have-no-recipe-yet) from Ha and Schmidhuber and Hafner et al., where agents learn a dynamics model and imagine trajectories inside it. Here there is no learned dynamics. The model is the empirical tree itself.

For teams running AI for science and systems, the budget implication is direct. In algorithm engineering on a Lasso path solver, Dream-RSI beat sklearn and cut agent calls by up to 162x versus SimpleTES and 1.7x versus fixed-exploration baselines. In mathematical optimization on sum-difference, autocorrelation and circle packing, it matched or beat baselines within 1,000 generations for over 50x savings versus SimpleTES. In GPU kernel engineering on KernelBench, it hit target speeds with 1.79x to 2.43x fewer generations, or delivered up to 2.09x faster kernels under the same budget.

Engineering leads and platform owners care because this shifts where spend happens. You still pay for the initial breadth to seed the simulator. After that, policy iteration moves offline. A kernel team optimizing for H100 or Blackwell can iterate exploration logic in simulation and reserve accelerator time for the redeployed policy, not for meta-search.

The limit is also clear. A replay simulator only covers the realized search space. If the good solution lies down an unvisited branch, no amount of dreaming will reveal it. The loop depends on online exploration to keep expanding coverage, and on an evaluator that gives reliable signal in the first place.

If that coverage grows, dreaming gets more powerful for free. That is the bet.
---
Original analysis from [startuphub.ai](https://www.startuphub.ai), the #1 AI startup directory.