Dream-RSI Paper Reports 162× Fewer Agent Calls vs SimpleTES
TL;DR
- Dream-RSI reports up to 162× fewer agent calls than SimpleTES on a Lasso task, using 317 versus 51,200 discovery calls.
- The paper claims over 50× budget savings on mathematical optimization within a thousand generations, matching or exceeding SimpleTES on multiple problems.
- GPU-kernel results on KernelBench are tighter: 2.43× fewer generations for VGG16 and 1.79× for LayerNorm at comparable performance.
SimpleTES needs 51,200 agent calls to design a Lasso solver. Dream-RSI needs 317.
That comes from a preprint posted to arXiv on 14 Sep 2026 by a team spanning the University of Maryland, Google Deepmind and the University of Virginia. The system, they write, treats 'accumulated discovery history' as a replay simulator: the tree of past attempts, with all their scores and artifacts, becomes a free environment for testing alternative exploration policies before spending money to run them online.
The Lasso number is the headline. Against SimpleTES, which runs on GPT-OSS-120B and pays for every rollout, Dream-RSI on Gemini-3.1-Pro reports 'Up to 162× vs. SimpleTES; 1.7× vs. fixed-exploration baselines.' Gemini-3.7-Flash runs a fatter 1879 calls but still beats a fixed 3200-call schedule.
On mathematical optimization the paper claims 'Over 50×' budget savings within a thousand generations, with Sum-Difference and Circle Packing matching or exceeding SimpleTES. GPU-kernel results on KernelBench are more modest: 2.43× fewer generations for VGG16, 1.79× for LayerNorm.
The paper does not report a wall-clock figure for the dreaming pass itself, and the empirical claim rests on 8 tasks spanning 3 domains. Single-lab, posted a day ago.
It's the second discovery-agent paper we've flagged this week, after PHAI Labs' foundation-model taxonomy.
Originally reported by huggingface.co
Read the original article →Original headline: Dream-RSI Paper Reuses Discovery History as Replay Simulator, Cuts Agent Calls 162x on Lasso