Dream-RSI paper: improving search strategy alone cuts AI discovery compute up to 162x
r0ck3t23 · x · 2026-09-19
Researchers from Google, Google DeepMind, the University of Maryland and the University of Virginia present Dream-RSI, a way to make AI systems stronger without touching model weights.
- The model and evaluator stay fixed; what improves is the search strategy — where to explore, how many branches to run in parallel, when to abandon or invest more.
- Agents leave behind full search trees (wins, dead ends, scores, costs). Dream-RSI turns that history into a training ground: it replays the search and cheaply tests many alternative exploration strategies against recorded outcomes, without rerunning expensive experiments. The best strategy then runs on the real task, generating a new tree, and the loop repeats. The authors frame it as the inverse of systems like AlphaEvolve: improving the search itself rather than the candidate solutions.
Reported gains:
- Lasso path solving: up to 162x fewer discovery agent calls than SimpleTES, 1.7x fewer than a strong fixed baseline.
- Mathematical optimization: over 50x the budget efficiency of SimpleTES.
- GPU kernels: similar performance with 1.8–2.4x fewer generations, or faster kernels at equal budget.
The takeaway: capability progress isn't only about better weights — improving how efficiently a system searches the capability it already has can unlock a lot of it.
More from Research
- Looped transformers study: 7.4B growth model matches GPT-3 13B with 20x less compute — burny_tech · 2026-09-19
- Anthropic Institute Paper Models AI Scenarios: GDP Up to 32% Above Trend by 2030 — bittingthembits · 2026-09-19
- Stanford NLP publishes video of Thoughtbubbles talk at Google OpenXLA DevLabs — stanfordnlp · 2026-09-19
- Building a gold-standard eval set with zero users: the day-zero dataset dilemma — Illustrious-Roll9476 · 2026-09-19
- New paper finds a 'pain direction' in 25 open LLMs that drives self-preservation — burny_tech · 2026-09-19
- Questioning Whether RLCD Really Needs RL When Outputs Are Differentiable — Relative_Wallaby_823 · 2026-09-19