PreSTO Speeds Up LLM Power Sampling Up to 2.33x via Predictive Subtree Prefetching
du_yilun · x · 2026-10-12
A new arXiv paper, PreSTO, accelerates power sampling — a test-time technique that boosts LLM reasoning by sampling from a sharpened model distribution without extra training — without changing the sampler's semantics.
Problem: The PowerMH sampler relies on sequential Metropolis–Hastings updates, where each accept/reject decision sets the state for the next proposal.
Method:
- Views future MH outcomes as a tree of accept/reject decisions and batch-generates a prefetched subtree of proposals per model call, provably leaving the MH transition kernel unchanged
- A predictive traversal rule exploits that MH only visits sparse, near-unary paths, choosing subtrees to maximize predicted completed transitions
Results: Across 7 datasets and 5 base LLMs, PreSTO achieves median speedups of 1.71x, 1.91x, and 2.33x on PowerMH, EntropyCut, and MultiTryMH; on the first two it makes 3.9x fewer model calls, each only 2.1x longer.
Related event: PreSTO Speeds Up LLM Power Sampling by Up to 2.33x(2 posts)→
More from Research
- Work2Skill Turns Human GUI Demos into Reusable Agent Skills, Beating No-Skill Baselines — nex-agi · 2026-10-12
- TouchScale: 500-Hour Vision-Tactile Dataset Lifts Robot Manipulation Success from 22.5% to 57.5% — 2077AIDataFoundation · 2026-10-12
- UCSD paper bakes theory of mind into RL reward, letting a 7B agent model partners solo — rohanpaul_ai · 2026-10-12
- All of Science Embedded: 203M Papers Free for AI Agent Search — evilsocket · 2026-10-12
- Physics law sets diffusion model's noise schedule for better particle spread prediction — bravo_abad · 2026-10-12
- Decision Index 0.3.1: vLLM d3 and Mapika Decider 31B tie for #1 — multimodalart · 2026-10-12