s1: Simple test-time scaling

Niklas Muennighoff, Zitong Yang, Weijia Shi, Xiang Lisa Li, Li Fei-Fei, Hannaneh Hajishirzi, Luke Zettlemoyer, Percy Liang, Emmanuel Candès, Tatsunori Hashimoto

Read on arXiv

Type: paper
Venue: arXiv / Stanford / Ai2
Year: 2026
Source: arxiv
Access: free
Language: en
Added: 2026-08-14T20:26:00Z
Verified: 2026-08-14T20:26:00Z

Summary

Stanford/Ai2/UW (cite Muennighoff et al.; arXiv 2501.19393). Curates s1K: 1,000 hard/diverse/quality questions with Gemini Thinking traces, distilled from a 59K pool (NuminaMATH, AIME 1983-2021, OlympicArena, OmniMath, AGIEval, plus original s1-prob/s1-teasers). SFT of Qwen2.5-32B-Instruct for 26 min on 16 H100s. Budget forcing: append end-of-think to cap tokens or suppress it and append "Wait" to extend thinking. s1-32B exceeds o1-preview on MATH/AIME24 by up to 27%; AIME24 50% without BF to 56.7% with BF (extrapolates 50%→57%). Ablations: random/diverse/longest 1K all worse (~−30% AIME24); 59K-full is not substantially better. Follow-up s1.1 regenerates traces with DeepSeek r1. Open model/data/code.

Keywords

s1 · test-time-scaling · budget-forcing · s1k · reasoning · qwen2.5 · distillation · stanford

Topics

test-time scaling, reasoning, sample-efficient SFT

Research notes

Primary: arxiv abs 2501.19393 (cs.CL; arXiv perpetual non-exclusive 1.0). Code https://github.com/simplescaling/s1 (Apache-2.0). Data https://huggingface.co/datasets/simplescaling/s1K (Apache-2.0, 1,000 rows). Model simplescaling/s1-32B. Discord posted HTML v1. s1K is a tiny SFT set, not a 27-field scorecard corpus.