按年被提到
20252025 · 10 次2026
大模型与语言模型›大模型推理与思维链·arXiv:2501.19393 · cs.CL

s1: Simple test-time scaling

s1:简单的测试时扩展
Niklas Muennighoff, Zitong Yang, Weijia Shi, Xiang Lisa Li, Li Fei-Fei, Hannaneh Hajishirzi, Luke Zettlemoyer, Percy Liang, Emmanuel Candès, Tatsunori Hashimoto · 2025

1000 条样本加 Wait 续写,让模型推理时多想

开始阅读 ↗PDFalphaXiv 讨论
被推荐 · 13 人首次 2025-02-04 · 最近 2026-07-20发表于 Conference on Empirical Methods in Natural Language Processing (EMNLP)
13不同的人
6活跃月份
13讨论串
77被引(OpenAlex)

作者构建了 1000 道题配推理轨迹的 s1K 数据集,用它对 Qwen2.5-32B-Instruct 做监督微调,并提出 budget forcing:强行终止模型的思考,或在它想结束时反复追加「Wait」来延长思考。得到的 s1-32B 在竞赛数学(MATH 和 AIME24)上最多比 o1-preview 高 27%,并且在 AIME24 上随思考预算增加从 50% 提升到 57%。

HN 为什么提它

评论者用它解释推理模型的「思考时长」可以靠强制追加 Wait 之类的 token 来控制,并以此说明模型为何能自我纠错。

方法难度 3 / 5约 35 分钟经得起时间
读之前最好知道
1大语言模型的监督微调(SFT)
2思维链(chain-of-thought)推理
3测试时计算(test-time compute)
4Transformer 语言模型

HN 书评

推荐在前,同类按时间 · 每篇最多保留 20 条,统计用全部提及
强烈推荐ozgune · 2025-02

“The abstract changed how I thought about this domain (advanced reasoning models).”

强烈推荐,称摘要改变了自己对推理模型这一领域的看法
S1: A $6 R1 competitor?原评论 ↗
引用xena · 2025-02

“The paper s1: Simple test-time scaling concludes that this is probably how OpenAI implemented the”

指出论文推断 OpenAI 可能就是这样实现 o1 的推理强度调节
Chat is a bad UI pattern for development tools原评论 ↗
引用wrenky · 2025-02

“Really, [what the underlying paper] shows is how much you can improve existing models with several of the techniques released by openai recently”

指出论文真正展示的是用若干技术能大幅提升现有模型
Researchers created an open rival to OpenAI's o1 'reasoning' model for under $50原评论 ↗
引用armcat · 2025-02

“* TinyZero: [LINK] * s1 Simple Test Time Scaling: s1”

与 TinyZero 并列列出,说明小数据加推理时扩展就能带来明显提升
LIMO: Less Is More for Reasoning原评论 ↗
引用lelag · 2025-03

“It's a method used to control thinking token generation showcased in this paper: s1”

指出这篇论文展示了控制 thinking token 生成的方法
QwQ-32B: Embracing the Power of Reinforcement Learning原评论 ↗
引用computerex · 2025-03

“These tokens DO extend the thinking time.”

指出追加这些 token 确实能延长思考时间
Tencent's 'Hunyuan-T1'–The First Mamba-Powered Ultra-Large Model原评论 ↗
看全部 10 条 →
引用mangoman · 2025-04

“a paper with a similar idea on scaling test time reasoning, this is sorta how all the thinking models work under the hood. s1”

贴出思路相似的论文,称思考模型的底层大体就是这种测试时推理扩展
Chain of Recursive Thoughts: Make AI think harder by making it argue with itself原评论 ↗
引用kgeist · 2026-04

“Reminded me of: s1”

联想到 budget forcing:追加 Wait 来延长思考,控制测试时算力
TinyLoRA – Learning to Reason in 13 Parameters原评论 ↗
引用mxwsn · 2026-07

“LLMs can self-correct, evidenced by the power of getting models to question themselves by emitting "Wait," in S1.”

拿 s1 里让模型输出 Wait 来质疑自己,说明 LLM 能自我纠错
The One-Step Trap (In AI Research)原评论 ↗
引用embedding-shape · 2026-07

“First described here I think, specifically for Qwen models: s1, although it seems unclear how much it actually improves the final response”

认为 budget forcing 最早见于此、针对 Qwen 模型,但对最终回答改善多少不清楚
Controlling Reasoning Effort in LLMs原评论 ↗
被提交为 story · 3 次