“The abstract changed how I thought about this domain (advanced reasoning models).”
s1: Simple test-time scaling
1000 条样本加 Wait 续写,让模型推理时多想
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%。
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。
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“The paper s1: Simple test-time scaling concludes that this is probably how OpenAI implemented the”
“Really, [what the underlying paper] shows is how much you can improve existing models with several of the techniques released by openai recently”
“* TinyZero: [LINK] * s1 Simple Test Time Scaling: s1”
“It's a method used to control thinking token generation showcased in this paper: s1”
“These tokens DO extend the thinking time.”
看全部 10 条 →
“a paper with a similar idea on scaling test time reasoning, this is sorta how all the thinking models work under the hood. s1”
“Reminded me of: s1”
“LLMs can self-correct, evidenced by the power of getting models to question themselves by emitting "Wait," in S1.”
“First described here I think, specifically for Qwen models: s1, although it seems unclear how much it actually improves the final response”
- Test-time scaling new approach: extra test-time compute improves LLM reasoning ↗TaurenHunter · 2025-02-08 · 2 分
- s1: Simple Test-Time Scaling ↗btilly · 2025-02-10 · 2 分
- S1: Simple Test-Time Scaling ↗bicepjai · 2025-06-18 · 3 分