s1: Simple test-time scaling

By Niklas Muennighoff, Zitong Yang, Weijia Shi, Xiang Lisa Li, Fei-Fei Li, Hanna Hajishirzi, Luke S. Zettlemoyer, Percy Liang, Emmanuel J. Candès, Tatsunori B. Hashimoto. Conference on Empirical Methods in Natural Language Processing, 2025

1,358 citations recorded by the source corpus.

Test-time scaling is a promising new approach to language modeling that uses extra test-time compute to improve performance. Recently, OpenAI's o1 model showed this capability but did not publicly share its methodology, leading to many replication efforts. We seek the simplest approach to achieve test-time scaling and strong reasoning performance. First, we curate a small dataset s1K of 1,000 questions paired with reasoning traces relying on three criteria we validate through ablations: difficulty, diversity, and quality. Second, we develop budget forcing to control test-time compute by forcefully terminating the model's thinking process or lengthening it by appending"Wait"multiple times to the model's generation when it tries to end. This can lead the model to double-check its answer, often fixing incorrect reasoning steps. After supervised finetuning the Qwen2.5-32B-Instruct language model on s1K and equipping it with budget forcing, our model s1-32B exceeds o1-preview on competition math questions by up to 27% (MATH and AIME24). Further, scaling s1-32B with budget forcing allows extrapolating beyond its performance without test-time intervention: from 50% to 57% on AIME24. Our model, data, and code are open-source at https://github.com/simplescaling/s1

Subfields: Artificial Intelligence, Control and Systems Engineering

View Paper: https://www.semanticscholar.org/paper/ef8a8bd193b1a0a5e2c834a7a28869a2ec85bab7

How researchers see this paper

The researchers whose trusted networks rank it highest, strongest advocate first. Every researcher gets a different order.

In Artificial Intelligence

1. Tong Lu
2. Zhaoyang Liu
3. Weijie Su
4. Dahua Lin
5. Yinan He
6. Xingwei Qu
7. Shaohui Lin
8. Conghui He

In Control and Systems Engineering

1. Dahua Lin
2. Zhaoyang Liu

Citations in the corpus

• L1: Controlling How Long A Reasoning Model Thinks With Reinforcement Learning (2025, neutral): “Another work S1 (Muennighoff et al., 2025) introduces “budget-forcing” by imposing a strict token limit: either truncating output at budget exhaustion or inserting a special token (“Wait”) to request continued generation until reaching the full length budget.”
• InternVL3.5: Advancing Open-Source Multimodal Models in Versatility, Reasoning, and Efficiency (2025, positive): “(2) Text-only data: this part of data is constructed based on the training corpora of InternLM series [9, 124] and is further augmented with open-source datasets [6, 73, 75, 94].”
• ReTool: Reinforcement Learning for Strategic Tool Use in LLMs (2025, positive): “Furthermore, when trained on DeepSeek-R1-Distill-Qwen-32B [4], our model demonstrates further improvements, surpassing competitive baselines such as QwQ-32B-Preview [23], s1-32B [10], and OpenAI o1-preview [11].”
• Vision-R1: Incentivizing Reasoning Capability in Multimodal Large Language Models (2025, positive): “…such as Chain-of-Thought (CoT) prompting methods [3], plan-based methods like Tree-of-Thought and Graph-of-Thought [4, 5], process-based reward models [6– 9], Monte Carlo Tree Search (MCTS) and Beam Search [10– 12], as well as constructing complex Supervised Fine-Tuning (SFT) datasets [13, 14].”
• Stop Overthinking: A Survey on Efficient Reasoning for Large Language Models (2025, neutral): “It includes GSM8K and AQuA for arithmetic problems, StrategyQA and HotPotQA for commonsense reasoning, ProntoQA for logical reasoning, Game of 24 and Bin Packing for algorithmic tasks, and BlocksWorld, Rubik’s Cube, TripPlan, and Calendar Plan for planning tasks.”
• Dream 7B: Diffusion Large Language Models (2025, neutral): “This timestep-based approach introduces an additional dimension for inference-time scaling (Snell et al., 2024; Muennighoff et al., 2025; Geiping et al., 2025), which works alongside existing techniques such as chain-of-thought reasoning used in large language models like OpenAI o1 (Jaech et al.,…”
• R&B: Domain Regrouping and Data Mixture Balancing for Efficient Foundation Model Training (2025, unknown): “We use S1-Reasoning dataset [25], which comprises reasoning traces from challenging math problems, drawn from 54 distinct sources.”
• Mieb: Massive Image Embedding Benchmark (2025, unknown): “We hypothesize that future models that better incorporate reasoning capabilities and test-time scaling techniques [20, 24, 37, 43, 62] may achieve better results on compositionality tasks.”