First, we curate a small dataset 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 doublecheck its answer, often fixing incorrect reasoning steps. After supervised finetuning [our] model on [our dataset] and equipping it with budget forcing, our model exceeds [OpenAI’s] o1-preview on competition math questions by up to 27%.
Muennighoff, N., Yang, Z., Shi, W., Li, X. L., Fei-Fei, L., Hajishirzi, H., … & Hashimoto, T. (2025). s1: Simple test-time scaling. arXiv preprint arXiv:2501.19393.
This paper demonstrates a remarkably efficient approach to improving AI language models’ reasoning capabilities. The key innovation is achieving strong performance with minimal training data and a simple methodology:
- Data efficiency: The researchers achieved impressive results by training on just 1,000 carefully selected examples, rather than the millions of examples typically used. This was accomplished by focusing on three key criteria when selecting training data:
- Quality: Using well-formatted, clear examples
- Difficulty: Including challenging problems that require significant reasoning
- Diversity: Covering a wide range of topics and problem types
- Simple approach: They developed a technique called “budget forcing” that controls how much time/computation the AI model spends thinking about a problem. This allows the model to scale its performance by spending more time on difficult problems when needed.
- Results: Using this approach, their model achieved competitive performance with more complex models while using far less training data. On certain mathematical reasoning tasks, it even outperformed some leading closed-source AI models.
The implications for education are significant, showing that AI systems can be trained to handle complex reasoning tasks with relatively small but well-curated datasets, rather than requiring massive amounts of data. This could make it more feasible for educational institutions to develop specialised AI tools tailored to specific educational contexts and needs.
There was a spate of news towards the end of last year, pointing to the lack of intelligence advances coming out of frontier AI labs. I said at the time that there are many routes to advancing the field further, even without an increase in baseline intelligence. These are the kinds of advances I was talking about.