s1: Simple test-time scaling
Muennighoff et al., 2025, s1 (Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing)
Subscriber deck — free during early access
2 cards in “2.2 Final selection of 1K samples” · whole deck
2.2 Final selection of 1K samples
- How did s1 filter 59K questions down to s1K?Quality: drop API errors and badly formatted samples. Difficulty: drop questions that Qwen2.5-7B-Instruct or Qwen2.5-32B-Instruct already solves. Diversity: repeatedly pick a domain (from the Mathematics Subject Classification) uniformly at random, then a question from it, favoring longer reasoning traces, until there are 1,000 across 50 domains.
- Muennighoff et al., 2025, s1 (Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing), §2.2, p. 20277
- About half of the s1K answers are wrong. Why did it still work?The model is fine-tuned to imitate the reasoning process, not to memorize answers: SFT here mainly teaches reasoning patterns, so traces with wrong final answers still help.
- Muennighoff et al., 2025, s1 (Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing), §2.2, p. 20277
Source
- Niklas Muennighoff, Zitong Yang, Weijia Shi, Xiang Lisa Li, Li Fei-Fei, Hannaneh Hajishirzi, Luke Zettlemoyer, Percy Liang, Emmanuel Candès and Tatsunori Hashimoto. 2025. s1: Simple test-time scaling. In Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing, pp. 20275–20321. License: CC BY 4.0. Flashcards adapted from this work. doi:10.18653/v1/2025.emnlp-main.1025 https://aclanthology.org/2025.emnlp-main.1025/