NaturalThoughts: Selecting and Distilling Reasoning Traces for General Reasoning Tasks
This work addresses the challenge of efficient knowledge distillation for general reasoning tasks, offering a method to enhance smaller models' performance in STEM domains, though it is incremental as it builds on existing distillation techniques.
The paper tackled the problem of improving student models' reasoning capabilities by systematically selecting high-quality reasoning traces from a teacher model, finding that selecting difficult examples with diverse reasoning strategies is more sample-efficient than random scaling, and training with NaturalThoughts outperformed existing datasets on benchmarks like GPQA-Diamond, MMLU-Pro, and SuperGPQA.
Recent work has shown that distilling reasoning traces from a larger teacher model via supervised finetuning outperforms reinforcement learning with the smaller student model alone (Guo et al. 2025). However, there has not been a systematic study of what kind of reasoning demonstrations from the teacher are most effective in improving the student model's reasoning capabilities. In this work we curate high-quality "NaturalThoughts" by selecting reasoning traces from a strong teacher model based on a large pool of questions from NaturalReasoning (Yuan et al. 2025). We first conduct a systematic analysis of factors that affect distilling reasoning capabilities, in terms of sample efficiency and scalability for general reasoning tasks. We observe that simply scaling up data size with random sampling is a strong baseline with steady performance gains. Further, we find that selecting difficult examples that require more diverse reasoning strategies is more sample-efficient to transfer the teacher model's reasoning skills. Evaluated on both Llama and Qwen models, training with NaturalThoughts outperforms existing reasoning datasets such as OpenThoughts, LIMO, etc. on general STEM reasoning benchmarks including GPQA-Diamond, MMLU-Pro and SuperGPQA.