Reasoning models tackle complex, multi-step tasks by producing intermediate reasoning before an answer. They can reduce the need for chain-of-thought examples in prompts, but generally require more processing time and longer context windows. section 4.1.1 section 5.2.2
Choose and evaluate them with the task and operating cost in view:
- For math, code, or planning, prefer a reasoning-specialist model as a distillation teacher so the student can learn reasoning traces, not only final answers. section 7.1.2
- Allow longer output lengths, filter for coherent reasoning, and evaluate intermediate steps as well as final answers. section 7.1.2
- Recheck few-shot baselines on current frontier models; reasoning models may need fewer examples, while added examples can increase inference compute. section 4.1.1
Selecting a teacher
For a math or planning distillation task, a DeepSeek-R1-style reasoning specialist may be a better teacher than a general-purpose model. Check its accuracy on curated examples before generating training data. section 7.1.2






