Preprint reports reasoning training with 0.22% of full-data optimization budget

Researchers report a method that maintained or improved language-model reasoning performance while using as little as 0.22% of the optimization budget of full-data supervised fine-tuning, according to an arXiv preprint. Their experiments span multiple reasoning benchmarks and model families, with over 450x less training exposure. The method first explores alternative reasoning paths around intermediate steps to identify uncertainty and select a compact pool of teacher-generated examples. During training, a second selection stage routes optimization according to the student model's evolving learning state. The reported reduction concerns the student optimization budget; the approach also requires constructing reasoning trees to select supervision.