A new training method teaches small AI models to reason better without needing more data to do it.
BoundaryEvo is a two-stage technique for distilling long reasoning chains from a large teacher model into a smaller student model. Most existing approaches pick a fixed set of training examples upfront, which ignores a basic problem: examples the student has already mastered teach it nothing new, and examples far beyond its current ability are too hard to learn from. BoundaryEvo first primes the student with a small batch of teacher-generated reasoning examples, then repeatedly swaps in new questions the student has not yet mastered but the teacher can still solve, updating the training set as the student improves. Using only 1,000 training examples, the method lifted Qwen3.5-4B's accuracy on IMO-AnswerBench from 62.9 to 75.3, a gain of about 20 percent, and on AMO-Bench from 26.8 to 46.0, a gain of about 72 percent, outperforming both static and dynamic data-selection baselines across 11 reasoning benchmarks.
The interesting part is what BoundaryEvo does not need: a bigger teacher dataset. If matching training data to a model's current ability matters more than raw volume, that changes the calculus for anyone trying to build capable small models on a budget, not just the labs that can afford to generate millions of reasoning traces.
The researchers have not yet released the code or data, so these numbers are still self-reported and worth treating as promising rather than proven.