A new training method teaches AI models to reason using a sliver of the data researchers used before.
Researchers propose an uncertainty-guided progressive contraction framework for reasoning post-training, the stage where a model learns to produce step-by-step answers after its initial training. Instead of training on every example a teacher model generates, the method builds Monte Carlo reasoning trees to find the moments where a student model is most uncertain, then narrows the training set to those high-value examples. A second stage keeps adjusting during fine-tuning, routing optimization based on how the student is learning in real time. Tested across multiple reasoning benchmarks and several model families, the approach used as little as 0.22% of the optimization budget of standard full-data training, over 450 times less exposure, while matching or improving performance.
Reasoning post-training is one of the most expensive steps in building today's AI systems, and most labs have scaled it by throwing more data and compute at the problem. This result suggests the real constraint isn't the volume of training examples but which examples teach a model something it doesn't already know, a different lever than the brute-force scaling race that has defined the field.
If it holds up outside benchmark conditions, this is the kind of efficiency gain that matters more to AI budgets than any new chatbot feature.