Learning More by Training Less: Compressing Reasoning Post-Training with Uncertainty-Guided Contraction

Summary

Large language models can gain strong reasoning abilities, but expanding post-training corpora raises costs without guaranteeing proportional improvements. This study asks how much reasoning supervision can be compressed without sacrificing capability. It introduces uncertainty-guided progressive contraction, which treats a teacher-generated chain-of-thought as more than one fixed path: Monte Carlo reasoning trees are built around intermediate states to expose process-level uncertainty. A teacher-informed contraction first produces a compact supervision pool, while a second learner-aware stage routes optimization online during supervised fine-tuning according to the student's changing learning state. Across diverse reasoning benchmarks and language-model families, the method uses as little as 0.22% of the optimization budget used by full-data SFT, corresponding to more than 450 times less training exposure, while maintaining or improving reasoning performance. The framework also generalizes across heterogeneous student families, black-box teachers, and weak-to-strong transfer. The results suggest that targeted supervision can reduce training cost and may itself improve reasoning capability, although the abstract does not specify the benchmark-level results or limitations in detail.