Uncertainty-Guided Progressive Contraction for Efficient Reasoning Training

This framework compresses reasoning supervision by using Monte Carlo reasoning trees to identify process-level uncertainty in teacher-generated CoT trajectories. It concentrates optimization on the most informative data points, reducing training costs without proportional performance loss.