Large language models have made remarkable strides in complex reasoning, yet scaling post‑training with ever‑larger corpora inflates cost without proportional gains. This work asks how far reasoning supervision can be compressed without hurting capability. We introduce an uncertainty‑guided progressive contraction framework that concentrates supervision and optimization on the most valuable reasoning trajectories. Instead of treating a teacher‑generated chain‑of‑thought (CoT) as a single path, we build Monte Carlo reasoning trees around intermediate states to obtain an uncertainty profile. This uncertainty first contracts the original corpus into a compact teacher‑informed pool. A second, learner‑aware contraction performs online routing during SFT, directing optimization according to the student’s evolving learning state. The two stages progressively shrink both supervision and optimization, dramatically cutting the student’s training budget. Experiments across diverse reasoning benchmarks and LLM families show the method uses only 0.22% of the full‑data SFT optimization budget—over 450× less training exposure—while maintaining or improving reasoning performance. The framework generalizes well across heterogeneous student families, black‑box teachers, and weak‑to‑strong transfer, indicating it not only speeds up training but can also enhance reasoning capability.
Review