The authors introduce an uncertainty‑guided progressive contraction framework to compress reasoning supervision during post‑training of large language models. First, they treat each teacher‑generated chain‑of‑thought not as a single path but as a set of Monte‑Carlo reasoning trees built around intermediate states, which reveals process‑level uncertainty. This uncertainty is used to contract the original supervision corpus into a compact, teacher‑informed pool that retains only the most informative trajectories. Second, during supervised fine‑tuning they perform an online, learner‑aware contraction that routes optimization effort toward the examples where the student model exhibits the highest uncertainty, thereby aligning updates with the model’s evolving learning state. The two stages jointly shrink both the supervision set and the optimization budget. Empirical results across multiple reasoning benchmarks and LLM families show that the method requires as little as 0.22 % of the optimization budget of conventional full‑data SFT—over a 450‑fold reduction in training exposure—while preserving or even improving reasoning performance. The approach generalizes to heterogeneous student models, black‑box teachers, and weak‑to‑strong transfer scenarios, indicating that it not only accelerates training but can also enhance reasoning capability.