LEARNING MORE BY TRAINING LESS: UNCERTAINTY-GUIDED PROGRESSIVE CONTRACTION FOR EFFICIENT REASONING POST-TRAINING | Suorong Yang et al. | ResearchPod

Transcript: LEARNING MORE BY TRAINING LESS: UNCERTAINTY-GUIDED PROGRESSIVE CONTRACTION FOR EFFICIENT REASONING POST-TRAINING

Sam: You can throw away more than ninety-nine percent of your training data and still get a better reasoner. That's the claim, and the trick is knowing which solutions to keep.

Alex: That's a big claim. Whose work is this, and what kind of paper?

Sam: It's a method paper from Suorong Yang and colleagues at Nanjing University, the National University of Singapore, Nanyang Technological University and others. It's on arXiv, marked as an ICLR 2027 conference paper. They train smaller language models to reason by learning from a bigger teacher model's worked solutions.

Alex: The usual recipe there is just more solutions, more training. Why should someone outside this corner care?

Sam: Because that recipe is expensive, and the gains don't keep pace with the cost. Their best setting uses about a fifth of one percent of the usual student training. That's over four hundred and fifty times less exposure to training examples, with reasoning scores that hold up or improve. So we'll cover the one idea behind it, the evidence, and the catch I'd want you to keep in mind.

Alex: Start with what people did before. Surely nobody trains on everything blindly.

Sam: No. There are two camps. One builds better supervision, with stronger teachers or augmented solutions. The other filters a fixed pile of solutions by correctness, difficulty, or quality. The authors argue neither alone gets you extreme efficiency. You need to find where the learning value actually sits, then aim training at it.

Alex: And where does it sit? If a solution is correct, what more is there to know?

Sam: That's their starting observation. Reasoning isn't uniform. A solution often hinges on a few intermediate steps where plausible next moves lead to different outcomes. They call these reasoning forks. A correct final answer can hide a fragile path through one of them.

Alex: How do you see a fork from the outside? The teacher just hands you one chain of thought.

Sam: They treat that chain as a backbone. At each step, they take the solution so far and ask the teacher to finish it several separate times. Then they check how often those finishes reach the right answer. Picture it as a small tree of sampled continuations branching off every step. They call it a Monte Carlo reasoning tree.

Alex: So if every continuation from a step gets it right, that step is safe. If every one fails, it's just hard.

Sam: And neither of those is the interesting case. The score peaks when continuations split roughly half right, half wrong. Formally, it's a scaled variance, which works out to twice the chance that two continuations from the same step disagree. So it flags instability, not difficulty. A solution's overall score is the average across its steps.

LEARNING MORE BY TRAINING LESS: UNCERTAINTY-GUIDED PROGRESSIVE CONTRACTION FOR EFFICIENT REASONING POST-TRAINING

Suorong Yang, Jiaxin Wang, Jiaming Su, Chentao Yue, Furao Shen, Peng Ye, Soujanya Poria

Oct 9, 2026·7:09·arxiv·Original Paper

Reasoning Post-TrainingUncertainty-Guided Progressive ContractionChain-of-Thought (CoT) SupervisionMonte Carlo Reasoning TreesTraining EfficiencyReasoning Forks

Listen

7 min

Summary

Abstract

Large language models have achieved remarkable advances in complex reasoning. However, scaling reasoning post-training with ever-larger corpora can substantially increase training cost without proportional performance gains. In this study, we therefore ask: how far can reasoning supervision be compressed without sacrificing capability? To this end, we propose an uncertainty-guided progressive contraction framework that concentrates supervision and optimization on the reasoning trajectories most valuable for learning. Specifically, rather than treating a teacher-generated chain-of-thought (CoT) as a single realized path, we construct uncertainty-profiled reasoning supervision by building Monte Carlo reasoning trees around intermediate reasoning states to expose process-level uncertainty. This uncertainty first contracts the original corpus into a compact teacher-informed pool. A second, learner-aware contraction then performs online routing during SFT, concentrating optimization according to the student's evolving learning state. Together, the two stages progressively contract both supervision and optimization, substantially reducing the student training budget. Experiments across diverse reasoning benchmarks and LLM families show that our method uses as little as 0.22% of the optimization budget of full-data SFT, over 450x less training exposure, while maintaining or improving reasoning performance. Notably, our framework generalizes well across heterogeneous student families, black-box teachers, and weak-to-strong transfer. This indicates that it is not only a way to accelerate training but can also enhance reasoning capability.

LEARNING MORE BY TRAINING LESS: UNCERTAINTY-GUIDED PROGRESSIVE CONTRACTION FOR EFFICIENT REASONING POST-TRAINING: Research Paper Podcast

LEARNING MORE BY TRAINING LESS: UNCERTAINTY-GUIDED PROGRESSIVE CONTRACTION FOR EFFICIENT REASONING POST-TRAINING

Listen