Learning More by Training Less: Uncertainty-Guided Progressive Contraction for Efficient Reasoning Post-Training
Quick Answer
The proposed uncertainty-guided progressive contraction framework reduces training costs by over 450x while maintaining or improving reasoning performance in large language models.
Quick Take
This method compresses reasoning supervision and optimization, using only 0.22% of the full-data optimization budget across diverse benchmarks. It effectively enhances reasoning capability and generalizes well across various student models and transfer scenarios.
Key Points
- Utilizes only 0.22% of the optimization budget compared to full-data training.
- Achieves over 450x reduction in training exposure while maintaining performance.
- Employs Monte Carlo reasoning trees to profile uncertainty in reasoning.
- Generalizes effectively across diverse student models and transfer scenarios.
- Focuses optimization on the most valuable reasoning trajectories for learning.
DeepSignal Analysis
What happened
A new framework called uncertainty-guided progressive contraction has been introduced to enhance reasoning in large language models while significantly reducing training costs. This method compresses reasoning supervision and optimization, achieving a reduction of over 450 times in training exposure.
Key evidence
- The framework uses only 0.22% of the full-data optimization budget during training, indicating a substantial reduction in resource requirements.
- Experiments show that the proposed method maintains or improves reasoning performance across diverse benchmarks and large language model families.
- The approach generalizes well across various student models and transfer scenarios, suggesting its versatility in different applications.
Why it matters
This development is significant as it addresses the escalating costs associated with training large language models, which often do not yield proportional performance improvements. By effectively compressing the training process, this framework could make advanced reasoning capabilities more accessible and efficient, potentially broadening the scope of applications for AI systems.
What to watch
Paper Resources
📖 Reader Mode
~2 min readAbstract: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.
| Subjects: | Artificial Intelligence (cs.AI) |
| Cite as: | arXiv:2610.12933 [cs.AI] |
| (or arXiv:2610.12933v1 [cs.AI] for this version) | |
| https://doi.org/10.48550/arXiv.2610.12933 arXiv-issued DOI via DataCite (pending registration) |
Submission history
From: Suorong Yang [view email]
[v1]
Fri, 9 Oct 2026 03:26:06 UTC (481 KB)
— Originally published at arxiv.org
Want this in your inbox every morning?
Daily brief at your local 8am — bilingual EN/中文, free.
More from arXiv cs.AI
See more →VEPA: Verified Evolving Formal Policy for Lifelong Agent Guardrails
VEPA introduces a framework for verified evolving formal policies for LLM agents, enhancing safety by adapting guardrails based on deployment feedback. It outperforms existing methods across benchmarks like ASSEBench-Safety and ASSEBench-Security, ensuring deterministic policy enforcement even under challenging conditions.