English

Learning to Think from Multiple Thinkers

Machine Learning 2026-04-28 v1 Artificial Intelligence Computational Complexity Machine Learning

Abstract

We study learning with Chain-of-Thought (CoT) supervision from multiple thinkers, all of whom provide correct but possibly systematically different solutions, e.g., step-by-step solutions to math problems written by different thinkers, or step-by-step execution traces of different programs solving the same problem. We consider classes that are computationally easy to learn using CoT supervision from a single thinker, but hard to learn with only end-result supervision, i.e., without CoT (Joshi et al. 2025). We establish that, under cryptographic assumptions, learning can be hard from CoT supervision provided by two or a few different thinkers, in passive data-collection settings. On the other hand, we provide a generic computationally efficient active learning algorithm that learns with a small amount of CoT data per thinker that is completely independent of the target accuracy ε\varepsilon, a moderate number of thinkers that scales as log1εloglog1ε\log \frac{1}{\varepsilon}\log \log \frac{1}{\varepsilon}, and sufficient passive end-result data that scales as 1εpolylog1ε\frac{1}{\varepsilon}\cdot poly\log\frac{1}{\varepsilon}.

Keywords

Cite

@article{arxiv.2604.24737,
  title  = {Learning to Think from Multiple Thinkers},
  author = {Nirmit Joshi and Roey Magen and Nathan Srebro and Nikolaos Tsilivis and Gal Vardi},
  journal= {arXiv preprint arXiv:2604.24737},
  year   = {2026}
}

Comments

Comments are welcome. There are 78 pages and 5 Figures

R2 v1 2026-07-01T12:37:39.853Z