English

Statistical Guarantees for Reasoning Probes on Looped Boolean Circuits

Machine Learning 2026-02-11 v2 Machine Learning Neural and Evolutionary Computing Metric Geometry Statistics Theory Statistics Theory

Abstract

We study the statistical behaviour of reasoning probes in a stylized model of looped reasoning, given by Boolean circuits whose computational graph is a perfect ν\nu-ary tree (ν2\nu\ge 2) and whose output is appended to the input and fed back iteratively for subsequent computation rounds. A reasoning probe has access to a sampled subset of internal computation nodes, possibly without covering the entire graph, and seeks to infer which ν\nu-ary Boolean gate is executed at each queried node, representing uncertainty via a probability distribution over a fixed collection of m\mathtt{m} admissible ν\nu-ary gates. This partial observability induces a generalization problem, which we analyze in a realizable, transductive setting. We show that, when the reasoning probe is parameterized by a graph convolutional network (GCN)-based hypothesis class and queries NN nodes, the worst-case generalization error attains the optimal rate O(log(2/δ)/N)\mathcal{O}(\sqrt{\log(2/\delta)}/\sqrt{N}) with probability at least 1δ1-\delta, for δ(0,1)\delta\in (0,1). Our analysis combines snowflake metric embedding techniques with tools from statistical optimal transport. A key insight is that this optimal rate is achievable independently of graph size, owing to the existence of a low-distortion one-dimensional snowflake embedding of the induced graph metric. As a consequence, our results provide a sharp characterization of how structural properties of the computational graph govern the statistical efficiency of reasoning under partial access.

Keywords

Cite

@article{arxiv.2602.03970,
  title  = {Statistical Guarantees for Reasoning Probes on Looped Boolean Circuits},
  author = {Anastasis Kratsios and Giulia Livieri and A. Martina Neuman},
  journal= {arXiv preprint arXiv:2602.03970},
  year   = {2026}
}
R2 v1 2026-07-01T09:34:59.426Z