AI has its own style in mathematical proving

Xiaokai Li, Ziyu Zhou, Qibin Cheng, Yangcheng Wu, An Zeng, Ying FanarXiv 2026 · arXiv

TL;DR

Researchers have discovered that Large Language Models (LLMs) possess a distinct, reproducible "proving style" in mathematics that differs significantly from human conventions. By reproving Euclid’s Elements, the study shows that AI tends to "build from the ground up," favoring foundational axioms and distant definitions over the high-level, "mature" theorems typically cited by human mathematicians.

What Problem Does This Paper Address?

As LLMs like Claude and GPT-6 begin to tackle frontier mathematical research—such as disproving the Erdős unit distance conjecture—a new friction has emerged: mathematicians often find AI-generated proofs difficult to interpret. They are frequently described as verbose, dwelling on trivialities, and "not written for humans."

This paper investigates whether this difficulty stems from a fundamental difference in how AI organizes and invokes mathematical knowledge. The authors seek to move beyond qualitative observations of "AI verbosity" to a systematic, quantitative characterization of AI’s deductive style. By comparing how AI and humans construct proofs within a closed axiomatic system, they aim to identify the "concrete signature" of machine reasoning.

Methods

The researchers used Euclid’s Elements as a controlled testbed, consisting of 465 propositions. They tasked three state-of-the-art models—Claude Opus 4.6, Claude Fable 5.1, and GPT-6 Astra—to independently reprove every proposition.

The experimental design (detailed in Experimental pipeline) ensured that models proved propositions sequentially, citing only previously established definitions, postulates, or theorems. To analyze the results, the team:

  • Calculated Jaccard Similarity: Measuring the overlap between the sets of citations used by AI vs. humans.
  • Complex Network Analysis: Treating the entire body of proofs as a directed graph where nodes are propositions and edges are citations.
  • Classification: An AI judge (validated by human experts) categorized proofs as Perfect matches (P), Granularity differences (G), Path differences (D), or Errors (E).
  • Ancestor Analysis: Tracing whether AI’s "extra" citations were actually the underlying logical dependencies of the theorems humans cited.
Experimental pipeline

Datasets

The study utilized a structured version of Heath’s classical English translation of Euclid’s Elements, sourced from the Wolfram Data Repository. This dataset provides a rigorous human baseline, including 5 postulates, 5 common notions, 132 definitions, and 465 propositions.

Key Insights and Ideas

  • "Building from the Ground Up": The most striking finding is that AI models prefer to cite foundational elements (definitions and postulates) rather than intermediate theorems. While humans cite foundations only 13% of the time, AI does so in roughly 40% of its citations.
  • Stable and Reproducible Style: This isn't random hallucination. The "AI style" is remarkably consistent across different models and runs. As shown in Jaccard similarity distributions, the similarity between different AI runs (~0.68) is nearly double the similarity between AI and human proofs (~0.37).
  • Complexity Drives Divergence: The more complex a proposition (measured by proof length or available tools), the more likely the AI is to diverge from the human path. Simple problems lead to convergence; complex ones allow the AI to explore its own idiosyncratic logical routes.
  • Reinventing the Wheel: About half of the "extra" citations AI uses are actually ancestors of the human citations. Essentially, where a human cites a high-level theorem to save cognitive effort, the AI "re-derives" that theorem by citing its constituent parts.
Jaccard similarity distributions

Key Results

  • Network Density: AI citation networks are substantially denser than human ones. The human network contains 1,889 edges, while AI networks range from ~3,500 to over 4,900 edges.
  • Citation Distance: AI tends to reach further back in the deductive chain. The mean citation distance for AI was 111.2, compared to 86.3 for humans (see Citation types and distances).
  • Verbosity vs. Efficiency: AI proofs are consistently longer. Human proofs average 314 words, while Claude Fable averaged 702. Interestingly, GPT-6 Astra was the most "terse" AI, producing the most concise reasoning per citation.
  • Path Divergence: In the final results, roughly 27% of AI proofs followed a valid but "genuinely different route" (Category D) from the human baseline.
Citation types and distances

Limitations and Future Work

The authors acknowledge several constraints:

  • Domain Specificity: The study is limited to the geometry and number theory of Euclid. It remains to be seen if this "ground-up" style persists in modern, abstract fields like topology or analysis.
  • Natural Language Focus: The research looked at natural language proofs. The authors suggest extending this comparison to formal proof environments (like Lean or Coq).
  • Steerability: Initial attempts to "steer" the AI toward human-like citation patterns using prompt engineering had negligible effects. Future work is needed to determine if AI can be trained to adopt a more "human-readable" hierarchical proving style without sacrificing its logical rigor.

Summary updated

Abstract

As large language models increasingly generate mathematical proofs, whether they develop a proving style distinct from that of human mathematicians remains an open question. This question is particularly pressing as mathematicians increasingly voice concerns that proofs generated by large language models are difficult to interpret. Here we ask three large language models to independently reprove all 465 propositions of Euclid's \textit{Elements} and compare their citation choices against the human-proof baseline through Jaccard similarity and complex network analysis. We find that AI-generated proofs differ from human proofs, and this difference is stable and reproducible: mean pairwise Jaccard similarity among AI runs (~0.68) is nearly twice that between AI and human proofs (~0.37), and cross-run classification consistency is also high. By examining citation structure in detail, we identify the concrete signature of this difference: approximately 40% of AI citations point to foundational elements (definitions, common notions, and postulates), compared with only 13\% in human proofs; even among theorem-to-theorem citations, AI tends to cite more distant results. Roughly half of AI's extra citations trace to the upstream dependencies of the human proof. Together, these findings suggest that the AI proving style is a tendency to construct arguments on its own from foundational elements rather than directly citing later, more established results, which we term "building from the ground up". Our findings offer a new perspective on how AI systems organise and invoke mathematical knowledge, with implications for understanding the interpretability of AI-generated proofs.
BibTeX
@misc{li2026ai,
  author = {Xiaokai Li and Ziyu Zhou and Qibin Cheng and Yangcheng Wu and An Zeng and Ying Fan},
  title = {{AI has its own style in mathematical proving}},
  year = {2026},
  eprint = {2611.16527},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2611.16527},
}
Figure