On November 12, 2026, researchers from Beijing Normal University, Microsoft Research Asia, and Renmin University of China published an arXiv preprint on AI mathematical proving styles (arXiv:2611.16527v1). Testing frontier language models across all 465 propositions of Euclid’s Elements, the authors found that AI systems systematically bypass established intermediate theorems to reconstruct proofs from foundational axioms, changing how mathematics departments, peer reviewers, and journal editors evaluate automated reasoning.
What happened
The study, titled “AI has its own style in mathematical proving,” was conducted by Xiaokai Li, Ziyu Zhou, Qibin Cheng, Yangcheng Wu, An Zeng, and Ying Fan. While recent breakthroughs have demonstrated that large language models can solve competition problems and assist with open mathematical questions, mathematicians have raised persistent concerns that AI-generated proofs are difficult to read, interpret, and mentally verify. To test whether these differences stem from systematic structural patterns rather than idiosyncratic phrasing, the research team used Euclid’s Elements as an axiomatic testbed spanning plane geometry, solid geometry, and number theory across 13 books and 607 total mathematical statements.
The researchers evaluated three frontier models: Anthropic’s Claude Opus 4.6 across three independent runs, Claude Fable 5.1, and OpenAI’s GPT-6 Astra, each configured with high reasoning effort. Using an orchestration workflow within the Claude Code platform, dedicated sub-agents solved each proposition in an isolated context, drawing only on available preceding propositions, definitions, and postulates. Generated proofs were evaluated through an automated pipeline and validated by human mathematicians, categorizing outcomes into perfect citation matches, granularity shifts, alternative deductive paths, and reasoning errors.
Methodology and key findings
Quantitative citation network analysis revealed that AI-generated proofs diverge sharply from human proofs while remaining highly consistent with each other. Across five AI runs, the mean pairwise Jaccard similarity among AI proofs was approximately 0.68—nearly twice the 0.37 mean similarity observed between AI and human proofs. AI models generated substantially denser citation graphs, averaging 3,575 edges for Opus and more than 4,900 edges for Fable and GPT-6, compared to 1,889 edges in the human baseline.
Examining citation choices uncovered what the authors termed a “building from the ground up” proving style. In human proofs, 86.8% of citations point to previously proven theorems, with foundational elements—definitions, common notions, and postulates—accounting for just 13.2%. In contrast, AI models directed approximately 40% to 47.5% of their citations to foundational elements (definitions ~24%, common notions ~10%, postulates ~6%), citing theorems only 52.5% to 60% of the time. Furthermore, ancestor analysis showed that approximately 47% to 52.7% of AI-only citations traced directly to upstream dependencies of the human proof. Instead of citing high-level intermediate lemmas that encapsulate earlier deductions, the models independently re-derived arguments from base principles. Explicit anti-grounding prompts instructing GPT-6 Astra to prefer mature theorems produced negligible change (Cohen’s d = 0.047), demonstrating that ground-up construction is an intrinsic property of current model reasoning.
Why it matters for research practice
These structural findings directly inform ongoing debates about transparency and legibility in AI-assisted scholarship. As highlighted by the Leiden Declaration on AI and Mathematics and statements from twenty-five Fields Medalists, mathematical proof serves two complementary purposes: establishing objective correctness and advancing human conceptual understanding. While ground-up proofs achieve formal validity, they bypass the modular abstraction layers that human mathematicians rely on to navigate complex arguments. Readers must follow extended dependency chains back to base axioms, significantly increasing cognitive evaluation costs.
For academic publishers and departmental hiring committees, this divergence alters how automated proofs must be audited. Human authors adhere to standard theorem-proof structure and style conventions that balance technical detail against conceptual clarity. When evaluating papers produced with AI assistance, editorial teams need tailored protocols for AI in peer review, including requiring a clear generative-AI disclosure statement specifying model versions, reasoning prompts, and human validation steps. In addition, integrating language models with interactive theorem provers like Lean or Isabelle will require intermediate translation layers that convert dense axiomatic derivations into human-readable lemma hierarchies, safeguarding long-term computational reproducibility across mathematics and computer science.
