Source-linked AI summary
The Gold Rush in AI4Math: Where Are We Now?
Jiashun Jin, Zheng Tracy Ke, Bingcheng Sui
TL;DR
The paper asks how AI is being adopted in mathematical research, where empirical evidence remains limited. It audits Mathematics arXiv submissions using disclosure-based classification and finds rapidly expanding but uneven adoption, while cautioning that disclosed use and reported outcomes are not complete or independently verified.
Problem
The paper examines limited empirical evidence about how mathematicians use AI in research, beyond whether systems can solve selected mathematical problems.
Method
The authors audit 32,944 Mathematics arXiv submissions using a questionnaire that classifies disclosed AI use, research roles, open-problem status, affiliations, countries, and systems.
Results
AI-assisted mathematics shows rapidly expanding but highly uneven adoption, with substantive use, field participation, open-problem involvement, authorship, institutions, and model use concentrated.
Takeaways & Limitations
The study provides a systematic empirical baseline for monitoring how, where, and by whom AI enters mathematical research practice.
Takeaways & Limitations
Reported prevalence measures disclosed use only, while classifications and open-problem outcomes rely on LLM review rather than independent mathematical verification.
Abstract
from arXiv · showhide
Recent advances in artificial intelligence (AI) have sparked growing interest in its use for mathematical research. While some view this as a major opportunity for discovery, others have raised concerns about its impact on traditional research practices. Despite extensive debate, empirical evidence on how AI is actually being used in mathematics remains limited. To address this gap, we collected all 32,944 arXiv submissions posted between March 1 and August 20, 2026, whose primary or secondary categories included Mathematics. We identified 3,575 submissions that explicitly disclosed author use of AI, of which 1,712 involved at least one substantive mathematical contribution. Our analysis reveals several broad patterns. First, disclosed AI use increased sharply over the study period, with substantive use growing from 1.39% of Mathematics submissions in March to 14.09% through August 20. Second, substantive AI use is highly uneven across fields: Combinatorics has the largest number of such papers, while Metric Geometry has the highest substantive-use rate. Third, substantive AI use is geographically concentrated: under weighted author counts, the United States and China together account for about two-thirds of the recognized country weight. Fourth, AI is already being applied to open research problems: among 717 named open-problem records associated with substantive use, 71% are labeled as fully resolved based on the authors' descriptions, with proofs of the conjectured statement more common than counterexamples or disproofs. Finally, AI-system use is also highly concentrated, with OpenAI systems appearing most frequently, followed by Anthropic. Together, these findings suggest that AI-assisted mathematics is expanding rapidly but remains at an early and uneven stage of adoption.
1 Introduction
The paper addresses limited empirical evidence about how AI is actually being adopted in mathematical research. Using recent arXiv submissions, it documents adoption patterns, research roles, open-problem involvement, and the systems used.
- Existing studies assess whether AI can solve or verify selected mathematical problems, but provide limited evidence on real-world research adoption.
- 32,944 Mathematics-category arXiv submissions yielded 3,575 papers disclosing author AI use, including 1,712 with substantive mathematical contributions.The study also identified 1,225 papers reporting AI involvement in proof construction.
- Substantive AI use increased rapidly during the study period, while Figure 1 tracks seven-day-window counts and shares alongside model releases and notable events.
- AI adoption is uneven across mathematics, concentrated in fields such as Combinatorics and Number Theory and among relatively few authors and institutions.
- AI is already being used on named open problems, including both unresolved questions and problems authors report as resolved.The paper cautions that these records are not independently verified AI solutions.
- The study establishes an empirical baseline for tracking how, where, and by whom AI is used in mathematical research.
2 Data and study design
The study systematically analyzes Mathematics arXiv submissions using a disclosure-based questionnaire. AI-assisted questionnaire responses classify use, open-problem status, affiliations, and countries, with manual validation of a sampled subset.
- The dataset contains 32,944 Mathematics-category submissions posted between March 1 and August 20, 2026, with PDFs and metadata collected for each.Article text was extracted page by page, with boundaries explicitly marked.
- The questionnaire separates substantive and non-substantive AI use into four non-exclusive subcategories each, assigning a category-level yes when any subcategory is positive.
- Open-problem records are coded by status as open, partially resolved, resolved true, resolved false, or unclear, alongside whether AI was critical to the reported resolution.
- Institutional and country information is extracted from affiliations or email addresses and supplemented by web searches when necessary.
- An AI system answers the questionnaire from authors’ disclosures, identifying statements and semantic meaning rather than assessing mathematical content.
- Manual validation samples articles found through keyword screening of parsed full texts, which identified approximately 5,000 candidate articles.
3 Results
Disclosed AI use in mathematics rose rapidly but remained uneven across use types, fields, open-problem activity, and authors. Substantive use extended beyond editorial assistance to proofs, formalization, and named research problems, though reported resolutions were not independently verified.
- 3.1 A sharp recent increase, from a small base: 14.09%: substantive AI use among Mathematics submissions by August 20, up from 1.39% in March.Overall confirmed AI use increased from 4.75% to 24.14% over the same period.
- 3.1 A sharp recent increase, from a small base: 1,225 papers reported AI involvement in proof construction, followed by 427 in formalization and verification, 289 in other research assistance, and 146 in problem formulation.These substantive subcategories are non-exclusive, so papers may appear in multiple categories.
- 3.2 Which areas of mathematics are most affected?: 397 substantive-use papers came from Combinatorics, the largest field count, while Metric Geometry had the highest rate at 13.11% (40 of 305 papers).Combinatorics ranked second by rate at 11.74%; large fields such as Numerical Analysis had lower rates.
- 3.3 What open problems have been solved?: 717 named open-problem records were identified among substantive-use papers, spanning problems proposed across many decades and multiple mathematical fields.Among 329 records with reported proposed years, 92 dated before 2000 and 36 to 2026.
- 3.3 What open problems have been solved?: 510 of 717 open-problem records were labeled fully resolved: 329 resolved true and 181 resolved false, while 103 remained open and 93 were partially resolved.The labels came from LLM-based review of authors’ descriptions and were not independent mathematical verification.
- 3.4 Who is using AI for mathematical research?: The United States accounted for 33.7% and China 32.9% of recognized weighted country assignments, together representing about two-thirds.Author-level participation was also concentrated: 83.6% of canonical authors appeared on one substantive-use manuscript and 11.4% on two.
4 Discussion and conclusion
AI-assisted mathematical research is expanding rapidly but remains unevenly distributed and difficult to validate independently. The authors call for continued disclosure audits, stronger human validation, and independent verification.
- Discussion: AI is moving beyond auxiliary assistance toward direct participation in mathematical research, including proof construction, formalization, and work on open problems.Hundreds of named open-problem records appear among substantive-use manuscripts, but they are not independently verified AI solutions.
- Discussion: Adoption is concentrated in particular fields, repeated users, institutions, countries, and model providers.The evidence also shows descriptive differences between substantive-use and non-substantive-only manuscripts, most clearly in manuscript length.
- Limitations: The study measures disclosed use rather than total use, and its LLM-based classifications are subject to classification error.Country and institution assignments also rely partly on normalized and inferred affiliation information.
- Limitations: Open-problem outcome labels are not substitutes for peer review, historical-priority research, or independent verification of mathematical correctness.They should be interpreted as evidence about the kinds of claims appearing in AI-assisted mathematics, not definitive counts of problems solved by AI.
- Conclusion: The authors recommend repeating disclosure audits alongside stronger human validation and independent verification to assess whether the current gold rush becomes durable transformation.This recommendation follows the observed sharp growth over six months and emergence of substantive use across fields and research problems.
Disclosure of AI Use
The authors used different AI systems for distinct manuscript-related tasks and reviewed all AI-assisted outputs themselves. They retain full responsibility for the final manuscript.
- ChatGPT assisted with language editing, organization, and manuscript preparation.
- Codex assisted with data collection, data processing, data analysis, and figure generation.
- The authors reviewed all AI-assisted outputs and take full responsibility for the final manuscript.