The Mathematician's Assistant: Integrating AI into Research Practice

• 252 billion USD in global private AI investment was recorded in 2024, including 33.9 billion USD for generative AI—8.5 times the 2022 investment.Source: §1 · p. 2
• 78 % of surveyed businesses reported using AI in 2024, up from 55 % in 2023 and 20 % in 2017.Source: §1 · p. 2
• The 2024 Nobel Prizes in Chemistry and Physics recognized work on AlphaFold and artificial neural networks, respectively.Source: §1 · p. 2
• The article focuses on responsible AI use in mathematical research and does not address societal risks such as bias, environmental impact, or economic displacement.Source: §1 · p. 2

• AI’s mathematical progress is best understood as new problem-solving tools, with examples in creative competition reasoning and complex optimization rather than automated discovery.Source: §2 · pp. 2, 4
• In 2025, Gemini Deep Think won IMO gold by solving problems autonomously in natural language within the 4.5-hour limit.The IMO problems test creative problem-solving, and the model’s full solutions were publicly available for analysis.Source: §2 · p. 3
• AlphaEvolve uses Gemini models to generate and test code variants, iteratively refining successful solutions after an initial human setup.Source: §2 · p. 3
• 20 % of AlphaEvolve’s tested cases yielded novel solutions surpassing the best-known results.The developers applied it to over 50 problems in areas including analysis, number theory, geometry, and combinatorics.Source: §2 · p. 4
• 48 scalar multiplications improved on Strassen’s 49-multiplication algorithm for 4 × 4 matrices, the first improvement in over 56 years.The paper notes that more efficient matrix multiplication can accelerate AI training and inference.Source: §2 · p. 4

3.1.1. Answer-Based Performance.

• On MathArena’s January–May 2025 competitions, Gemini 2.5 Pro Exp, o3, and o4 mini high surpassed the top 1 % of human participants.The evaluation used numerical final answers; the benchmark drew problems from recent competitions to reduce contamination concerns.Source: §3.1 · pp. 4–5, §3.1.1 · p. 5

3.1.2. Proof-Based Performance.

• On the 2025 IMO, the top-performing accessible model Gemini 2.5 Pro scored 31.55%, while leading models scored around 30% on the USAMO; a gap remains to human medal-winning performance.Source: §3.1.2 · p. 6

3.2.1. Overall Model Performance in Proof Generation.

• Human evaluators judged 43 % of OPC-generated proofs correct, with Gemini 2.5 Pro and o3 leading across both dataset partitions.The second partition included more challenging problems, and the model roster differed between partitions.Source: §3.2.1 · p. 6, Figure 3 · p. 7

3.2.2. Discrepancy Between Final-Answer Accuracy and Proof Validity.

• Final-answer accuracy poorly predicts proof validity: the performance drop was approximately 8 % for Gemini 2.5 Pro and nearly 30 % for o3.The comparison suggests a correct final result can coexist with flawed logical justification.Source: §3.2.2 · p. 6, Figure 4 · p. 7

3.2.3. LLMs as Evaluators of Mathematical Proofs.

• Gemini 2.5 Pro achieved 85.4 % proof-judgment accuracy against a 90.4 % human baseline, but models were worse at identifying errors in their own proofs.The authors suggest using different models for proof generation and verification in iterative refinement.Source: §3.2.3 · p. 7, Figure 6 · p. 8

3.2.4. Common Failure Modes in Proof Generation.

• OPC proof errors included overgeneralization and flawed logical steps, while models often produced faulty proofs rather than admitting inability, even when prompted to express uncertainty.The cited failure patterns include errors with inequalities and geometric properties.Source: §3.2.4 · p. 8

3.2.5. The Impact of Best-of-n Sampling on Proof Quality.

• For o4 mini, ranking-based best-of-eight proof selection raised success from 26 % at pass@1 to 47 %.The method generates multiple candidate solutions and selects the most promising one.Source: §3.2.5 · pp. 8–9, Figure 5 · p. 8

3.4. FrontierMath: A Benchmark at the Research Frontier.

• 4.2 % success on FrontierMath Tier 4 was reported for Gemini 2.5 Pro and Claude Opus 4, compared with under 2 % for older models.The authors set aside OpenAI model results on this benchmark because of concerns about transparency and possible data contamination.Source: §3.4 · pp. 9–10

4.1. Guiding Principles for AI Collaboration.

• The mathematician remains the pilot: verify every output, account for persistent errors, choose prompts and models strategically, and experiment critically.The guide also cautions against treating models as human-like reasoners.Source: §4.1 · pp. 10–11

4.2.1. Google DeepMind’s Portfolio.

• Google pairs premium Gemini Deep Think with free Gemini 2.5 Pro, whose million-token context, adjustable creativity, and proof abilities support distinct research tasks.Gemini Deep Research creates cited literature summaries after a reviewable search plan.Source: §4.2.1 · pp. 11–12

4.2.2. OpenAI’s Product Stack.

• OpenAI’s pre-August 7, 2025 stack combined benchmark-leading reasoning models with Deep Research summaries and Agent Mode for autonomous multi-step tasks.The product stack described here formed the basis for the paper’s benchmark results.Source: §4.2.2 · pp. 12–13

4.2.3. The GPT 5 Consolidation.

• GPT 5 consolidated OpenAI’s selectable reasoning models into three tiers, so earlier o-series benchmark results no longer map directly to one selectable model.The new tiers require independent benchmarking.Source: §4.2.3 · pp. 13–14

4.2.4. xAI’s Grok Family.

• Grok 4 rose to benchmark leadership; Grok-4-Heavy parallelizes agents, but smaller context, mandatory web search, and documented safety vulnerabilities call for caution.For deep, uninterrupted analysis of large self-contained texts, the guide recommends Gemini 2.5 Pro.Source: §4.2.4 · p. 14

4.2.5. Specialized Auxiliary Tools.

• DeepL Write is presented as a specialized tool for polishing academic English, refining sentence structure, and improving linguistic precision.Source: §4.2.5 · p. 14

4.2.6. A Note on Coding.

• Manually executing AI-generated code preserves user control, supports debugging, and makes computational work reproducible; assistants can also guide novices through setup.Source: §4.2.6 · p. 15

4.2.7. Data Privacy and Security.

• Cloud AI may transmit sensitive research to third-party servers, where platform terms may allow training use, so researchers must follow data-security policies.Source: §4.2.7 · p. 15

5. Seven Ways of AI Usage in Mathematical Research

• The seven strategies are intended to remain useful as technology evolves, grounding practical advice in benchmarked models and translating collaboration principles into workflows.The paper uses Gemini 2.5 Pro, Grok 4, and legacy OpenAI o-series models as examples, while encouraging readers to map capabilities to current counterparts.Source: §5 · pp. 15–16

5.1.1. Research.

• For research ideation, broad exposure to scientific literature can help models suggest tractable questions, relevant connections, and examples, including beyond a researcher’s specialty.Source: §5.1.1 · p. 16, §5.1 · p. 16
• A model’s final-answer intuition can generate conjectures for rigorous testing, while best-of-n sampling and code support exploration of alternative paths and empirical checks.Promising ideas can then be investigated through targeted literature search and analysis.Source: §5.1.1 · p. 16

5.1.2. Teaching.

• For teaching, LLMs can assist curriculum development and individualized materials, while instructors iteratively generate diverse problems and refine selected ideas into syllabus-calibrated exercises.The instructor filters ideas for course scope and can use lower-temperature outputs for precise problems with model solutions.Source: §5.1.2 · pp. 16–17

5.2. Literature Search.

• For an initial topic overview, Grok 4 or OpenAI’s o3 can combine reasoning with real-time web search.Source: §5.2 · p. 17
• Google and OpenAI Deep Research iteratively search and analyze sources, producing cited summaries that users can verify; Google’s output tends to be more expansive than OpenAI’s.The described searches take 5–15 minutes.Source: §5.2 · p. 17
• Summary quality depends on retrieved sources, and models may present flawed information from processed articles as factual.Source: §5.2 · p. 17

5.3.1. Analysis of Provided Documents: Capabilities and Workflow.

• Gemini 2.5 Pro’s 1-million-token context supports analyzing multiple papers or textbooks, summarizing arguments, tracking notation, and finding theorems and cross-references.Unlike exact-string search, it can identify concepts across different variables or phrasings; the paper advises low creativity and disabling Google search.Source: §5.3.1 · pp. 17–18
• Interactive discussion of supplied material lets researchers clarify difficult passages, examine theorem implications, and compare authors’ approaches.The paper describes this as using the AI as a tireless sparring partner.Source: §5.3.1 · p. 18

5.3.2. A Critical Limitation: Internal Knowledge vs Provided Text.

• Internal recall can produce plausible but incorrect theorem citations, and persistent memory can make a model cling to an initial error even after the source is supplied.The model may then fail to find the relevant passage or hallucinate that the information is present.Source: §5.3.2 · p. 18
• Complex mathematical graphs and diagrams are often interpreted unreliably, so claims based on figures require independent verification.A plausible-sounding interpretation does not establish that the model has accurately read the figure.Source: §5.3.2 · p. 18

5.4. Interdisciplinarity.

• LLMs can translate research papers between languages and concepts between fields, helping experts communicate across differences in technical language and notation.They can rephrase concepts for non-experts or clarify discipline-specific terminology.Source: §5.4 · pp. 18–19
• AI can make unfamiliar areas more accessible, but validating complex ideas still requires expert knowledge and collaboration with human specialists.The paper emphasizes critical verification when exploring a new domain with AI.Source: §5.4 · p. 18
• LLMs may suggest analogies between mathematical fields that could transfer proof techniques, with a second model offering a first layer of verification.The paper frames the multi-model check as a way to cross-check the plausibility of a suggested connection.Source: §5.4 · p. 19
• AI can translate theoretical problems into code for computational exploration and conjecture formation, and can interface with CAS it has often encountered in training.The workflow uses empirical results to form new conjectures.Source: §5.4 · p. 19

5.5. Mathematical Reasoning.

• Reasoning quality is model-dependent, so serious mathematical work requires selecting a proven high-performing model and may benefit from combining models’ different strengths.The paper notes that models can combine known arguments in novel ways to construct proofs for unfamiliar statements.Source: §5.5 · p. 19

5.5.1. The Interactive Workflow and Principles of Verification.

• Researchers must critically check every AI-generated proof or calculation and learn each model’s biases; persistent memory may require a fresh chat to avoid repeating an early error.This workflow applies the Copilot, Not the Pilot Principle and the Experimental Mindset.Source: §5.5.1 · p. 19
• A second model can review and refine a proof sketch, but models may also be overskeptical and flag correct steps as potential mistakes.The paper presents multi-model peer review as a possible response to self-critique blindness.Source: §5.5.1 · p. 19
• Best-of-n sampling can increase the chance of finding a correct proof, while translating a logical step into code offers computational verification.Multiple solution paths also provide different strategies to consider.Source: §5.5.1 · p. 20

5.5.2. Reasoning as a Tool for Exploration and Diagnosis.

• AI dialogue can help researchers test lines of reasoning and diagnose weaknesses or missing links in a partial proof.The paper casts the model as a sparring partner for exploring and understanding arguments.Source: §5.5.2 · p. 20

5.6.1. The AI as a Personal Research Sparring Partner.

• An always-available, non-judgmental AI sparring partner can give researchers, especially at smaller institutions, a space to discuss doubts and test unconventional ideas.It can also support exploration of calculation results and analysis of literature.Source: §5.6.1 · p. 20

5.6.2. Enhancing Human and Interdisciplinary Collaboration.

• Free tools can give researchers at different institutions comparable computational support, while a trusted model can assess technical disagreements as a neutral third party.The paper points to top models’ near-human accuracy as evaluators.Source: §5.6.2 · p. 20

5.6.3. Implications for Teaching and Learning.

• AI can help instructors create routine exercises and give students one-on-one tutoring and space to explore concepts at their own pace.For instructors, automating routine exercises can free time for mentoring or seminars.Source: §5.6.3 · p. 20
• Because students may struggle to distinguish correct reasoning from plausible falsehoods, educational use calls for reliable, fact-grounded models and mindful oversight.The paper says safe and effective use depends on a critical approach from both students and instructors.Source: §5.6.3 · pp. 20–21

5.6.4. Warning: The Non-Human Nature of the AI Partner.

• Fluent, convincing output is not necessarily correct, and a flawed conclusion can resurface because LLMs may be reluctant to admit failure.This behavior is particularly noticeable in models with large, persistent memory contexts.Source: §5.6.4 · p. 21

5.7. Writing.

• AI can assist across the writing phase, from structuring initial drafts to preparing the final manuscript for dissemination.Source: §5.7 · p. 21

5.7.1. Structuring and Refining the Core Argument.

• Writing tools can help transform rough notes, partial proofs, and intuitive ideas into a coherent draft, then check consistency across a lengthy manuscript.The paper also describes LLMs as useful for refining language, while warning that precise prompting is needed to avoid changing mathematical content.Source: §5.7.1 · p. 21

5.7.2. Auxiliary Tasks and Ethical Considerations.

• LLMs can help find references, maintain bibliography formatting, draft abstract versions, and adapt technical writing for different audiences.Source: §5.7.2 · p. 21
• Authors must verify and internalize AI-generated text before submission, and mentioning tools in acknowledgements is becoming a transparency best practice.Source: §5.7.2 · p. 21

• The section raises authorship, plagiarism, and researcher responsibility as unresolved questions when AI generates a key proof step.Source: §5.8 · p. 22
• The paper treats an LLM as an interactive instrument, not an independent author, with the human retaining ownership and responsibility for verification and contextualization.It compares using an LLM to using a computer algebra system and argues that oversight may remain necessary because some mathematical problems resist purely algorithmic resolution.Source: §5.8 · p. 22
• Responsible AI use is framed as a transparent continuation of scientific tool adoption, conditional on rigorous human verification of the final work.The paper says clear community standards for transparent acknowledgment will be essential as these tools become more integrated into research.Source: §5.8 · p. 22

• AI's current and foreseeable role for working mathematicians is primarily augmentation, with humans retaining strategic direction, critical judgment, and intellectual ownership.Source: §6 · p. 22
• Affordable state-of-the-art models perform strongly on pre-Olympiad problems, while exceptionally difficult Olympiads mark their current frontier; Gemini Deep Think's premium-tier gold-medal success suggests a rapidly maturing field.Source: §6 · pp. 22–23
• The Seven Ways of AI Usage offer a durable workflow framework centered on rigorous, independent verification of AI-produced arguments.The paper says critical oversight can empower mathematicians to explore ambitious questions, test cross-field connections, and verify otherwise unexplored reasoning.Source: §6 · p. 23
• Future research practice may integrate natural-language models with formal proof systems and specialized agents, requiring mathematicians to develop human-AI interaction skills.The paper identifies strategic prompting, critical output evaluation, and ethical navigation as skills for future training.Source: §6 · p. 23