Co-Mathematician: Collaborative Math Practice

• Co-Mathematician is a role that involves collaborative, error-prone, and exploratory mathematical work rather than only presenting finished proofs.
• It integrates human intuition with AI capabilities in literature search, computational exploration, and theorem development, as seen in platforms like MathOverflow and Polymath.
• Agentic systems such as AIM and AI co-mathematician orchestrate human oversight with automated proof assistance, marking a new paradigm for mathematical research workflows.

A co-mathematician is a collaborator in the production of mathematical knowledge whose role is not limited to routine calculation, formal proof checking, or isolated question answering. In contemporary research discourse, the term is most naturally situated at the intersection of mathematical practice, sociotechnical collaboration, and AI-assisted discovery. It denotes a participant—human, machine, or composite sociotechnical system—that can contribute to the exploratory, error-prone, concept-forming, and socially negotiated processes by which mathematics is actually developed, rather than only to the polished frontstage of finished proofs (Martin et al., 2013). Recent AI work sharpens this into an explicit research paradigm: the co-mathematician is an agentic or interactive partner for open-ended mathematical work, supporting ideation, literature search, computational exploration, theorem proving, and theory building under human oversight (Zheng et al., 7 May 2026, Liu et al., 30 Oct 2025). At the same time, whether such a collaborator is recognized as a “mathematician” depends on broader disputes about identity, function, qualification, and power within the mathematical community (Buckmire et al., 2023).

1. Conceptual emergence

The contemporary concept of a co-mathematician emerges from a rejection of the image of advanced mathematics as a solitary thinker producing a finished proof for later judgment. Martin and Pease argue that mathematics is at an “inflexion point,” because crowdsourcing, symbolic computation, proof checking, archives, and online platforms make visible and technologically amplify the backstage of mathematical work: the informal, speculative, error-tolerant process in which examples, conjectures, proof ideas, and revised concepts are developed (Martin et al., 2013). Within this perspective, deduction and proof remain central, but they are no longer treated as the whole of mathematical production. Analogy, creativity, concept revision, error handling, and social interaction become constitutive features of research practice.

This shift is closely linked to the notion of the “social machine.” In this usage, a mathematical social machine is a combination of people, computers, and mathematical archives functioning as a single problem-solving entity. A co-mathematician, therefore, need not be a single autonomous AI agent. It may instead be a composite arrangement of people, computational services, formal libraries, archives, discussion platforms, and governance structures that jointly create and apply mathematics (Martin et al., 2013). This broader framing distinguishes the co-mathematician from narrower notions such as theorem prover, computer algebra system, or search engine.

A dialogical variant of the same idea appears in work that stages mathematics and physics as mutually illuminating rather than strictly separated enterprises. In Michel Planat’s dialogue between “Phys” and “Math,” abstract arithmetic, group theory, modular forms, dessins d’enfants, and moonshine are treated as structures that recur on both sides of the mathematics–physics interface (Planat, 2015). This is not an AI framework, but it extends the co-mathematical idea beyond human–machine collaboration to collaborative reasoning across disciplinary modes of thought.

2. Empirical anatomy of collaborative mathematics

The empirical basis for the co-mathematician concept is supplied most clearly by analyses of MathOverflow and Polymath. In a sampled study of MathOverflow, about 90% of sampled questions received an answer that moved knowledge forward, with 78% yielding reasonable answers to the original question and another 12% yielding partial or helpful responses. The same study found literature citations in 56% of responses, explicit examples in 34%, and acknowledged and corrected error in 37% of discussions (Martin et al., 2013). These figures matter not merely as platform metrics, but because they expose a characteristic interactional norm: mathematical progress often proceeds through partial answers, reference retrieval, example generation, polite correction, and rapid repair of misunderstandings.

Polymath makes the same point even more strongly by exposing collaborative proof discovery on open problems. In Mini-Polymath 3, an Olympiad geometry problem was solved in 74 minutes by 27 participants through 174 comments. The coding reported 33% examples, 20% conjectures, 14% proof contributions, 10% concept work, and 23% other material, including clarifications, cross-references, and social glue (Martin et al., 2013). Only a minority of comments directly advanced the final proof. Much of the mathematical work consisted in exploring cases, trying interpretations, generating sub-conjectures, and stabilizing a productive environment.

These observations have direct implications for the meaning of co-mathematician. A genuine collaborator in research mathematics must be able to operate in settings where claims are provisional, partial answers are valuable, and corrected error is productive. It must also handle semantic drift and concept negotiation. Martin and Pease connect these phenomena to Lakatosian themes such as monster-barring, monster-adjusting, exception-barring, lemma-incorporation, and proofs and refutations, using examples around Euler’s formula

$V - E + F = 2$

and the interpretation-dependent status of controversial polyhedra (Martin et al., 2013). In this sense, a co-mathematician must engage mathematics as practice, not only as formal derivation.

3. Agentic architectures and research workflows

Recent AI systems operationalize the co-mathematician idea in two distinct but related ways. One line treats AI as a research partner inside a tightly supervised human–AI proof process. Another treats it as an agentic workbench that organizes long-horizon mathematical workflows.

System	Stated role	Characteristic mechanisms
AIM	“research partner rather than a mere problem solver”	iterative decomposition, targeted human interventions, validation of intermediate results
AI co-mathematician	“agentic workbench for mathematical research”	asynchronous, stateful workspace, hierarchical delegation, reviewed LaTeX reports

In the AIM case study on homogenization theory, the authors describe an “AI Mathematician as a Partner” paradigm centered on sustained human–AI co-reasoning rather than autonomous solution of a monolithic problem. The proof architecture was decomposed into six subproblems: two-scale expansion; cell problem and homogenized equation; existence and uniqueness; ellipticity of the effective operator; error estimation and control; and regularity of the cell problem. Humans retained responsibility for decomposition of the global problem, theorem-condition discipline, diagnosis of unsound steps, and formal correctness. AIM contributed by suggesting intermediate lemmas, identifying the Babuška–Brezzi route for well-posedness, producing a relatively complete ellipticity argument, generating candidate error-estimate decompositions, surfacing the hidden need for a strong regularity statement, and eventually succeeding with a Schauder-theory-based route under sparse theoretical guidance. The collaboration produced a complete and verifiable proof, including the quantitative bound

$$\|u_{\varepsilon}-u_{\lim}\|_{H^1(\Omega)} \le C \varepsilon^{1/2}\|u_0\|_{H^2(\Omega)}.$$

(Liu et al., 30 Oct 2025)

The AI co-mathematician system generalizes this into an environment for open-ended research. It is described as a stateful, multi-agent workspace in which a user-facing project coordinator agent clarifies the problem, formulates goals, launches workstreams, and integrates outputs; workstream coordinators manage subgoals; and specialized sub-agents perform literature review, coding, proof attempts, or search. The system is explicitly asynchronous rather than a single conversational chatbot, and all project information is stored in a shared workspace and file system. Its design principles include embracing mathematics beyond proofs; supporting iterative refinement of intent; producing native mathematical artifacts; enabling asynchronous interaction and flexible steering; managing cognitive load through progressive disclosure; tracking, managing, and communicating uncertainty; and preserving failed explorations. The primary outputs are reviewed LaTeX reports with margin notes and internal links into the workspace (Zheng et al., 7 May 2026).

These systems instantiate a recurrent division of labor already anticipated in earlier discussions of mathematical social machines. Machines are strong at routine calculation, exhaustive search, formal checking, storage, retrieval, and large-scale coordination; humans remain stronger at contextual interpretation, concept formation, analogical transfer, informal explanation, theorem-condition discipline, and social negotiation (Martin et al., 2013). The co-mathematician paradigm does not abolish that asymmetry. It organizes it.

4. Recognition, identity, and power

The term co-mathematician also raises the prior question of what counts as a mathematician at all. The framework developed in “On definitions of ‘mathematician’” distinguishes three axes: identity, function, and qualification. A function-based definition asks what a person does; a qualification-based definition asks what normative bar they have passed; and an identity-based definition asks who the person or others say they are (Buckmire et al., 2023). This framework is directly relevant because collaborative mathematical roles often satisfy some of these criteria but not others.

Under a broad function-based definition, “A mathematician is a person who uses mathematical concepts, tools or techniques to study and solve problems.” Under a narrow function-based definition, “A mathematician is a person who proves theorems using proof techniques.” Between them is the hybrid definition: “A mathematician is a person who, as part of their daily work, employs mathematical techniques and tools to solve mathematical and other problems” (Buckmire et al., 2023). The co-mathematician concept aligns most naturally with the broad and hybrid forms, because many collaborators in modern mathematics—applied mathematicians, data scientists, statisticians, teachers, computational scientists, interdisciplinary researchers, and AI systems embedded in research workflows—contribute mathematically without centering proof-writing.

The same paper shows that identity-based and qualification-based definitions complicate recognition further. Internal identity-based definition centers self-identification; external identity-based definition relies on collective recognition; narrow qualification-based definition maps mathematician status to institutional credentials such as a Ph.D. in mathematics and a current research position. The authors also define “compuesto definitions” by union and intersection, noting that the narrow intersection “could result in the empty set” (Buckmire et al., 2023). This is highly pertinent to co-mathematics: if one requires a collaborator to satisfy every function, qualification, and identity criterion simultaneously, many genuine contributors to mathematical work disappear from view.

The paper treats this not as a semantic triviality but as a question of power. Narrow definitions can support clear membership boundaries and data collection on structural inequity, yet they can also reinforce the implication that “real mathematicians are...” only those who satisfy elite, proof-centered, or prestige-centered criteria. Broad definitions can widen belonging, but can also “artificially erase” barriers if they flatten differences in authorship, legitimacy, resources, and voice (Buckmire et al., 2023). A co-mathematician, in this sense, is not merely a functional role in a workflow; it is also a contested status within a community.

5. Foundational scope and the “core mathematician”

One recent proposal approaches the co-mathematician problem from foundations. Mumford and Friedman argue that the right default universe for mainstream mathematics is not unrestricted ZFC as a final ontology, but a real-based inner universe tailored to “core mathematics,” understood as mathematics centered on $N$, $R$, and structures built from them. Their proposed canonical model is

$M(R)=L_{\alpha^{\mathrm{ZF--}}}(R),$

and, in the paper’s implications for AI-assisted mathematics, this is presented as a suitable default stance for a collaborator intended to work on ordinary mathematics (Mumford et al., 5 May 2026).

On this view, a co-mathematician should treat most standard mathematical assertions as ranging over natural numbers, reals, separable or Polish spaces, Borel and measurable structures, and objects coded by countable or real parameters. The paper explicitly associates “almost all core math” with mathematics already formalizable in second-order arithmetic $Z_2$, including standard analysis, Banach spaces, PDEs, Borel sets via codes, second countable spaces, manifolds, probability, and dynamical systems, while using the real-based set theory to permit actual sets of reals and sets of sets of reals (Mumford et al., 5 May 2026).

This proposal does not claim to settle all of set theory. Rather, it offers a heuristic boundary for co-mathematical reasoning. Claims involving arbitrary subsets of $R$, well-orderings of $R$, nonconstructive choice of representatives, high cardinal arithmetic, strong regularity principles for all projective sets, or global category-of-all-categories constructions are treated as potentially outside core mathematics and therefore potentially model-dependent (Mumford et al., 5 May 2026). For a co-mathematician, the practical significance is that foundational sensitivity becomes part of the system’s epistemic responsibilities: not every mathematical statement should be handled as if it were equally canonical.

6. Performance, limits, and disputed questions

The strongest evidence for present-day co-mathematician systems is methodological rather than metaphysical. The AI co-mathematician paper reports early use on open-ended research tasks, including a reported resolution of Kourovka Notebook Problem 21.10, proofs for two conjectures about Stirling coefficients that are “currently in detailed human review,” and a proof-related subproblem in Hamiltonian systems. On a benchmark of 100 unleaked, research-level mathematics problems with code-checkable answers, the system significantly outperformed single-call Gemini baselines. In Epoch AI’s blind evaluation on FrontierMath Tier 4, it correctly solved 23 out of 48 non-sample problems, yielding 48% accuracy, which the paper describes as a new high score among evaluated AI systems (Zheng et al., 7 May 2026).

The AIM case study offers a different form of evidence. Its importance lies less in benchmark accuracy than in the fact that the AI contributed at the level of proof organization, subgoal generation, dependency discovery, and theory-conditioned derivation on a research-level homogenization problem. The authors explicitly describe AIM not as an infallible theorem prover or black-box solver, but as an exploratory collaborator whose intermediate traces—including false starts—were often informative search trajectories (Liu et al., 30 Oct 2025). A co-mathematician, on this model, is valuable not because it is always right, but because it can supply cumulative quanta of progress.

Several misconceptions follow from this literature. One is that a co-mathematician is simply an autonomous replacement for a human mathematician. The recent AI papers reject that framing: the strongest workflows arise when humans assign tasks selectively, enforce theorem-condition discipline, diagnose failure modes, and preserve ownership of rigor, while AI contributes search breadth, persistence, and rapid derivational adaptation (Liu et al., 30 Oct 2025, Zheng et al., 7 May 2026). Another misconception is that the term is socially neutral. The definitional framework around mathematician status shows that collaborative participation does not automatically entail recognition, and that broadening the label can either support inclusion or mask persistent inequalities (Buckmire et al., 2023).

The current systems also have clear limitations. In the AI co-mathematician workbench, all proofs are still purely informal; the paper warns about “Reviewer-Pleasing Bias (False Consensus),” non-termination or “death spirals,” and the danger that polished LaTeX can overstate rigor (Zheng et al., 7 May 2026). In the AIM case study, the verification mechanism called Pessimistic Rational Verification is explicitly said not to be fully reliable, and there is no formal proof-assistant certificate for the end-to-end derivation (Liu et al., 30 Oct 2025). At the foundational level, the real-based proposal for core mathematics depends on Freiling’s axiom, which the paper itself notes is philosophically controversial and insufficient to settle all regularity questions for projective sets (Mumford et al., 5 May 2026).

Taken together, these strands define the co-mathematician as a technical, social, and epistemic role rather than a single device or ontology. It is technical because it integrates search, proof, computation, literature, and stateful workflow; social because participation and recognition are mediated by communal norms, attribution, and power; and epistemic because it must manage uncertainty, partial progress, and foundational scope rather than merely output finished proofs. In current research, the term most often names a collaborator embedded in a mathematical social machine: a participant in mathematics in the making.