Research and development initiative · Feedback and collaboration welcome

Mithohne

AI-Assisted Mathematical Understanding and Knowledge Development

No Myth. No Y. Then Know Why.

About the name Mithohne

The name joins two words: MITH and ohne.

It begins with a transformation: replace the Y in MYTH with I, and MYTH becomes MITH.

The Y sounds like “why”, while I represents us — the people who remain in control of mathematics and seek to understand it.

M Y T H M I T H

“Why”: the question of mathematical understanding.

Us: the mathematicians who direct, control, and seek to understand the work.

MYTH becomes MITH. The Y that asked “why” is replaced by I, which stands for us.

The name ends in German. Ohne means “without”.

MITH

No myth, and the people who ask “why” stay part of the process.

OHNE

German for “without”: no blind trust, whether a claim comes from a person or from a machine.

Read together, the name states the intention. AI can assist mathematical discovery and verification, but people remain in control of how mathematical knowledge is understood, organized, and developed.

Mithohne is an AI-assisted mathematical research initiative focused on understanding, verifying, organizing, and extending mathematical knowledge.

As AI becomes increasingly capable of producing mathematical results, we believe there is an important opportunity to improve how mathematical knowledge is preserved, explained, connected, and reused.

Our goal is to develop tools and methods that support mathematical understanding while making reliable knowledge more accessible and useful for future research.

AI
Processes and connects mathematical knowledge.
Humans
Guide, evaluate, and develop mathematical understanding.

Explore the research capabilities Applications and research directions

Why this matters

Mathematics has developed an enormous body of knowledge, but it is distributed across specialized areas, papers, books, and mathematical languages. Researchers can only master part of it. Useful abstractions and relationships between fields may therefore be difficult to recognize, explain, or reuse.

AI is also changing how mathematical work is produced. An AI system may deliver a result or proof sketch without preserving the useful mathematical representations, reductions, connections, assumptions, and dependencies identified along the way in a form others can inspect and reuse. Those ideas may remain scattered across temporary conversations and documents, requiring later researchers or AI agents to reconstruct them.

The same AI technology creates an opportunity to address this problem. We want to investigate systems in which AI agents actively process mathematical work, explicitly construct reusable knowledge records, and connect them to existing literature and verification tools. The goal is not merely to produce more theorems, longer proofs, or larger formal libraries. It is to make mathematical development cumulative through reliable, connected, and shared understanding.

The core idea: two outputs from mathematical work

An AI-assisted mathematical investigation could yield two complementary outputs. The first is the one we habitually keep. The second is the one Mithohne is about.

Output 1

A mathematical result

An answer, theorem, example, counterexample, or proposed argument addressing the immediate question.

Output 2

Reusable mathematical knowledge

Explicit records of useful abstractions, reformulations, mathematical-object mappings, assumptions, dependencies, supporting sources, explanations, and verification status.

The second output is the distinctive research goal of Mithohne.

Two outputs from one investigation. The result answers the question; the knowledge records what was learned on the way there, in a form other people and later AI agents can inspect and reuse.

The reusable knowledge from an investigation may include:

Instead of preserving only a final answer or an informal conversation, Mithohne aims to preserve this information explicitly, so that useful mathematical connections do not stay buried in individual AI conversations or research outputs. The system would aim to keep the mathematical bridges that others can understand, check, and build upon.

This does not mean extracting a model's private internal reasoning, or assuming that a post-hoc narrative faithfully reconstructs it. The proposed AI workflow must deliberately create external, mathematically meaningful intermediate artifacts and independently check important claims.

From AI results to reusable mathematical knowledge

The distinction above is easy to state and harder to picture. This section follows one small investigation through the proposed workflow — from a question, to an answer, to a record, to a later problem that reuses it. The mathematics is deliberately elementary; the point is the form of the output rather than its difficulty.

Illustrative, not implemented. The record below is a hand-written sketch of the artifact a Mithohne workflow would aim to produce, not a live database entry. The identity it records is standard and well known; it is used to show the shape of the workflow, not presented as a new AI discovery.

  1. Step 1 · A mathematical question

    A researcher asks about two classical objects

    How are random walks on graphs related to graph Laplacians?

    Take a finite undirected graph $G=(V,E)$ with no isolated vertices. Let $A$ be the adjacency matrix, $D$ the diagonal degree matrix, $P = D^{-1}A$ the simple random-walk transition matrix, and $L = D - A$ the combinatorial Laplacian.

  2. Step 2 · The AI produces a result

    The answer is an exact identity between two operators

    A conventional assistant can answer the question and stop here:

    $$L = D(I-P).$$

    The identity holds because $D(I-P) = D - DD^{-1}A = D - A = L$. The answer is correct and useful. When the conversation ends, however, the connection it established usually ends with it.

  3. Step 3 · Mithohne preserves the connection

    The answer is accompanied by a structured record

    In addition to the answer, the workflow constructs an explicit mathematical record: an inspectable object with an identity, stated assumptions, a derivation, evidence, and a review status.

    Mathematical record · concept example

    Graph Laplacian ↔ random walk

    Illustrative record Not a live database entry

    Mathematical objects
    Graph Laplacian $L = D - A$; simple random walk with transition matrix $P = D^{-1}A$
    Exact transformation
    $L = D(I-P)$
    Assumptions
    Finite undirected graph; every vertex has positive degree, so $D^{-1}$ exists.
    Abstraction
    Represent graph dynamics using linear operators, which makes spectral and algebraic questions available.
    Derivation
    $D(I-P) = D - DD^{-1}A = D - A = L$
    Evidence
    Direct algebraic derivation, checked step by step. No computational test and no formal proof is claimed.
    References
    Precise source links to be added when identified and checked against the exact statement and assumptions.
    Provenance
    Constructed during the investigation above. An external artifact, not a transcript or reconstruction of private model reasoning.
    Status
    Illustrative record. Not reviewed, not published, not a live entry.
  4. Step 4 · A later investigation reuses the record

    A second question retrieves and specializes the connection

    What is the relationship between the eigenvalues of the random-walk transition matrix and the graph Laplacian for a regular graph?

    Mithohne retrieves the earlier record and reuses both parts of it: the transformation $L = D(I-P)$ and the requirement that vertex degrees be positive. The new question adds one further assumption — that the graph is $d$-regular with $d > 0$ — so $D = dI$ and the recorded transformation specializes to

    $$L = d(I-P).$$

    Therefore, if $Pv = \mu v$, then $Lv = d(I-P)v = d(1-\mu)v$: the vector $v$ is an eigenvector of both operators, and the eigenvalues are related by

    $$\lambda = d(1-\mu).$$

    The second investigation did not have to rediscover the basic correspondence. It retrieved a connection established earlier, checked the assumptions attached to it, and added the new hypothesis.

One investigation, two outputs, and a later reuse. Step 3 is the record that Mithohne would keep; step 4 shows a later investigation retrieving it and adding a regularity assumption.

The assumptions are doing real work. The eigenvalue relation $\lambda = d(1-\mu)$ uses $d$-regularity through $D = dI$. For a general graph the identity $L = D(I-P)$ still holds, but $D$ is not a scalar multiple of the identity, and the eigenvalues of $L$ and $P$ are not related by this formula. A record that carries its hypotheses can be reused soundly; a record that drops them cannot.

How this differs from an ordinary AI assistant

Both workflows begin in the same place: a researcher asks a mathematical question and an AI assistant produces a solution. The difference is what survives the conversation.

Ordinary AI interaction Proposed Mithohne workflow
The user asks a mathematical question. The user asks the same mathematical question.
The AI produces a solution. The AI produces a solution.
Useful abstractions may appear in the explanation. Useful abstractions are additionally recorded explicitly.
Connections remain primarily in the conversation. Connections become persistent mathematical records.
Future work may need to reconstruct the connection. Future work can retrieve, check, and reuse the connection.
Corrections depend on individual conversations. Records can be reviewed, corrected, and maintained.

The distinction is not about remembering more text. Existing AI systems can already save conversations and retrieve past messages. The proposed contribution is an explicit, mathematically structured representation of reusable knowledge — objects, transformations, assumptions, evidence, and review status — that a later person or agent can inspect and build on.

This comparison describes a proposed workflow. It is not a claim that current assistants lack memory or retrieval, nor that any system already produces records of this kind at scale. In both columns the researcher remains the one who decides what matters and what to trust; the proposed workflow is meant to give that judgment more to work with.

Four complementary research capabilities

Mithohne's work is organized around four capabilities. They are complementary rather than separate products: verification limits what can be claimed, organization keeps claims connected and traceable, explanation makes them understandable, and discovery proposes new connections for the other three to examine, record, and check.

Capability 1

Mathematical verification

Connect mathematical claims with formal verification tools, including Lean, while distinguishing formal correctness from the interpretation of informal mathematical statements. Verification is one kind of evidence among several: independent review, source checking, and computational tests each establish something different.

Capability 2

Knowledge organization

Use structured, evidence-backed mathematical records to preserve definitions, statements, dependencies, citations, and relationships. Records are designed to outlast any particular model or conversation, so that later researchers and AI agents can inspect, correct, and build on them.

Capability 3

Mathematical understanding

Develop explanations that help mathematicians understand important constructions, proof mechanisms, assumptions, and conceptual connections at different levels, from intuition and motivation to precise statements and technical detail.

Capability 4

Mathematical discovery

Explore how AI can identify useful abstractions, discover connections between mathematical areas, and suggest potential generalizations or new research directions. Exploratory suggestions are kept clearly separate from established results and are evaluated with researchers.

Each capability is a research goal. None of them is a claim that a complete system of this kind already exists. Together they are meant to extend a researcher's reach without removing the researcher from the process.

Applications and research directions

Mithohne is the overarching initiative. Individual applications may have their own names and specific purposes. The four directions below describe intended areas of work, not released products.

First application · current engineering priority

Verified Proof Atlas

An interactive environment connecting mathematical papers, Lean formalizations, dependencies, explanations, and verification evidence.

Proposed application

Research Lab

A proposed environment for exploring modifications to mathematical arguments, alternative assumptions, and new mathematical possibilities.

Longer-term direction

Formalization Studio

A longer-term direction for AI-assisted mathematical formalization.

Broader research direction

AI-Native Mathematical Knowledge Processing

A broader research direction exploring how AI can extract, organize, and preserve mathematical abstractions and connections across different sources. This work may eventually support additional applications.

An initial research opportunity: existing mathematical resources

Existing resources such as Wikipedia, mathematical research literature, and formal libraries already contain extensive mathematical knowledge.

Mithohne is interested in exploring how AI could process these materials to identify underlying mathematical structures, make connections between different areas explicit, and preserve them in reusable forms.

Such connections would be accompanied by source references, assumptions, explanations, and appropriate verification or review information.

This is a proposed research direction, not an implemented feature.

How these parts fit together

The intended interaction is best understood as a loop rather than a pipeline. In outline:

  1. Researchers set objectives, while source materials and AI-assisted investigations supply mathematical statements, problems, and candidate results.
  2. AI agents perform the main knowledge-processing work: proposing abstractions and transformations, identifying concepts and connections, and constructing explicit intermediate records.
  3. Each investigation may produce both a mathematical answer and a separate reusable knowledge artifact carrying assumptions, dependencies, citations, conceptual explanations, and uncertainty labels.
  4. Independent tools, precise source checks, and targeted human mathematical review establish the evidential status of important claims. The system must never equate agreement between language models with proof.
  5. Reviewed records enter a persistent mathematical knowledge structure, which supports human-readable pages, graph exploration, explanations, and future AI tasks.
  6. Human researchers assess significance, guide research priorities, correct errors, and provide feedback that improves subsequent AI knowledge processing.

This is a conceptual architecture, not a claim that a functioning system has already been implemented.

What Mithohne could become

Our longer-term vision is an AI-native, human-directed, evidence-checked mathematical knowledge platform. It could combine the readability of an encyclopedia, the explicit links of a mathematical knowledge graph, and the exploratory abilities of an AI research assistant.

Interfaces

Readable mathematical pages Knowledge graph AI research assistant

Engine

AI agents that process sources, propose abstractions and connections, and construct explicit knowledge records

Persistent output

A structured mathematical knowledge base that outlasts any particular model or conversation

Verification tools Human direction and review
These are complementary views of one system, not separate products. The AI agents are the main knowledge-processing engine; the structured knowledge base is its persistent output; verification tools help establish reliability; and mathematicians guide, review, and interpret the work. A conceptual sketch of a direction, not an implemented architecture.

A wiki-like interface would be one way to view and contribute to the knowledge, not the primary innovation.

We imagine a platform where AI-assisted work can add well-described mathematical connections to a shared body of knowledge, rather than losing them when a chat ends. Proposed relations would remain clearly distinguished from source-supported, reviewed, or formally checked results.

This is a vision to investigate — not a claim that full autonomous mathematical knowledge curation has already been solved.

What needs to be built

The following is a dependency-informed list of work, not a schedule. Items may be revisited as the research evolves.

  1. Choose representative mathematical problems. Collect a small, carefully selected set of questions where navigating existing results or finding connections is genuinely difficult.
  2. Define the two-output knowledge-processing contract. Specify which reusable abstractions, transformations, connection records, explanations, and proof obligations an AI workflow should produce alongside an answer.
  3. Define a mathematical knowledge model. Specify how to identify objects, notation contexts, statements, hypotheses, source locations, versions, transformations, and relationship types; ensure records can outlast a model or chat session.
  4. Build a curated evidence base. Ingest permitted mathematical sources, normalize references, preserve provenance, and curate initial connections with mathematicians.
  5. Develop the AI knowledge-processing engine. Build an agent workflow that extracts and proposes mathematical knowledge, creates explicit intermediate artifacts, traces supporting material, and maintains candidate records for validation.
  6. Explore AI-assisted connection discovery. Generate candidate reductions, translations, and conceptual bridges, and record their assumptions and uncertainty.
  7. Develop layered checking. Validate source attribution and the semantics of proposed links; use appropriate independent mathematical reviews, tests, and selective formal verification.
  8. Develop explanations and navigation. Make statements, dependencies, conceptual connections, and explanations accessible at different levels of mathematical background.
  9. Evaluate with real mathematical tasks. Compare with ordinary search and general-purpose LLM interaction, checking correctness, usefulness, understanding, and time spent, as well as whether the extracted connections remain reusable — not just counts of extracted or formalized theorems.
  10. Support responsible contributions. Explore how the community might propose, review, correct, version, and reuse mathematical records.

How knowledge and evidence are recorded

Accurate references are central to Mithohne, but a traditional bibliography is not sufficient. A research system of this kind should distinguish at least the following relation types.

Relation Meaning
CITES A publication cites another material, perhaps for attribution or context.
SUPPORTS_CLAIM A precise source passage or statement supplies evidence for a specific claim.
DEPENDS_ON A proof or theorem logically relies on another mathematical result.
ABSTRACTS_AS / TRANSFORMS_TO A problem, object, or argument is deliberately represented in a mathematically useful alternative form, with explicit conditions.
REFORMULATES / EQUIVALENT_TO A claim or object is translated into another representation, under explicit assumptions.
GENERALIZES / SPECIAL_CASE_OF A mathematical relationship of scope or strength.
FORMALIZES A formal statement or proof corresponds to an informal mathematical claim.
CORRECTS / SUPERSEDES A later source changes, repairs, or replaces an earlier statement.
SUGGESTS_CONNECTION An exploratory, unverified relationship proposed by AI or a researcher.
PRODUCED_DURING A reusable mathematical artifact was explicitly constructed during an AI-assisted investigation. This is a provenance record, not a claim to reconstruct hidden model reasoning.

Each proposed relation should carry, when available: a source identifier (DOI, arXiv ID, or book edition), an exact locator (theorem, lemma, section, or page), the source version, the statement and its assumptions, a link to the underlying evidence, the method of checking, the identity of the reviewer or tool, the date of review, and an explicit status.

Several kinds of evidence are involved, and the design keeps them separate:

These categories are not interchangeable: a bibliographic reference is not a proof; a computational test is not a general theorem; an AI model agreeing with itself is not independent verification; and a checked formal proof still needs to be aligned with the intended informal statement. AI-generated explanations and proposed relationships are not presented as independently verified mathematical facts. A link is never shown as established merely because a model generated it, and references that cannot be confirmed are marked as unconfirmed rather than guessed.

On this static proposal page it is enough to state these distinctions clearly. No citation database or knowledge graph is operated here; the design above describes what a future implementation would need to represent.

Open questions

Principles

This division of responsibility does not mean that every mathematician must personally verify every result. It means designing the whole system to strengthen human mathematical agency and shared understanding.

This research direction was previously presented under the working title MathBridge. That page now lives at /mithohne. The old address redirects here permanently.

Help shape Mithohne

Mithohne is an early research and development initiative. We welcome mathematical examples, critical feedback, suggestions for existing resources, and potential collaboration from mathematicians, AI researchers, and formal-methods practitioners. If you are interested in the problem of connecting and organizing mathematical knowledge, we would be happy to hear your perspective.

The initiative is deliberately human-directed: AI processes and connects mathematical knowledge, while researchers decide what matters, what to trust, and what to pursue.

Get in touch