MATH · IN · MODELS
Why this map exists

How reproducible is the mathematics found inside trained models?

Deep networks trained on very different data and objectives keep converging on a small vocabulary of mathematical structures — linear feature directions, low-dimensional manifolds, belief-state simplices, concept lattices, attractors, collapse geometries. The evidence is spread across machine learning, computational neuroscience and statistical physics and described in incompatible language, so how much of it holds up cannot be answered from the literature as it stands.

The map is the instrument for answering it. Every entry names a structure, names the model it was found in, points at the paper that reported it, and is read to primary source before it is written.

700papers
703findings
816models
331families
249methods
66theory nodes
The question

Reproducibility and equivalence.

The question splits in two, and each half decomposes into things a corpus can actually count. Every axis below is computed from the records, never asserted by a curator.

Reproducibility — did it happen more than once?

  • The same structure reported in more than one paper.
  • Reported by more than one research group — papers that share no author.
  • The overall frequency with which a structure appears in the corpus at all.

Equivalence — does it hold across kinds of model?

  • The same structure across modalities — language, vision, audio, protein, control.
  • Across model families, trained independently by different labs.
  • Across architecture classes — transformer, CNN, RNN, diffusion, GNN.

Two by-products. The corpus is a source of hypotheses about where structures come from — the empty cells show where to look. And it is an instrument for analysing the field: which methods dominate, which structures are the mastodons, whether models are linear or linear methods are simply used more often.

The object

The map as an object.

A domain knowledge base in the form of a graph of claims, decomposed into structures, models, methods, families and hypotheses.

A continuously updated map of empirically reported mathematical structure in trained models.

The rule

What gets into the map.

A paper enters the corpus if and only if all five conditions hold. They are checked in order, and the first failure ends the assessment.

01
Object
A trained model.
The claim is about the activations, weights or gradients of a model that was actually trained — any modality, any architecture, any scale. Analytic toy models and synthetic point clouds do not count as the object.
02
Claim
A specific mathematical structure.
A shape, an operator, an order, or a measurable structural property of the representation — something that can be named in mathematics and, in principle, falsified. A benchmark score, an accuracy delta, or a ranking of which concept is encoded more strongly is not a structure.
03
Evidence
An original measurement or intervention, reported in that paper.
The paper itself must contain the evidence. It has to satisfy at least one of the three qualifying criteria (a), (b) or (c) below. Citing someone else’s measurement does not make a paper an entry; it makes it context.
04
Discovery
The structure is found, not built in.
If the geometry is forced by the architecture or the objective, the observation is rejected.
05
Verification
Read to primary source.
The reference must resolve to a real paper whose title matches, and the full text must have been read before the record is written, with the exact quantitative claim and its conditions extracted. No entry is ever created from an abstract.

Qualifying evidence: criteria (a), (b), (c)

Slot 03 requires at least one of three kinds of evidence. Together they define what counts as a measurement of structure, and the boundaries between them are what keeps the corpus from drifting.

(a)Shape measurement
A quantitative measurement of a geometric shape or property on the activations or weights of a real trained model.
Typically: a circle recovered for days of the week; a torus for modular arithmetic; a simplex of belief states; the dimension of a linear concept subspace.
(b)Geometry-linked causal intervention
A causal intervention whose handle is a geometric object, producing a measurable behavioural or representational change.
Typically: steering along a direction; ablating a subspace; a projector applied at a layer; a rank-one weight edit — weight- and gradient-space interventions count.
(c)Structural characterisation
A measured structural, spectral or dynamical characterisation that is not reducible to one static shape and is not causal.
Typically: eigen- and covariance spectra of embeddings; intrinsic-dimension and curvature profiles across depth; input-space partition geometry; neural- and dimensional-collapse; latent metrics and geodesics; how geometry evolves during training; convergence between independently trained models.

Standing exclusions

Each of these was decided once, on a real case, and is not reopened per paper. The reason is recorded so the boundary can be argued with rather than guessed at.

Relative encoding strengthA claim that X is encoded more strongly than Y is about magnitude, not about shape.
Method papers with no structural claimA technique that is never used to establish a structure inside a real model is not a finding.
Imposed geometryShape that follows from the architecture or the loss by construction.
Behaviour and benchmarksCapability, robustness, pruning-survival and evaluation results with no representational claim attached.
Theory with no measurementPurely analytic results about what a network could represent. These can appear as context on a theory page; they do not become paper records.
Position and survey papersNo original experiments, no entry. They may be cited on an existing node as background.

Two things that look like exclusions and are not. A paper whose full text cannot be obtained is deferred, never rejected on its abstract. And a negative or critical result — a failed replication, a demonstration that a reported structure is an artefact — is not rejected at all: it is recorded as a caveat on the structure it bears on, because that is exactly the evidence a map like this exists to keep.

Three outcomes, and only three

Every screened candidate ends in exactly one of these, dated and logged in public. There is no silent middle.

Added
Full record: a paper file, at least one observation, and links to the structure, method, model and family nodes it touches. New nodes only where genuinely warranted — the default is to reuse an existing node. Must pass the content validator before it is marked done.
Rejected
Zero corpus files, but a dated one-line reason in the public log. Rejections are recorded, not silently dropped — the log is part of the evidence.
Deferred
The paper is left untouched and a specific open question is logged: full text unobtainable, geometry ambiguous between two existing nodes, unclear whether a new node is distinct. Deferral is a normal outcome, not a failure. Nothing is fabricated to force a decision.
Metrics on the papers themselves

How far it has been reproduced.

Computed from the authors of the papers and from each model's architecture class, domain and family. The unit here is a node, not a finding: every count below is out of the 66 structures and hypotheses that have at least one paper behind them.

axisbreadthreplication
papers with no shared author31
architecture classes3829
domains3328
model families4830

Breadth — two or more values turned up at all, even if one study supplied them. Replication — the values came from papers with no shared author. Only the second is called replication.

Read the denominator: 35 of these 66 nodes rest on a single paper, so the ceiling for every row above is 31.

Process

How a paper becomes an entry.

Four stages, run separately and logged separately, so that discovery bias, reading errors and classification errors stay distinguishable from one another.

1
Discovery
Sweep arXiv, OpenAlex and the ACL Anthology for candidates and tag each against criteria (a), (b), (c) from title and abstract. Candidates are numbered in a single continuous public ledger, so nothing is screened twice and nothing disappears.
2
Verification
Confirm the identifier resolves and the title matches — the guard against hallucinated references — then read the primary source in full and extract the exact claim, the model, the layer or site, the metric and the conditions.
3
Classification
Assign the finding a mathematical role plus independent facets, following a written rule set, and grade its evidence status. Role first, shape second: what the object does mathematically decides the type; curvature, topology, linearity and specification are facets.
4
Audit
A separate pass re-reads each structure node for over-claims — the standing rule being that strength is not shape — and reclassifies or demotes where the write-up outran the evidence.
1047candidates screened
700papers included
703findings extracted
66structures and hypotheses
Limits

What this map does not yet do.

These are the places a reviewer will press, so they are named here first — each one paired with what is being done about it.

  • Coverage is skewed. The discovery sweep runs mostly on arXiv, and arXiv is where recent machine learning lives; non-arXiv venues and older work are under-represented. A separate back-fill track for 2000s–2020 results and non-arXiv venues.
  • Replication is tracked, not tested. The map records when two papers report the same structure. Nobody here re-runs the experiment, and a paper may also ship code the map does not yet point at. A layer of links to executable replications.
  • Curation is single-reader. Inclusion decisions run through one pipeline with one person's audit pass, not two independent screeners with a measured agreement rate. A human-verification loop with a published inter-rater agreement.
  • Thin on theory. The structure pages record what was measured, but rarely build the intuition or give an exercise, so the map informs without teaching. A track of worked problems and intuitions.

One schema change is also outstanding: how the evidence was obtained — criteria (a), (b), (c) — is decided during screening but not stored on the finding, the way the Gene Ontology stores evidence codes. Two grades were retired for a related reason: canonical, which graded consensus in the field rather than evidence in the corpus, and replicated, where a hand-set grade duplicated, worse and with drift, what the corpus computes for itself.

Access

The corpus is not public yet.

The corpus is plain text in Git — 2 500+ YAML and Markdown files with a schema and a validator — but the repository is closed while the schema is still moving. Every record is readable through this site: each structure, method, model family and paper has a page carrying the same fields the files do.

The method is written down here so that it can be checked. If a finding is wrong, or the inclusion rule was applied incorrectly, that is a defect worth reporting.

Suggestions and structural feedback
sadsobr7@gmail.com
  • The classification. A structure filed under the wrong role or type; two entries that are the same object under different names; a class the tree has no slot for.
  • The inclusion rule. A paper that passes the five slots and is missing, or one that is here and should not be.
  • The replication scheme. The four axes, the thresholds, the author-disjointness test — cases where they read the corpus wrongly.
  • A misread paper. A claim attributed to a paper that the paper does not make.
  • Missing work. Especially pre-2020, non-arXiv, and results outside language models.

Built and curated by Sabrina Sadiekh.