Papers in three-dimensional topology often contain sentences like "after replacing $M$ by a finite cover, we may assume…". From that point on the letter $M$ names a different space. The new space has a different volume, a different fundamental group, and possibly a different genus, yet it has the same symbol, the same type ("a closed hyperbolic 3-manifold"), and often the same sentence frames as before. Current systems for math-symbol disambiguation, which decide what each occurrence of a letter refers to, have never been tested on this pattern. Its main signal suggests they will fail on it. On the existing datasets, the best predictor of a symbol's meaning is its most recent previous occurrence. In covering-space arguments the most recent occurrence of $M$ is exactly the wrong referent once the author has silently swapped spaces.
We propose to build and test a disambiguator for this setting. It reads a geometric topology paper and keeps a running record of every space the argument introduces: the original manifold, its finite covers, the pieces obtained by cutting along tori, Dehn fillings, and the tetrahedra of a decomposition. The record also stores how these spaces map to one another. Each symbol occurrence is then resolved to a node in this record rather than to a dictionary concept. Because the spaces in small examples can be computed exactly from census manifolds with SnapPy and Regina, we can also generate synthetic arguments whose referents are known by construction. Numeric claims in those arguments, such as volumes that scale with covering degree, give a check on any proposed grounding that can be verified by machine.
Core Claim
The claim has two parts. First, a "renaming under substitution" event is common in arguments about virtual properties. In such an event a symbol keeps its letter and its type but is rebound to a new object linked to the old one by a map such as a covering. This produces the long-range disambiguation errors that the cascade heuristic cannot fix. Second, a resolver that tracks a typed graph of spaces and maps, and treats substitution phrases as explicit rebinding operations, fixes these errors. Long-context language models and position-plus-affix classifiers do not.
The work is new in three ways. No existing benchmark covers geometric topology. Existing annotation schemes merge "$M$ before the cover" and "$M$ after the cover" into one concept. And the referents can be checked against computed invariants from ideal triangulations.
Grounding in Prior Work
Three lines of work show the gap.
Asakura, Miyao and Aizawa annotated document-wide coreference among identifiers (Building Dataset for Grounding of Formulae). They found 2.09 meanings per identifier per paper and found that 89.5% of meaning changes happen inside a single section. In a follow-up, Asakura and Miyao showed that position and formula affixes beat text context and that models fail on "revived topics" (What Is Needed for Intra-document Disambiguation of Math Identifiers?). Both studies use NLP, astronomy, economics and physics papers. Their concept-level labels would not separate a manifold from its cover when the author calls both $M$. We reuse their annotation tool (MioGatto) and their position and affix features. We change the label space to referents in a graph of spaces and maps.
Dev, Asakura and Sætre showed that GPT-4 can build concept dictionaries and coreference clusters (An Approach to Co-reference Resolution and Formula Grounding for Mathematical Identifiers Using Large Language Models). This is the strongest general-purpose competitor and the main threat to the proposal.
STEM-POM shows that language models struggle with symbol polymorphism even with full-manuscript context. Its paper set excludes pure mathematics.
On the topology side, the corpus that motivates the work is the virtual-domination series. Sun proved that every closed hyperbolic 3-manifold has a finite cover mapping onto every closed 3-manifold with nonzero degree (Virtual domination of 3–manifolds). Liu and Sun strengthened this to degree one (Virtual 1-domination of 3-manifolds). Sun extended it to positive simplicial volume (II) and to cusped manifolds with degree 8 (III). Lackenby's work on Heegaard gradients in towers of covers (Heegaard splittings, the virtually Haken conjecture and Property (τ)) adds whole sequences $M_i$ of covers, which makes the referent structure dense.
The only machine-learning work on triangulations, Learning 3-Manifold Triangulations, trains classifiers on isomorphism-signature strings, which are compact text encodings of how tetrahedra are glued. It never touches covers or prose, and it reports that BERT did worse than simple feedforward networks. We use its software stack (Regina, SnapPy, the 4,815-manifold cusped census) only to generate ground truth, not as a learning target.
Müller and Kaliszyk's translation of informal LaTeX into a typed formal language (Disambiguating Symbolic Expressions in Informal Documents) constrains the design. Even after pre-training, only 59.6% of outputs passed type checking, on elementary material. Full autoformalization of topology is therefore not a realistic route.
Research Program
The first component is an annotated corpus. We will annotate roughly 30 arXiv papers from the virtual-domination, covering-tower, triangulation-complexity and subgroup-distortion literature cited above, using MioGatto. The new annotation layer marks two kinds of event. Space-introduction events are "let $\tilde M$ be a degree-$d$ cover", "cut along the JSJ tori", "Dehn fill" and "triangulate". Rebinding events are "replacing $M$ by…" and "we still denote by…". Each occurrence of a space-valued symbol is linked to a node in a per-paper directed graph whose edges are labelled maps (covering of degree $d$, inclusion, filling, quotient). Annotators will be topology graduate students. Agreement will be reported as Cohen's κ, so it can be compared with the 0.75–0.94 range in the grounding dataset.
The second component is a synthetic generator with verified referents. We will enumerate low-degree covers of census manifolds with SnapPy and compose template-free arguments by prompting a language model with a scripted sequence of operations. Each argument is then rendered with deliberate symbol reuse. Ground-truth referents come from the script. Each node also has computed invariants (volume from the ideal triangulation, first homology, number of cusps), so numeric claims in the text check the grounding. For example, the volume of a degree-$d$ cover is $d$ times the base volume, because hyperbolic volume is multiplicative under finite covers.
The third component is the resolver. A tagger detects introduction and rebinding events. A small state machine maintains the space graph. A pointer classifier, which uses the Asakura–Miyao position and affix features plus a language-model encoding of the local sentence, scores each occurrence against the current graph nodes. An optional consistency step rejects groundings that contradict a stated invariant.
Baselines
The naive reference point is the cascade rule: assign each occurrence the referent of the most recent prior occurrence of the same symbol. Asakura and Miyao identify this signal as dominant. It costs nothing to run. Any method that does not beat it on rebinding events has failed.
The strongest feature-based method is the Asakura–Miyao classifier using position and affix features. It reports 85% accuracy (κ = 0.73) on their 40-paper set (What Is Needed for Intra-document Disambiguation of Math Identifiers?). We will retrain it on our training papers with the same split and tuning budget as our resolver. We will report it both on our new labels and on concept-level labels, so the comparison is fair.
The strongest general method is long-context language-model coreference in the style of Dev et al. That paper reports a GPT-4 CoNLL score of 80.15 on the full set and 87.92 on a subset (An Approach to Co-reference Resolution and Formula Grounding). We will run a current long-context model on the full paper with the same instructions plus a description of the space-graph label scheme and a few worked examples. We will also run an open-weight model under LoRA fine-tuning on the same training papers. Beating this baseline means a statistically significant gain in referent accuracy on rebinding-affected spans, measured with a paired bootstrap over papers. A gain only on easy spans does not count.
The ablation that isolates the key ingredient is our resolver with the rebinding operation disabled, so substitution phrases are treated as ordinary mentions. A second ablation removes the invariant-consistency step. If the full model's advantage disappears without rebinding, the explicit state is doing the work.
Why not just…?
Why not just prompt a long-context language model with the whole paper? This is the real threat. GPT-4 already reached a CoNLL score above 80 on general papers, and rebinding phrases are explicit in the text. The evidence against it is indirect. STEM-POM's best full-manuscript accuracy is 68.5%, and topology papers chain many rebinding events across pages. Our experiments settle the question directly by reporting accuracy separately on spans after the first, second and third rebinding of a symbol. If the language model matches the resolver there, we drop the resolver and publish the corpus and the synthetic generator as a benchmark.
Why not just use existing concept-level annotation and treat this as ordinary coreference? Under concept labels, $M$ and its cover may be marked as the same concept "the hyperbolic manifold". Every downstream use breaks under that merge. Retrieving the statement "vol$(M) \ge \dots$", checking a degree claim, or formalizing a lemma all depend on which space is meant, because volume, homology rank and Heegaard genus all change under covers. We will measure how many rebinding events concept-level annotators collapse when they label the same papers. That number shows whether the distinction is real or whether we are inventing it.
Why not just autoformalize into a proof assistant and let the type checker resolve symbols? The best reported translator passed type checking on 59.6% of outputs even on elementary algebra and computer-science notes (Disambiguating Symbolic Expressions). The file contains no formal library for finite covers or hyperbolic invariants. Our invariant-consistency check is a cheap, partial substitute that works today, and its ablation shows how much this kind of semantic checking buys.
Proposed Experiments
The main experiment trains on synthetic arguments plus about 20 annotated papers and tests on about 10 held-out papers. It reports referent accuracy overall, on rebinding-affected spans, and on revived-topic spans, against the baselines above.
The first decisive check is whether rebinding-affected spans exist in meaningful numbers at all. If they make up under 5% of space-symbol occurrences in the corpus, the motivating hypothesis is wrong.
The second decisive check is whether cascade accuracy drops sharply on these spans relative to the roughly 85% general figure. If it does not drop, existing methods already handle the phenomenon.
The third check is transfer. A resolver trained only on synthetic SnapPy-grounded arguments should beat the cascade rule on real papers. This shows that computed referents give useful supervision. For calibration we will also report the resolver's numbers on the concept-level 40-paper dataset, so we can see whether specializing to topology costs accuracy elsewhere. A reported number for disambiguation on geometric topology text does not exist in the file: no reported result found on Lacuna for symbol disambiguation in geometric topology papers.
References
- Building Dataset for Grounding of Formulae - Annotating Coreference Relations Among Math Identifiers
- What Is Needed for Intra-document Disambiguation of Math Identifiers?
- MioGatto: A Math Identifier-oriented Grounding Annotation Tool
- An Approach to Co-reference Resolution and Formula Grounding for Mathematical Identifiers Using Large Language Models
- STEM-POM: Evaluating Language Models Math-Symbol Reasoning in Document Parsing
- Virtual domination of 3–manifolds
- Virtual 1-domination of 3-manifolds
- Virtual domination of 3‐manifolds II
- Virtual domination of 3-manifolds, III
- Heegaard splittings, the virtually Haken conjecture and Property (τ)
- Learning 3-manifold triangulations
- Disambiguating Symbolic Expressions in Informal Documents