Public scorecard

How often the link is right

GlossMath claims one thing: that the symbol under your cursor is linked to the place this paper defined it. This page is where we keep that claim honest. Every number below is either measured or marked n/a. A metric we have not measured is never shown as a zero.

Last evaluated 2026-08-11 10:31:39 UTC

Papers evaluated

checked link by link

Links checked

424

individual symbol→definition links

Links correct

95.8% 406/424

the link points at the real definition

Abstention rate

77.7% 3104/3997

3104/3997 across the papers listed

Flag rate

0.0% 0/893

reader reports, over the links we showed

What we count as correct

A link is correct when the place it jumps to is where this paper actually introduces that symbol: the sentence that names it, the equation that sets it, or the line that gives its type. A link that lands on a later re-use of the symbol counts as wrong, even though the symbol is there. A link that lands in the right paragraph but on the wrong symbol counts as wrong.

Abstention is a feature, and we would rather it went up than down

When we cannot find where a symbol was defined, we list it under “not linked” with the reason and we do not link it. That is the abstention rate above. We could push it to zero tomorrow by guessing, and the correctness number would fall. The number we optimise is correctness among the links we do show.

What we link, and what we hold back

The model grades every definition it finds. It calls a passage high when the paper states the meaning outright at that spot, and medium when the meaning is clear from context but never actually said. We link only the first kind. 422 located passages across the papers below are held back for being the second kind.

That is a coverage decision and it costs real links. We made it because “medium” is defined, in the same prompt, as a meaning the paper never states outright — and this product’s entire claim is that a link points at where the paper defines the symbol. A link we would have to argue for is not one of those.

What a person said

Not run yet. The numbers above come from an independent model, which is a real check with a real blind spot: it reads the same passage our own pass read and answers the same question about it. Neither has tried to read the paper. A 400-question human evaluation is built and waiting on a labeller; nothing here will claim a human number until it has one.

Per paper

Paper	Symbols	Linked	Checked	Correct
Attention Is All You Need `arXiv:1706.03762`	54	18	18	100.0% 18/18
Denoising Diffusion Probabilistic Models `arXiv:2006.11239`	56	10	10	100.0% 10/10
Auto-Encoding Variational Bayes `arXiv:1312.6114`	90	33	34	97.1% 33/34
Neural Ordinary Differential Equations `arXiv:1806.07366`	69	10	10	100.0% 10/10
BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding `arXiv:1810.04805`	25	7	7	100.0% 7/7
An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale `arXiv:2010.11929`	58	13	n/a	n/a
Deep Residual Learning for Image Recognition `arXiv:1512.03385`	9	4	n/a	n/a
Generative Adversarial Networks `arXiv:1406.2661`	59	6	n/a	n/a
Language Models are Few-Shot Learners `arXiv:2005.14165`	99	8	n/a	n/a
Distilling the Knowledge in a Neural Network `arXiv:1503.02531`	39	15	n/a	n/a
Neural Machine Translation by Jointly Learning to Align and Translate `arXiv:1409.0473`	103	54	n/a	n/a
Batch Normalization: Accelerating Deep Network Training by Reducing Internal Covariate Shift `arXiv:1502.03167`	65	16	n/a	n/a
The entropy formula for the Ricci flow and its geometric applications `arXiv:math/0211159`	210	48	n/a	n/a
Calculation of prompt diphoton production cross sections at Tevatron and LHC energies `arXiv:0704.0001`	282	58	n/a	n/a
Model Evaluation, Model Selection, and Algorithm Selection in Machine Learning `arXiv:1811.12808`	106	42	n/a	n/a
Learning Transferable Visual Models From Natural Language Supervision `arXiv:2103.00020`	31	2	n/a	n/a
Layer Normalization `arXiv:1607.06450`	95	27	n/a	n/a
Efficient Estimation of Word Representations in Vector Space `arXiv:1301.3781`	33	8	n/a	n/a
Deep contextualized word representations `arXiv:1802.05365`	44	16	n/a	n/a
Densely Connected Convolutional Networks `arXiv:1608.06993`	16	9	n/a	n/a
Playing Atari with Deep Reinforcement Learning `arXiv:1312.5602`	40	11	n/a	n/a
Chain-of-Thought Prompting Elicits Reasoning in Large Language Models `arXiv:2201.11903`	16	1	n/a	n/a
Neural Discrete Representation Learning `arXiv:1711.00937`	33	11	n/a	n/a
An Algorithm for Optimal Partitioning of Data on an Interval `arXiv:math/0309285`	32	10	n/a	n/a
Universal Model Routing for Efficient LLM Inference `arXiv:2502.08773`	190	66	n/a	n/a
U-Net: Convolutional Networks for Biomedical Image Segmentation `arXiv:1505.04597`	22	6	n/a	n/a
Sentence-BERT: Sentence Embeddings using Siamese BERT-Networks `arXiv:1908.10084`	19	11	n/a	n/a
Model-Agnostic Meta-Learning for Fast Adaptation of Deep Networks `arXiv:1703.03400`	36	14	n/a	n/a
Differential Transformer `arXiv:2410.05258`	74	22	n/a	n/a
The Era of 1-bit LLMs: All Large Language Models are in 1.58 Bits `arXiv:2402.17764`	13	0	n/a	n/a
Improving neural networks by preventing co-adaptation of feature detectors `arXiv:1207.0580`	33	4	n/a	n/a
Very Deep Convolutional Networks for Large-Scale Image Recognition `arXiv:1409.1556`	7	4	n/a	n/a
Unsupervised Representation Learning with Deep Convolutional Generative Adversarial Networks `arXiv:1511.06434`	3	1	n/a	n/a
A sparsity result for the Dynamical Mordell-Lang Conjecture in positive characteristic `arXiv:2012.13711`	73	22	22	100.0% 22/22
Projected Stochastic Gradient Langevin Algorithms for Constrained Sampling and Non-Convex Learning `arXiv:2012.12137`	198	43	49	87.8% 43/49
Effective approximation of heat flow evolution of the Riemann ξ function, and a new upper bound for the de Bruijn-Newman constant `arXiv:1904.12438`	285	47	49	95.9% 47/49
Supermembranes and domain walls in 𝒩=1, D=4 SYM `arXiv:1905.02743`	269	47	48	97.9% 47/48
Dirichlet forms and polymer models based on stable processes `arXiv:1905.00181`	295	90	91	98.9% 90/91
Localizing virtual cycles for Donaldson-Thomas invariants of Calabi-Yau 4-folds `arXiv:2012.13167`	402	79	86	91.9% 79/86
Extremal eigenvalues of critical Erdős-Rényi graphs `arXiv:1905.03243`	414	n/a	n/a	n/a
Total across 40 papers	3997	893	424	95.8% 406/424

“Symbols” is every symbol we found. “Linked” is the subset we were confident enough to link. “Checked” is how many of those links an independent model from a different family has judged, span by span, without being shown our gloss or our confidence. It also still counts every link that check rejected and we then took off the page, so cleaning the showcase cannot raise this number.

What we do not know yet

Whether accuracy holds outside the papers listed here. Everything measured so far is machine learning and adjacent maths, mostly recent arXiv preprints, and notation habits vary a lot by field.
How we do on papers with no LaTeX source, where we fall back to a rougher conversion and the symbol boundaries are less reliable.
Whether a definition we picked is the one a reader wanted. A symbol can be introduced loosely in the intro and pinned down properly three sections later. We prefer the pinned-down one, and we have not measured how often readers agree.

Found a link that is wrong? Every symbol card has a quiet wrong? button. Those reports come straight here as the flag rate.