CAP: Multi-Domain Definitions & Applications

• CAP is a term with multiple, unrelated definitions, spanning mathematical constructs such as cap sets, cap products, and cap amplitudes, as well as machine learning frameworks like data contamination detection and visual classification.
• In mathematics, CAP covers areas including additive combinatorics, algebraic topology, and convex geometry, offering insights through constructs like cap sets with nontrivial arithmetic progression restrictions and cap bodies as convex unions.
• In machine learning, CAP refers to innovative approaches such as consistency amplification for LLM contamination detection and prompt generation for copyright auditing, emphasizing model evaluation and data integrity.

Searching arXiv for papers relevant to the multiple meanings of “CAP” referenced in the provided data. In contemporary arXiv usage, CAP is not a single technical notion but a family of domain-specific terms spanning random matrix theory, additive combinatorics, algebraic topology, convex geometry, large-language-model evaluation, copyright auditing of generative models, and fine-grained visual classification. In some settings, “cap” is a mathematical noun, as in cap sets, cap product, cap amplitudes, and cap bodies; in others, CAP is an acronym, as in Consistency Amplification–based Data Contamination Detection, Copyright Audit via Prompts generation, and Context-aware Attentional Pooling. The underlying concepts are unrelated except lexically, and the cited literature treats them as distinct research programs rather than variants of a common framework (Okuyama, 4 Sep 2025, Zhao et al., 2024, Gallo et al., 2024, Tyrrell, 2022, Kable et al., 28 Apr 2026, Hanamura, 2016, Arman et al., 29 Oct 2025, Behera et al., 2021).

1. Disambiguation and domain structure

The mathematical uses of cap are older and structurally heterogeneous. In additive combinatorics, a cap set is a subset of $\mathbb{F}_3^n$ with no nontrivial solution to $x+y+z=0$; in the affine-geometry analogue over $\mathbb{F}_{2^n}$, the constraint becomes the absence of forbidden four-point additive relations (Tyrrell, 2022, Kable et al., 28 Apr 2026). In algebraic topology, the cap product is a bilinear operation pairing homology with cohomology, and Hanamura studies its supported form on Borel–Moore homology (Hanamura, 2016). In random matrix theory, the cap amplitude $\psi(b)$ is introduced as the expansion coefficient of the 1-form $y\,dx$ on the spectral curve of a one-matrix model (Okuyama, 4 Sep 2025). In convex geometry, a cap is the convex hull of the unit ball and one external point, while a cap body is any convex union of finitely many such caps (Arman et al., 29 Oct 2025).

The acronymic uses are specific to machine learning. CAP in LLM evaluation denotes Consistency Amplification–based Data Contamination Detection, a benchmark-level contamination detector based on a statistic called the Performance Consistency Ratio (Zhao et al., 2024). A different CAP denotes Copyright Audit via Prompts generation, a black-box method for testing whether a generative model has been trained on unauthorized data by learning prompts that induce reproduction of suspected samples (Gallo et al., 2024). In computer vision, Context-aware Attentional Pooling is a plug-in module for fine-grained classification that combines region extraction, bilinear pooling, attention, LSTM encoding, and differentiable clustering (Behera et al., 2021).

A common misconception is to treat these occurrences as terminological variations of a shared idea. The cited works do not support such a unification. The overlap is nominal; the mathematical objects, problem settings, and proof or evaluation methodologies are distinct.

2. Cap sets and cap bodies in discrete and convex geometry

In additive combinatorics, a cap set in $\mathbb{F}_3^n$ is a subset containing no nontrivial three-term arithmetic progression, equivalently no nontrivial solution of $x+y+z=0$ (Tyrrell, 2022). Tyrrell gives a new constructive lower bound by exhibiting a cap set in $\mathbb{F}_3^{56232}$ of cardinality

$$\binom{11}{7}^{141}\,\cdot\,6^{572}\,\cdot\,12^{572}\,\cdot\,112^{8800}\,\cdot\,37\,\cdot\,142,$$

and therefore establishing that for all sufficiently large $n$,

$x+y+z=0$0

(Tyrrell, 2022). The construction builds on Edel’s extended product construction, introduces recursively admissible sets and meta-extensions of admissible sets, and uses Boolean satisfiability with symmetry-breaking heuristics to generate admissible families such as $x+y+z=0$1, $x+y+z=0$2, and $x+y+z=0$3 (Tyrrell, 2022). The paper situates this lower bound against earlier constructive constants of Pellegrino, Calderbank–Fishburn, and Edel, while noting the gap to the Ellenberg–Gijswijt upper bound $x+y+z=0$4 (Tyrrell, 2022).

A related but structurally different line identifies cap sets as multiplicative subgroups of finite fields. Kable, Mills, and Wright show that the subgroup of $x+y+z=0$5 nonzero fourth powers in $x+y+z=0$6 is a cap set, and likewise the subgroup of $x+y+z=0$7 nonzero seventh powers in $x+y+z=0$8 is a cap set (Kable et al., 28 Apr 2026). These correspond to the card games SET and EvenQuads: in $x+y+z=0$9, the four cosets of $\mathbb{F}_{2^n}$0 partition $\mathbb{F}_{2^n}$1 into four disjoint maximal $\mathbb{F}_{2^n}$2-caps, while in $\mathbb{F}_{2^n}$3 the seven cosets of $\mathbb{F}_{2^n}$4 partition $\mathbb{F}_{2^n}$5 into seven disjoint maximal $\mathbb{F}_{2^n}$6-caps (Kable et al., 28 Apr 2026). The same paper proves that for $\mathbb{F}_{2^n}$7, the subgroup $\mathbb{F}_{2^n}$8 has no four distinct elements summing to zero, and that the $\mathbb{F}_{2^n}$9 cosets of $\psi(b)$0 partition $\psi(b)$1 into $\psi(b)$2-caps of size $\psi(b)$3 (Kable et al., 28 Apr 2026).

In convex geometry, the noun cap has a different meaning. Arman, Kaire, and Prymak define a cap as $\psi(b)$4 with $\psi(b)$5, and a cap body as any convex union of finitely many such caps (Arman et al., 29 Oct 2025). They prove Hadwiger’s covering conjecture for cap bodies in all dimensions: for the class $\psi(b)$6 of $\psi(b)$7-dimensional cap bodies, $\psi(b)$8 for all $\psi(b)$9 (Arman et al., 29 Oct 2025). For $y\,dx$0 the argument combines a probabilistic technique with reduction to linear programming performed with computer assistance; for $y\,dx$1 the paper gives an explicit illumination bound, and for $y\,dx$2 this bound is shown to be $y\,dx$3 (Arman et al., 29 Oct 2025). The shared word “cap” here refers to a convex-geometric primitive, not to arithmetic-progression-free sets.

3. Cap product in Borel–Moore homology

Hanamura studies the supported cap product on Borel–Moore homology and compares three models: sheaf-theoretic Borel–Moore homology, locally finite singular homology, and locally finite simplicial homology (Hanamura, 2016). For a locally finite, countable, finite-dimensional simplicial complex $y\,dx$4, the Borel–Moore homology is defined as

$y\,dx$5

where $y\,dx$6 consists of locally finite infinite $y\,dx$7-linear combinations of simplices (Hanamura, 2016). If $y\,dx$8 is locally compact Hausdorff and locally contractible and $y\,dx$9 is closed and locally contractible, the supported cap product induces a pairing

$\mathbb{F}_3^n$0

(Hanamura, 2016).

On chains, the operation is given by the Alexander–Whitney-type formula

$\mathbb{F}_3^n$1

for $\mathbb{F}_3^n$2 and $\mathbb{F}_3^n$3, with the standard compatibility

$\mathbb{F}_3^n$4

(Hanamura, 2016). The main theorem states that, for a locally finite, countable, finite-dimensional simplicial complex $\mathbb{F}_3^n$5 and subcomplex $\mathbb{F}_3^n$6, the sheaf-theoretic, singular, and simplicial supported cap products are all identified under the canonical isomorphisms

$\mathbb{F}_3^n$7

(Hanamura, 2016).

The significance of this comparison is methodological rather than terminological. It establishes that the supported cap product is independent of the chosen model and, in particular, independent of triangulation (Hanamura, 2016). The paper also extends the comparison to relative theories and shows compatibility with localization isomorphisms, which supports standard applications such as Poincaré duality for noncompact oriented manifolds (Hanamura, 2016).

4. Cap amplitudes in random matrix models

In the one-cut Hermitian matrix model, the entire topological expansion is built from the single 1-form $\mathbb{F}_3^n$8 on the classical spectral curve, and the paper “Cap amplitudes in random matrix models” introduces the cap amplitude $\mathbb{F}_3^n$9 as the coefficient with which $x+y+z=0$0 decomposes into Fourier modes on the uniformizing variable $x+y+z=0$1 (Okuyama, 4 Sep 2025). Using the Joukowsky map

$x+y+z=0$2

the decomposition is

$x+y+z=0$3

(Okuyama, 4 Sep 2025). Expanding around $x+y+z=0$4 or $x+y+z=0$5 determines the sequence $x+y+z=0$6, while the large-$x+y+z=0$7 resolvent normalization fixes $x+y+z=0$8 and vanishing at the branch points $x+y+z=0$9 yields the sum rules

$\mathbb{F}_3^{56232}$0

(Okuyama, 4 Sep 2025). For $\mathbb{F}_3^{56232}$1, the coefficients can equivalently be written as

$\mathbb{F}_3^{56232}$2

(Okuyama, 4 Sep 2025).

The central structural claim of the paper is that the Eynard–Orantin dilaton identity for $\mathbb{F}_3^{56232}$3 can be rewritten as a purely combinatorial gluing formula for the discrete volumes $\mathbb{F}_3^{56232}$4: $\mathbb{F}_3^{56232}$5 (Okuyama, 4 Sep 2025). In this interpretation, capping one of the $\mathbb{F}_3^{56232}$6 boundaries by gluing on the cap amplitude reduces the number of boundaries by one. The same mechanism produces the genus-$\mathbb{F}_3^{56232}$7 free energy from the one-boundary volume,

$\mathbb{F}_3^{56232}$8

(Okuyama, 4 Sep 2025).

The paper further states that once $\mathbb{F}_3^{56232}$9 is known, one may reconstruct the eigenvalue density $$\binom{11}{7}^{141}\,\cdot\,6^{572}\,\cdot\,12^{572}\,\cdot\,112^{8800}\,\cdot\,37\,\cdot\,142,$$0 via its Fourier series, recover the potential by

$$\binom{11}{7}^{141}\,\cdot\,6^{572}\,\cdot\,12^{572}\,\cdot\,112^{8800}\,\cdot\,37\,\cdot\,142,$$1

hence

$$\binom{11}{7}^{141}\,\cdot\,6^{572}\,\cdot\,12^{572}\,\cdot\,112^{8800}\,\cdot\,37\,\cdot\,142,$$2

and obtain the moments $$\binom{11}{7}^{141}\,\cdot\,6^{572}\,\cdot\,12^{572}\,\cdot\,112^{8800}\,\cdot\,37\,\cdot\,142,$$3 as linear combinations of $$\binom{11}{7}^{141}\,\cdot\,6^{572}\,\cdot\,12^{572}\,\cdot\,112^{8800}\,\cdot\,37\,\cdot\,142,$$4 (Okuyama, 4 Sep 2025). Because the local behavior of $$\binom{11}{7}^{141}\,\cdot\,6^{572}\,\cdot\,12^{572}\,\cdot\,112^{8800}\,\cdot\,37\,\cdot\,142,$$5 governs the full hierarchy $$\binom{11}{7}^{141}\,\cdot\,6^{572}\,\cdot\,12^{572}\,\cdot\,112^{8800}\,\cdot\,37\,\cdot\,142,$$6, the paper concludes that knowledge of $$\binom{11}{7}^{141}\,\cdot\,6^{572}\,\cdot\,12^{572}\,\cdot\,112^{8800}\,\cdot\,37\,\cdot\,142,$$7 completely fixes $$\binom{11}{7}^{141}\,\cdot\,6^{572}\,\cdot\,12^{572}\,\cdot\,112^{8800}\,\cdot\,37\,\cdot\,142,$$8 by topological recursion and therefore all free energies $$\binom{11}{7}^{141}\,\cdot\,6^{572}\,\cdot\,12^{572}\,\cdot\,112^{8800}\,\cdot\,37\,\cdot\,142,$$9 (Okuyama, 4 Sep 2025). The term “cap” here names a disc-like gluing building block in the genus expansion.

5. CAP as consistency amplification for LLM contamination detection

In large-language-model evaluation, CAP stands for Consistency Amplification–based Data Contamination Detection, a framework designed to determine whether an LLM has merely been fine-tuned on a benchmark’s training set or has memorized held-out test samples (Zhao et al., 2024). The method takes a training set $n$0 and test set $n$1, applies a consistency-preserving transformation $n$2 to each split, runs the model $n$3 on the original and modified data, and compares the resulting Performance Consistency Ratio values (Zhao et al., 2024). With

$n$4

the bounded PCR is defined as

$n$5

(Zhao et al., 2024). The decision variable is

$n$6

with positive values interpreted as fine-tuning and negative values as contamination (Zhao et al., 2024).

A key claim of the paper is that CAP is, to the authors’ knowledge, the first method to explicitly differentiate between fine-tuning and contamination, a distinction they regard as crucial for domain-specific models (Zhao et al., 2024). The method requires only inference access, a task metric $n$7, a consistency criterion $n$8, and a consistency-preserving transformation $n$9, so it applies in both white-box and black-box settings (Zhao et al., 2024).

The evaluation uses seven LLMs and four financial benchmarks: FinEval, FinQA, AlphaFin, and ECTSum (Zhao et al., 2024). The reported findings include strongly negative $x+y+z=0$00 on FinEval for Baichuan-13B and DISC-Fin-13B, approximately $x+y+z=0$01, later tied to overlap with the C-EVAL corpus; large positive $x+y+z=0$02 on FinQA for FinMA-Full and FinMA-NLP, approximately $x+y+z=0$03 to $x+y+z=0$04, matching known fine-tuning on FinQA’s training split; mild contamination signals on AlphaFin due to overlap with FinQA; and near-zero $x+y+z=0$05 on ECTSum for most LLMs, with LLaMA-8B only slightly positive and inconclusive (Zhao et al., 2024). The paper argues that composite benchmarks from multiple dataset sources are particularly prone to unintentional contamination (Zhao et al., 2024).

A recurring misconception in benchmark auditing is that any unusually high test performance must reflect contamination. CAP explicitly rejects that simplification: the sign test is intended to separate legitimate train-split exposure from leaked test-split exposure (Zhao et al., 2024). The paper nonetheless notes limitations, including benchmark-level rather than sample-level granularity and possible sensitivity of $x+y+z=0$06 and $x+y+z=0$07 scaling in very low-consistency regimes (Zhao et al., 2024).

6. CAP in generative-model auditing and fine-grained visual classification

A second machine-learning acronym uses the same letters but addresses a different problem. CAP: Detecting Unauthorized Data Usage in Generative Models via Prompt Generation defines Copyright Audit via Prompts generation as a black-box auditing framework for determining whether a generative model $x+y+z=0$08 has been trained on unauthorized data (Gallo et al., 2024). Given suspected copyrighted samples $x+y+z=0$09, CAP trains a prompt generator $x+y+z=0$10 so that, for a target sample $x+y+z=0$11, the prompt $x+y+z=0$12 steers $x+y+z=0$13 to regenerate $x+y+z=0$14 or a close approximation (Gallo et al., 2024). A violation is declared when

$x+y+z=0$15

where $x+y+z=0$16 and $x+y+z=0$17 (Gallo et al., 2024). The prompt generator is trained by minimizing

$x+y+z=0$18

over minibatches (Gallo et al., 2024). To accelerate training, the framework periodically fits a Generalized Pareto Distribution to the tail of per-sample errors and drops the top $x+y+z=0$19 worst-fitting samples when loss stagnates and at least one-third of $x+y+z=0$20 remains (Gallo et al., 2024).

The reported experiments use four IoT scenarios and two synthetic datasets, with both $x+y+z=0$21 and $x+y+z=0$22 implemented as standard Transformer encoder–decoder models with $x+y+z=0$23 layers each, $x+y+z=0$24 attention heads, and embedding dimension $x+y+z=0$25 (Gallo et al., 2024). On real datasets, Precision@5 reaches $x+y+z=0$26 on Pump Sensor, Elevator Failure, and Head Posture, while Electric Power Consumption is lower; AUC-Gain values range from about $x+y+z=0$27 to $x+y+z=0$28 depending on dataset and optimization setting (Gallo et al., 2024). The optimized procedure reduces $x+y+z=0$29’s wall-clock training time by roughly $x+y+z=0$30–$x+y+z=0$31 across datasets with minimal loss in detection performance (Gallo et al., 2024). On synthetic data, CAP attains AUC-Gain $x+y+z=0$32 on the non-overlapping dataset and $x+y+z=0$33 on Synthetic-Overlap, which the paper presents as evidence that heavy distribution overlap is a fundamental limitation (Gallo et al., 2024).

In computer vision, Context-aware Attentional Pooling is another unrelated CAP. The method is proposed as a plug-in module for fine-grained recognition, designed to capture subtle changes via sub-pixel gradients, attend informative integral regions, and encode the consistency between region informativeness and spatial structure (Behera et al., 2021). Starting from a backbone feature map $x+y+z=0$34, the model first applies self-attention,

$x+y+z=0$35

then extracts a hierarchy of integral regions, uses bilinear sampling to obtain sub-pixel-sensitive region features, computes context-aware region-to-region attention, feeds the ordered region summaries through an LSTM, and aggregates hidden states with a NetVLAD-style differentiable clustering before classification (Behera et al., 2021). The paper reports evaluation on eight fine-grained benchmarks and six backbone networks, with gains such as $x+y+z=0$36 versus $x+y+z=0$37 on Aircraft, $x+y+z=0$38 versus $x+y+z=0$39 on Food-101, $x+y+z=0$40 versus $x+y+z=0$41 on Cars, and $x+y+z=0$42 versus $x+y+z=0$43 on NABirds (Behera et al., 2021). It states that CAP significantly outperforms prior approaches on six datasets and is very competitive on the remaining two (Behera et al., 2021).

The juxtaposition of these two acronymic CAPs illustrates the instability of acronym-based terminology in contemporary ML literature. One CAP is an audit mechanism for unauthorized-data usage in generative models; the other is a feature-aggregation module for fine-grained visual classification. The acronym alone therefore does not identify a research object without immediate domain context.