GRAS: Polysemous Terms in Research

• GRAS is a polysemous term with distinct definitions in number theory, statistics, astrophysics, and machine learning.
• In algebraic number theory, GRAS refers to conjectures linking unit groups and ideal class groups, while in high-dimensional statistics it underpins efficient edge screening in graphical models.
• In astrophysics and fairness evaluations, GRAS describes gravitational anti-screening effects and serves as a benchmark for demographic bias in vision-language models.

GRAS denotes several unrelated constructs in contemporary research literature. In algebraic number theory, it appears in the Gras conjecture, which relates unit groups and ideal class groups in abelian extensions (Oukhaba et al., 2011). In high-dimensional statistics, the closely related acronym GRASS denotes graphical sure screening for Gaussian graphical models (Luo et al., 2014). In astrophysics, GRAS denotes GRavitational Anti-Screening, a dark-matter alternative formulated through vacuum-induced dipole effects (Penner, 9 Feb 2026). The string also appears in Multi-GraS, a multiplex graph model for extractive summarization (Jing et al., 2021), and in the GRAS benchmark for demographic bias in vision-LLMs, where it expands to Gender, Race, Age, and Skin Tone (Malik et al., 26 Aug 2025). The term is therefore intrinsically polysemous, and its meaning is fixed by disciplinary context.

1. Gras conjecture in algebraic number theory

In its classical number-theoretic usage, GRAS refers to the Gras conjecture, which predicts a precise relationship between the unit group and ideal class group of an abelian extension, measured using the so-called “index.” For a rational character $\chi$ of $G=\mathrm{Gal}(K/k)$, the conjecture is summarized in the form

$$[\mathbb{Z}[G]e_\chi:\ell_K(E_K)e_\chi]=|\mathrm{Cl}(K)[e_\chi]|.$$

The central objects are the ideal class group $\mathrm{Cl}(K)$), the unit group $\mathcal{O}_K^\times$, and distinguished unit subgroups generated from elliptic, Stark, or circular units, depending on the arithmetic setting (Oukhaba et al., 2011).

For abelian extensions of an imaginary quadratic field $k$, “On Gras conjecture for imaginary quadratic fields” proves the previously unresolved case where the prime $p$ divides the number of roots of unity in $k$ (Oukhaba et al., 2011). The paper extends Rubin’s methods, defines $E_K$ as the subgroup of $K^\times$ generated by the roots of unity in $G=\mathrm{Gal}(K/k)$0 and all possible norms of Stark units, and combines Euler-system constructions with elementary group-theoretic and index calculations via the generalized index of Sinnott. Its main theorem states that if $G=\mathrm{Gal}(K/k)$1 divides $G=\mathrm{Gal}(K/k)$2 and does not divide $G=\mathrm{Gal}(K/k)$3, and if $G=\mathrm{Gal}(K/k)$4 is a nontrivial irreducible $G=\mathrm{Gal}(K/k)$5-character of $G=\mathrm{Gal}(K/k)$6, then

$G=\mathrm{Gal}(K/k)$7

This completes the proof of the Gras conjecture for all primes in that setting (Oukhaba et al., 2011).

A function-field analogue is established in “The Gras conjecture in function fields by Euler systems” (Oukhaba et al., 2011). There the ambient objects are a global function field $G=\mathrm{Gal}(K/k)$8, a finite abelian extension $G=\mathrm{Gal}(K/k)$9, the unit group $$[\mathbb{Z}[G]e_\chi:\ell_K(E_K)e_\chi]=|\mathrm{Cl}(K)[e_\chi]|.$$0, the class group $$[\mathbb{Z}[G]e_\chi:\ell_K(E_K)e_\chi]=|\mathrm{Cl}(K)[e_\chi]|.$$1, and a Stark-unit subgroup $$[\mathbb{Z}[G]e_\chi:\ell_K(E_K)e_\chi]=|\mathrm{Cl}(K)[e_\chi]|.$$2. The proof uses Euler systems constructed from torsion points of sign-normalized Drinfeld modules and adapts the techniques of Thaine, Kolyvagin, and Rubin. Under the paper’s explicit hypotheses on $$[\mathbb{Z}[G]e_\chi:\ell_K(E_K)e_\chi]=|\mathrm{Cl}(K)[e_\chi]|.$$3 and $$[\mathbb{Z}[G]e_\chi:\ell_K(E_K)e_\chi]=|\mathrm{Cl}(K)[e_\chi]|.$$4, the main equality is

$$[\mathbb{Z}[G]e_\chi:\ell_K(E_K)e_\chi]=|\mathrm{Cl}(K)[e_\chi]|.$$5

This places Stark units in global function fields in the same structural role that elliptic and cyclotomic units occupy in number fields (Oukhaba et al., 2011).

A further extension appears in “Gauss Sums, Stickelberger’s Theorem, and the Gras Conjecture for Ray Class Groups” (All, 2015). For a real abelian number field $$[\mathbb{Z}[G]e_\chi:\ell_K(E_K)e_\chi]=|\mathrm{Cl}(K)[e_\chi]|.$$6, an odd prime $$[\mathbb{Z}[G]e_\chi:\ell_K(E_K)e_\chi]=|\mathrm{Cl}(K)[e_\chi]|.$$7 not dividing $$[\mathbb{Z}[G]e_\chi:\ell_K(E_K)e_\chi]=|\mathrm{Cl}(K)[e_\chi]|.$$8, the unit group

$$[\mathbb{Z}[G]e_\chi:\ell_K(E_K)e_\chi]=|\mathrm{Cl}(K)[e_\chi]|.$$9

the subgroup $\mathrm{Cl}(K)$0 of $\mathrm{Cl}(K)$1-circular units, and the ray class group $\mathrm{Cl}(K)$2, the paper proves a ray-class version of the conjecture:

$\mathrm{Cl}(K)$3

when the ramification index of $\mathrm{Cl}(K)$4 in $\mathrm{Cl}(K)$5 is less than $\mathrm{Cl}(K)$6 (All, 2015). It also constructs explicit Galois annihilators of $\mathrm{Cl}(K)$7 akin to the classical Stickelberger theorem. Across these variants, the common theme is the comparison of explicit unit subgroups with class-theoretic invariants via characterwise index formulas.

2. GRASS in Gaussian graphical models

In high-dimensional statistics, the near-homographic term GRASS stands for graphical sure screening (Luo et al., 2014). It is a very simple and computationally efficient screening procedure for recovering the structure of a Gaussian graphical model in the high-dimensional setting. Rather than estimating the precision matrix directly, GRASS thresholds the elements of the sample covariance or correlation matrix. With standardized variables and data matrix $\mathrm{Cl}(K)$8, the empirical correlation for features $\mathrm{Cl}(K)$9 is

$\mathcal{O}_K^\times$0

and the estimated edge set is

$\mathcal{O}_K^\times$1

An edge is declared between $\mathcal{O}_K^\times$2 and $\mathcal{O}_K^\times$3 if and only if $\mathcal{O}_K^\times$4 (Luo et al., 2014).

Its principal theoretical result is the sure screening property: with very high probability, the GRASS estimated edge set contains the true edge set. Under the condition that, for a true edge $\mathcal{O}_K^\times$5, the population covariance satisfies $\mathcal{O}_K^\times$6 with $\mathcal{O}_K^\times$7, and when $\mathcal{O}_K^\times$8 with $\mathcal{O}_K^\times$9, choosing

$k$0

yields

$k$1

The paper also shows that, under an additional eigenvalue-growth condition, the size of the estimated neighborhood is controlled, which makes the screening step computationally useful in ultra-high dimensions (Luo et al., 2014).

The method includes an analytical threshold for expected false positive rate control. If $k$2 is the maximum number of acceptable false positives, one sets

$k$3

and under mild assumptions the expected false positive rate satisfies

$k$4

The computational cost is $k$5, in contrast with the $k$6 cost stated for the graphical lasso and neighborhood selection. The paper further notes that the connected components produced by thresholding the sample covariance at level $k$7 correspond exactly to those recovered by the graphical lasso with $k$8 (Luo et al., 2014). Simulation studies and a gene-expression example show that GRASS performs competitively with more complex methods, especially when the sparsity pattern of the correlation matrix mirrors that of the precision matrix (Luo et al., 2014).

3. GRAS as gravitational anti-screening

In astrophysics, GRAS denotes GRavitational Anti-Screening, a theory presented as an alternative to dark matter and to MOND’s AQUAL formulation (Penner, 9 Feb 2026). Its core physical picture is that the gravitational field of baryonic matter induces mass dipole moments in quantum vacuum fluctuations, producing an anti-screening effect that boosts the effective observed gravitational mass. The field equation is written as

$k$9

with dipole density

$p$0

and effective dipole contribution

$p$1

Under spherical symmetry this reduces to

$p$2

or equivalently

$p$3

A specific interpolating function cited as well matched to the galactic radial acceleration relationship is

$p$4

The paper explicitly states that GRAS and AQUAL are mathematically identical after identifying $p$5 (Penner, 9 Feb 2026).

The same manuscript argues that both AQUAL and standard GRAS fail to explain the bulk of the missing mass in galaxy clusters, even though they are in excellent agreement with observations of galaxies, including the galactic RAR (Penner, 9 Feb 2026). The proposed modification introduces sensitivity to the radial variation of the field through

$p$6

and replaces the standard spherical equation by

$p$7

The corresponding dipole interpretation becomes

$p$8

The modified field is solved with an outer boundary condition chosen to match the galactic RAR regime (Penner, 9 Feb 2026).

According to the paper, the modified equation has just a single free parameter, the universal acceleration scale $p$9, and is applied to a sample of six relatively bound galaxy clusters (Penner, 9 Feb 2026). It is reported to fit the observationally inferred dynamical mass profiles far better than the unmodified equation, while retaining galactic behavior and satisfying solar-system constraints. A central interpretation is that the full baryonic density profile, rather than only the enclosed baryonic mass, becomes decisive for the gravitational response. This suggests a route for explaining the difference between the galactic RAR and the cluster RAR within a single modified-field framework (Penner, 9 Feb 2026).

4. GRAS-related graph models in machine learning and computational biology

The string “GraS” also appears in Multi-GraS, a graph-based model for extractive text summarization (Jing et al., 2021). Multi-GraS is built on a Multiplex Graph Convolutional Network (Multi-GCN) that jointly models multiple types of relationships among sentences and words. At the word level, it uses intra-sentential syntactic and semantic graphs; at the sentence level, it uses inter-sentential natural-connection and semantic-similarity graphs. The sentence-level natural connection is defined through shared keywords,

$k$0

while semantic sentence similarity is

$k$1

For each relation $k$2, Multi-GCN applies a separate GCN with an inner skip connection,

$k$3

followed by relation aggregation and an outer skip connection (Jing et al., 2021).

The model is evaluated on the CNN/DailyMail benchmark. The reported ROUGE scores are 43.16 for R-1, 20.14 for R-2, and 39.49 for R-L, compared with 42.95, 19.76, and 39.23 for HSG in the reported table (Jing et al., 2021). The paper states that removing the multiplex GCN from either the word or sentence block reduces performance, that both inner and outer skip connections are crucial, and that all modeled relation types contribute to the final result. In that literature, “GraS” is therefore a model name tied to graph-based summarization rather than to arithmetic or gravity (Jing et al., 2021).

A related acronymic development is GRASMOS, short for Graph Signage Model Selection for Gene Regulatory Networks (Brilliantova et al., 2022). GRASMOS addresses signed directed networks by fitting sign-generation models to a fixed graph topology through maximum likelihood. The latent structure assigns each node a group label $k$4, typically in $k$5, and the likelihood is

$k$6

The framework defines Node-Oblivious, Source-Consistent, Target-Consistent, and Bi-Node-Consistent models, with algorithms ranging from closed-form expressions to MCMC sampling. The paper reports evaluation on synthetic datasets and on real-world GRNs including E. coli and Bacillus subtilis, with BNC models achieving the best negative log-likelihood on the cited real datasets (Brilliantova et al., 2022). Although GRASMOS is not itself “GRAS,” it illustrates how the same letter sequence recurs in graph-centric modeling nomenclature.

5. GRAS as a benchmark for demographic bias in vision-LLMs

In multimodal ML, GRAS expands to Gender, Race, Age, and Skin Tone and denotes a benchmark for measuring demographic bias in vision-LLMs (Malik et al., 26 Aug 2025). The benchmark is designed for visual question answering and is presented as offering the most diverse coverage to date across four demographic attributes. Its dataset consists of 5,010 stratified face images from FairFace and AI-Face, 100 personality trait words—50 positive and 50 negative—and five linguistically diverse, yet semantically equivalent templates for each image–trait pair. The total prompt count per evaluated model is therefore

$k$7

(image, trait, template) combinations (Malik et al., 26 Aug 2025).

Bias is evaluated through the probability assigned to “Yes” for each prompt,

$k$8

followed by between-group tests: Welch’s ANOVA for race, age, and skin tone, and Welch’s $k$9-test for gender (Malik et al., 26 Aug 2025). The paper introduces the GRAS Bias Score, defined as

$E_K$0

so that the score is the percentage of attribute–trait–template cases in which no statistically significant between-group bias is detected. The scale is interpreted from 0 (maximal bias) to 100 (unbiased) (Malik et al., 26 Aug 2025).

The benchmarking results are severe. The abstract states that the least biased evaluated model attains a GRAS Bias Score of only 2 out of 100, and the detailed table reports scores of 2.00 for llava-1.5-7b-hf, 1.75 for paligemma2-3b-mix-224, 1.00 for Qwen2.5-VL-3B-Instruct, 0.25 for blip2-opt-2.7, and 0.00 for Phi-4-multimodal-instruct (Malik et al., 26 Aug 2025). The paper further reports that darker skin tones receive higher mean probabilities for more than 80% of negative traits, that lighter skin tones are more likely to be assigned positive traits, and that question formulation materially changes responses. A central methodological conclusion is that evaluating bias in VLMs with VQA requires considering multiple formulations of a question (Malik et al., 26 Aug 2025).

6. Disambiguation and cross-domain significance

A common source of confusion is that GRAS is not a single standardized research object. The same letter string, or a one-letter variant such as GRASS, identifies unrelated theories, conjectures, benchmarks, and models across mathematics, statistics, astrophysics, NLP, and fairness evaluation.

Term	Referent	Representative paper
Gras conjecture	Relationship between unit-group indices and class groups in abelian extensions	(Oukhaba et al., 2011, Oukhaba et al., 2011, All, 2015)
GRASS	Graphical sure screening for Gaussian graphical models	(Luo et al., 2014)
GRAS	GRavitational Anti-Screening	(Penner, 9 Feb 2026)
Multi-GraS	Multiplex Graph Summarization for extractive summarization	(Jing et al., 2021)
GRAS	Gender, Race, Age, and Skin Tone benchmark for VLM bias	(Malik et al., 26 Aug 2025)

The disciplinary separation is substantial. In number theory, the term is tied to unit groups, Stark units, Euler systems, class groups, and ray class groups (Oukhaba et al., 2011). In statistics, it refers to thresholding the sample covariance matrix to recover a conditional dependence graph with sure screening and false-positive-rate guarantees (Luo et al., 2014). In astrophysics, it denotes a modified gravitational response that is mathematically equivalent to AQUAL in its standard form and modified to address cluster-scale discrepancies (Penner, 9 Feb 2026). In machine learning, the same string appears in graph architectures for summarization and in a benchmark for demographic bias in VLMs (Jing et al., 2021). This suggests that any technical use of “GRAS” should be read locally, with explicit attention to the surrounding field, notation, and cited literature.