GASP: Versatile Uses in Astronomy, NLP & Beyond

• GASP is a recurring acronym defined by domain-specific expansions, each featuring distinct technical vocabularies and evaluation regimes.
• It ranges from astronomical instrumentation (full-Stokes polarimetry and gas stripping surveys) to NLP applications like scientific abstract generation and grounding diagnostics.
• GASP also represents geometry-based methods in graphics and vision, signed-graph clustering in optimization, and quantum circuit synthesis, underscoring its interdisciplinary impact.

In arXiv usage, GASP is not a single standardized concept but a recurrent acronym applied to unrelated instruments, surveys, algorithms, and pre-training frameworks. It names, among other things, an astronomical full-Stokes polarimeter, a MUSE survey of gas stripping in galaxies, a text-to-text task for abstract generation from cited abstracts, several geometry- and Gaussian-based methods in graphics and vision, a signed-graph clustering framework, a quantum circuit synthesis algorithm, and a self-supervised driving representation (0905.0084, Poggianti et al., 2017, Zanzotto et al., 2020, Borycki et al., 2024). The commonality is therefore lexical rather than genealogical: each occurrence is domain-specific, with its own technical vocabulary, assumptions, and evaluation regime.

1. Major expansions and research domains

The literature uses the acronym for multiple independent expansions.

Expansion	Domain	Brief characterization
Galway Astronomical Stokes Polarimeter	Astronomical instrumentation	Ultra-high-speed, full-Stokes astronomical imaging polarimeter based on Division of Amplitude Polarimetry (0905.0084)
GAs Stripping Phenomena in galaxies with MUSE	Extragalactic astronomy	MUSE integral-field survey of gas removal processes in galaxies (Poggianti et al., 2017)
Generating Abstracts of Scientific Papers from abstracts of cited papers	NLP / scientific text generation	Text-to-text task that maps cited-paper abstracts to a target abstract (Zanzotto et al., 2020)
Generalized Algorithm for Signed graph Partitioning	Graph clustering / vision	Hierarchical agglomerative framework for signed graphs and instance segmentation (Bailoni et al., 2019)
Gated Attention For Saliency Prediction	Saliency prediction	Multimodal dynamic saliency model integrating gaze and affect cues (Abawi et al., 2022)
Genetic Algorithm for State Preparation	Quantum computing	Evolutionary synthesis of low-depth state-preparation circuits (Creevey et al., 2023)
Gaussian Splatting for Physics-Based Simulations	Graphics / physics simulation	Reparameterizes flat Gaussians as triangles for black-box physics engines (Borycki et al., 2024)
Gaussian Avatars with Synthetic Priors	Neural rendering / avatars	Monocular fitting of real-time 360° Gaussian head avatars (Saunders et al., 2024)
Geometric and Semantic self-supervised Pre-training	Autonomous driving	Continuous 4D occupancy and semantic-field pre-training (Ljungbergh et al., 19 Mar 2025)
Grounding-Aware Sensitivity by Perturbation	RAG hallucination detection	Span-level detector based on context sensitivity of likelihood and JSD (Bouke, 5 Jul 2026)
Geometric-Aware Spatial Priors	Vision-LLMs	Injection of geometric priors into transformer layers for 3D reasoning (Yeh et al., 28 May 2026)
Gradient-Aware Shortest Path	Topological visualization	Boundary-confined Reeb-graph embedding aligned with scalar-function gradient (Rahman et al., 7 Aug 2025)
Generative Adversarial Suffix Prompter	LLM red-teaming	Black-box generation of human-readable jailbreak suffixes (Basani et al., 2024)

This distribution suggests that “GASP” functions as a reusable acronymic form rather than a stable technical term. Disambiguation therefore depends on the expansion, the paper title, and the host field.

2. Astronomy: polarimetry and environmental galaxy evolution

In observational instrumentation, GASP most prominently denotes the Galway Astronomical Stokes Polarimeter, an ultra-high-speed, full-Stokes astronomical imaging polarimeter based on a Division of Amplitude Polarimeter (DOAP). Its defining property is simultaneous measurement of the complete Stokes vector from a single exposure, with no moving parts or temporal modulation. The basic retrieval equation is

$\mathbf{I}=\mathbf{A}\mathbf{S},$

with $\mathbf{S}=(I,Q,U,V)^T$, and Stokes recovery obtained by inversion of the calibrated system matrix $\mathbf{A}$. The optical design centers on a modified Fresnel rhomb acting both as a highly achromatic quarter-wave plate and a beamsplitter, followed by Wollaston prisms that generate four analyzer channels (0905.0084).

Calibration became a central technical issue in the later study of the instrument. The 2015 investigation of the Eigenvalue Calibration Method (ECM) applied to GASP treated both the non-imaging APD configuration and the imaging EMCCD configuration, emphasizing that ECM compensates for systematic errors in calibration optics while recovering the instrument matrix in the exact in-use configuration. In laboratory APD tests, the reported errors were 0.2% for the degree of linear polarisation and 0.1° for the polarisation angle, while imaging performance was limited primarily by channel-to-channel image registration and transmitted-path aberrations (Kyne et al., 2015).

A second, entirely unrelated astronomical usage is GASP = GAs Stripping Phenomena in galaxies with MUSE, an ESO Large Programme designed to study gas removal in galaxies through deep MUSE integral-field spectroscopy. The survey targets 114 galaxies at $z=0.04\text{–}0.07$, spanning $M_\star \sim 10^{9.2}\text{–}10^{11.5}\,M_\odot$ and environments from groups to rich clusters. Its science program centers on where, why, and how gas is removed, how that affects star formation and quenching, the interplay with AGN activity, and the stellar and metallicity histories of stripped and control galaxies. The introductory paper uses JO206 as a textbook ram-pressure-stripping case, with ionized gas tails extending to about 90 kpc and clear gas–stellar kinematic decoupling (Poggianti et al., 2017).

Later papers in that survey retained the acronym while narrowing the scope. GASP XXI studied 54 ram-pressure-stripped cluster galaxies and showed that no single parameter dominates the amount of star formation in stripped tails. Its analytic model links stripped-gas fraction and tail SFR to cluster velocity dispersion, galaxy stellar mass, clustercentric distance, and speed in the intracluster medium, and concluded that the star formation efficiency in tails is lower than in galaxy discs by a factor of about 5. It also estimated an average present-day tail SFR of about 0.22 $M_\odot\,\mathrm{yr}^{-1}$ per cluster and an integrated contribution of about $4\times10^9\,M_\odot$ of stars formed in RPS tails since $z\sim1$ (Gullieuszik et al., 2020).

GASP XXX shifted from global to spatially resolved star formation, analyzing the $\Sigma_{\rm SFR}$–$\Sigma_M$ relation in 40 stripping cluster galaxies. Even on $\mathbf{S}=(I,Q,U,V)^T$0 kpc scales, stripping galaxies were found to be offset upward in $\mathbf{S}=(I,Q,U,V)^T$1 at fixed $\mathbf{S}=(I,Q,U,V)^T$2, with an enhancement of about 0.35 dex at $\mathbf{S}=(I,Q,U,V)^T$3 relative to undisturbed controls. The effect persisted across the disks and was interpreted as consistent with ram-pressure-induced compression waves rather than a purely localized starburst (Vulcani et al., 2020).

3. Scientific text generation, grounding, and adversarial prompting

In NLP, GASP first appears as “Generating Abstracts of Scientific Papers from abstracts of cited papers”, proposed as a text-to-text task for studying scientific creativity. The formal instance is written as

$\mathbf{S}=(I,Q,U,V)^T$4

where the input is the set of cited-paper abstracts and the target is the abstract of the citing paper. The paper frames this as more than multi-document summarization because the target abstract is dominated by novel contribution rather than background. The corpus was built from Semantic Scholar, with intended splits of 100,000 training, 10,000 validation, and 10,000 test instances, and baseline experiments with TextRank and a BiLSTM pointer-generator yielded low ROUGE scores, underscoring the difficulty of the task (Zanzotto et al., 2020).

A later usage in retrieval-augmented generation is Grounding-Aware Sensitivity by Perturbation. Here GASP is a span-level hallucination detector that measures how strongly each answer sentence depends on retrieved evidence. The method keeps the answer fixed, rescoring it under full context, no context, and leave-one-out context ablations, then aggregates log-likelihood gaps and Jensen–Shannon divergences into four grounding features: overall context dependence, full-versus-no-context divergence, maximal leave-one-out likelihood drop, and maximal leave-one-out divergence. On RAGTruth it reaches a response-level AUC of about 0.73 and a span-level AUC of about 0.67, and the paper argues that the method is most effective when outputs must be constructed from retrieved context rather than recalled from parametric memory (Bouke, 5 Jul 2026).

A further language-model usage is Generative Adversarial Suffix Prompter, a black-box jailbreak framework. It trains a separate SuffixLLM, uses latent Bayesian optimization to search a continuous embedding space for adversarial suffixes, evaluates them with GASPEval, and refines the generator with ORPO. The emphasis is on human-readable suffixes rather than token-level gibberish. Reported results include higher attack success rates than GCG, AutoDAN, and AdvPrompter on several open-source and closed models, together with lower training time than AdvPrompter by about 1.75× and high readability in both LLM-based and human evaluation (Basani et al., 2024).

Taken together, these three NLP usages illustrate a broad shift in the meaning of the acronym within language research: from creative scientific text generation, to factual grounding diagnostics, to adversarial prompt construction.

4. Graphics, vision, and geometric representation

Several recent vision and graphics papers use GASP for geometry-aware methods built around explicit spatial structure. In “Gaussian Splatting for Physics-Based Simulations”, the method bridges 3D Gaussian Splatting and standard physics engines by reparameterizing each flat Gaussian as a triangle defined by three 3D points, simulating those points in a point-based engine such as MPM, and reconstructing Gaussians after deformation. The forward geometry is

$\mathbf{S}=(I,Q,U,V)^T$5

with inverse reconstruction of $\mathbf{S}=(I,Q,U,V)^T$6 from the deformed triangle. Because the physics engine operates on points, it is treated as a black box; no solver modification is required (Borycki et al., 2024).

In neural rendering, “Gaussian Avatars with Synthetic Priors” trains a synthetic prior over Gaussian avatar parameters and then fits it to a single photo or short monocular video. The prior factorization is

$\mathbf{S}=(I,Q,U,V)^T$7

where $\mathbf{S}=(I,Q,U,V)^T$8 is a canonical Gaussian template, $\mathbf{S}=(I,Q,U,V)^T$9 is a learnable per-Gaussian feature, and $\mathbf{A}$0 is an identity latent. After a three-stage fitting pipeline, the resulting avatar supports 360° rendering and real-time animation at about 70 fps on commercial hardware, while the prior itself is only needed during fitting, not inference (Saunders et al., 2024).

In saliency prediction, “Gated Attention For Saliency Prediction” uses social cues as explicit input modalities. A first stage extracts gaze following, 3D gaze direction, facial expression representations, and audiovisual saliency maps; a second-stage GASP module applies a Directed Attention Module and gated multimodal fusion, including GMU, RGMU, and attentive LSTM variants. The inversion used in DAM is

$\mathbf{A}$1

The reported result is that gaze direction and affective representations improve prediction-to-ground-truth correspondence by at least 5% relative to dynamic saliency models without social cues, with different fusion strategies favored in static and dynamic settings (Abawi et al., 2022).

Topological visualization supplies another distinct expansion: “A Gradient-Aware Shortest Path Algorithm for Boundary-Confined Visualization of 2-Manifold Reeb Graphs.” This GASP decomposes the manifold into topological cylinders for each Reeb-graph edge, samples candidate points on contour levels, and computes shortest paths constrained to progress monotonically in scalar value. The method is explicitly designed to satisfy three properties: boundary confinement, compactness, and alignment with the function gradient, and is evaluated against the geometric barycenter algorithm in TTK (Rahman et al., 7 Aug 2025).

A related but separate geometric-ML usage is “Geometric-Aware Spatial Priors” for VLMs. Here GASP attaches a small correspondence head to every transformer layer and applies deep supervision using a contrastive correspondence loss and a depth consistency loss. The paper reports that standard VLMs have internal correspondence matching accuracy often below 5%, whereas GASP boosts peak layer-wise correspondence to over 70% and temporal robustness to over 85%, with downstream gains including +18.2% on All-Angles Bench and +29.0% on VSI-Bench, all without any 3D VQA training (Yeh et al., 28 May 2026).

5. Graph partitioning, quantum state preparation, and autonomous driving

Outside vision and astronomy, the acronym also names several optimization-oriented frameworks. “Generalized Algorithm for Signed graph Partitioning” presents GASP as a unifying agglomerative framework for signed graphs with attractive and repulsive edges. It generalizes hierarchical clustering, recovers existing algorithms such as GAEC and Mutex Watershed, introduces new constrained variants, and proves that some instantiations define an ultrametric. Its importance in computer vision comes from a simple bottom-up instance-segmentation pipeline that, when combined with CNN predictions, achieved state-of-the-art accuracy on the CREMI 2016 EM segmentation benchmark without domain-specific superpixels (Bailoni et al., 2019).

In quantum computing, “A Genetic Algorithm for State Preparation” uses a hybrid search over circuit structure and continuous gate parameters to prepare target quantum states on NISQ hardware. Individuals are encoded as sequences of $\mathbf{A}$2, $\mathbf{A}$3, $\mathbf{A}$4, and CNOT gates, and fitness is state fidelity,

$\mathbf{A}$5

The method adaptively increases circuit length only when necessary. On benchmark Gaussian and W states, GASP produced circuits with lower depth and gate count than Qiskit’s exact synthesis, and under noise its shorter circuits yielded higher practical fidelity on IBM simulators and hardware (Creevey et al., 2023).

In autonomous driving, “Geometric and Semantic self-supervised Pre-training” uses GASP to learn a continuous 4D world model from unlabeled logs. Given past LiDAR scans, the model predicts at a queried future spacetime point $\mathbf{A}$6: general occupancy, ego occupancy, and distilled DINOv2 features. The training signal combines future occupancy supervision from LiDAR rays, ego-path occupancy, and semantic feature distillation into a unified representation of scene geometry and evolution. The paper reports gains on semantic occupancy forecasting, online mapping, and ego trajectory prediction, arguing that continuous 4D geometric and semantic occupancy prediction is a scalable pre-training paradigm for autonomous driving (Ljungbergh et al., 19 Mar 2025).

6. Disambiguation and recurrent patterns in the literature

The literature suggests that treating GASP as a single lineage is misleading. Some usages denote long-running programs with sequenced papers—most clearly the galaxy-survey series GASP I, GASP XXI, and GASP XXX—whereas others denote single methods or instruments with no genealogical relation to the survey usage (Poggianti et al., 2017). Other expansions recur around particular lexical themes: geometry, Gaussian representations, grounding, and generation appear repeatedly in recent machine-learning papers, but these similarities are nominal rather than methodological. A plausible implication is that acronym disambiguation in arXiv practice depends less on the acronym itself than on the paper title’s expansion and the surrounding technical context.

A second recurrent pattern is structural. Many GASP papers, despite being unrelated, map naturally onto one of three technical templates. One is measurement and calibration, exemplified by the polarimeter’s system matrix and ECM procedure. Another is latent or field-based representation learning, as in Gaussian avatars, geometric spatial priors, and autonomous-driving occupancy fields. The third is search or optimization over constrained spaces, as in signed-graph agglomeration, latent Bayesian optimization for jailbreak suffixes, shortest-path Reeb-graph layout, and genetic circuit synthesis. This suggests that the acronym has often been attached to methods that convert an ill-posed problem into a structured representation plus an explicit optimization or inference procedure.

For bibliographic and scholarly use, the practical consequence is straightforward: “GASP” should always be expanded on first mention. Without expansion, the acronym is underdetermined across astronomy, NLP, computer vision, quantum computing, graphics, and autonomous driving. In arXiv-centered discourse, the stable identifier is therefore the expansion and the arXiv record, not the acronym alone.