AQuA and AQUA: Research Systems and Frameworks

• AQuA and AQUA are domain-specific names for unrelated research systems, datasets, models, and software frameworks rather than a single unified technology.
• Applications range from GPU-memory offloading, multimodal question answering, video and label-quality assessment, water equations of state, quantum annealing, and cybersecurity evaluation.
• Researchers should identify the field, acronym expansion, and source paper before comparing results, since methods, metrics, limitations, and use cases differ substantially across implementations.

AQuA and AQUA are names used by multiple unrelated research systems, datasets, physical models, and software frameworks. The acronym denotes, among other subjects, Aqua Computing, art and audio question answering, analytical quality assessment, label-quality benchmarking, deliberative-quality scoring, quantum-annealing process engineering, multimodal copyright watermarking, GPU-memory management, attention approximation, cybersecurity evaluation, a free-electron-laser beamline, and recursive quantitative-trading research. The meanings are domain-specific and should not be conflated.

1. Computing, networking, and infrastructure

Aqua Computing

“Aqua Computing: Coupling Computing and Communications” defines Aqua Computing as a paradigm in which computing services are delivered jointly with networking services rather than independently (Magurawalage et al., 2015). Its central process, computation dissolving, couples computation, storage, radio resources, and communication through joint or cooperative resource allocation, service integration, and hardware integration.

The architecture is intended to enhance connected devices automatically when they attach to an Aqua-enabled network. Computational functions may remain on the device, move to a transient mobile-edge cloud, or reside in a persistent central cloud. The “Aqua” terminology reflects the water cycle: central clouds correspond to centralized data centers, fogs to mobile-edge clouds, and liquid water on the ground to local device computation.

A user-specific clone is the principal architectural abstraction. A clone is a segregated virtualized compute-and-storage space, proxy, cache, communication intermediary, and context-aware representative of a subscriber. Persistent clones are maintained in a central cloud; transient clones may be spawned near a serving base station, BBU pool, Wi-Fi access point, or broadband gateway and later migrated or destroyed.

The Mobile Cloud Controller monitors radio resources, physical machines, hypervisors, virtual machines, applications, and traffic. It allocates radio resources through a base-station API, cloud resources through a Mobile Cloud API, and traffic through an SDN controller. The prototype combines C-RAN, OpenStack, Android-x86 clones, ThinkAir offloading, sFlow monitoring, OpenDaylight, and a Pica8 SDN switch.

The paper gives the latency model

$T=\frac{D-S}{r}+\frac{F}{f},$

where $D-S$ is the data transferred, $r$ is communication bitrate, $F$ is the instruction count, and $f$ is allocated processing speed. The model expresses the trade-off between communication and computation, but the paper does not provide a complete multi-user optimization formulation or named joint scheduling algorithm.

The prototype’s clearest demonstrated result concerns RTMP streaming: an edge clone buffers video and handles retransmissions locally, making RAN-backhaul retransmission traffic “almost” zero under packet loss. The work is therefore primarily an architectural vision with prototype validation, not a complete formal theory of coupled resource allocation.

Network-accelerated LLM memory

“AQUA: Network-Accelerated Memory Offloading for LLMs in Scale-Up GPU Domains” addresses GPU-memory contention during bursty LLM inference (Kumar et al., 2024). It combines time-sliced, preemptive prompt scheduling with paging of KV-cache and LoRA state between GPUs over NVLink or NVSwitch.

AQUA introduces migratable AQUA Tensors whose physical storage can reside in local HBM, neighboring producer-GPU HBM, or host DRAM. A consumer GPU hosts a memory-intensive workload; a producer GPU contributes otherwise-unused HBM. A centralized coordinator manages allocation, leasing, reclamation, tensor movement, and producer–consumer relationships.

The scheduler is inspired by Linux’s completely fair scheduler and uses generated-token count as its service measure. At slice boundaries, fragmented context tensors are gathered into contiguous buffers, transferred to remote GPU memory, and later restored. Custom CUDA kernels implement gather and scatter operations to avoid inefficient small NVLink transfers.

The reported results include approximately fourfold lower TTFT for bursty CodeLlama-34B inference, six times more generated tokens than FlexGen for 8,000-token OPT-30B prompts, and up to 1.8-fold lower request-completion time for LoRA workloads. The principal limitation is dependence on scale-up GPU connectivity and available producer memory; AQUA falls back to host DRAM when no remote HBM is available and provides no formal worst-case latency bound.

Attention approximation

“AQUA: Attention via QUery mAgnitudes for Memory and Compute Efficient Inference in LLMs” is a distinct inference-time attention approximation method (S et al., 14 Sep 2025). It first computes an offline SVD-based rotation matrix from calibration activations containing both queries and keys. During inference, queries and keys are rotated, and the dimensions with the largest query magnitudes are selected dynamically.

The approximate score calculation uses $k$ dimensions instead of all $d_{\mathrm{head}}$ dimensions. Its modeled complexity is

$$C_{\mathrm{AQUA}}=O\left(d_{\mathrm{head}}^2+(i+1)k\right),$$

compared with standard attention’s

$$C_{\mathrm{std}}=O\left((i+1)d_{\mathrm{head}}\right).$$

The projection cost must therefore be amortized over sufficiently long contexts. For Llama-3.1-8B-Instruct, retaining 75% of projected dimensions removes 25% of attention dot-product computation with negligible reported benchmark changes. More aggressive pruning causes sharp degradation, particularly on reasoning tasks.

AQUA can be combined with H2O token eviction, which reduces the number of attended tokens, and with AQUA-Memory, which statically reduces KV-cache dimensionality before applying dynamic query-magnitude selection.

2. Question answering and multimodal interaction

Art question answering

“A Dataset and Baselines for Visual Question Answering on Art” introduces AQUA, or Art QUestion Answering, as a benchmark based on paintings and associated comments from the SemArt dataset (Garcia et al., 2020). It separates questions into visual questions, answerable from the painting, and knowledge questions, answerable from textual art-historical comments.

The final dataset contains 79,848 cleansed question–answer pairs: 32,345 visual and 47,503 knowledge examples. Questions are generated automatically using object detection, image captioning, and textual question-generation methods, then reviewed through Amazon Mechanical Turk for grammaticality, answer existence, answer correctness, modality dependence, and reasonability.

The VIKING baseline contains a modality selector, a visual branch, and a knowledge branch. BERT and ResNet-152 determine whether a question is visual or knowledge-based. The visual branch uses iQAN; the knowledge branch retrieves comments with TF-IDF and BERT reranking, then extracts an answer span with XLNet.

VIKING obtains an exact-match score of 0.555, compared with 0.352 for the knowledge-only variant and 0.204 for the visual-only variant. The modality selector reaches approximately 0.996 accuracy, although this result is likely facilitated by artifacts of the distinct question-generation procedures. The benchmark is preliminary: its visual questions are noisy, its answer vocabulary is restricted, and most questions require only one modality rather than integrated visual and historical reasoning.

Audio question answering

“AQUALLM: Audio Question Answering Data Generation Using LLMs” presents a framework for generating large AQA datasets from audio-caption corpora (Behera et al., 2023). AQUALLM uses four modules: the Candidate Answer Extraction Module, Question Generation Module, Question–Answer Filtering Module, and Question Paraphrasing Module.

Candidate answers are extracted from captions using spaCy. T5 models generate questions, another T5 model performs caption-conditioned answer verification, and a token-level $F_1$ threshold of 0.55 filters inconsistent pairs. A T5 model trained on Quora Question Pair then produces the top five similar questions for each retained item.

The framework produces three datasets:

Dataset	Audios	Unique questions	Unique answers
AQUALLM-Clotho	5,929	438,600	25,235
AQUALLM-AudioCaps	51,308	728,310	35,008
AQUALLM-MACS	3,930	268,875	11,874

Using MWAFM, binary accuracy rises from 68.75% on ClothoAQA to 95.58–96.84% on the AQUALLM datasets. These results demonstrate high internal learnability, but they do not by themselves establish open-domain audio grounding: the generated examples depend heavily on captions, round-trip verification, and synthetic paraphrases.

Unanswerability in audio QA

“AQUA-Bench: Beyond Finding Answers to Knowing When There Are None in Audio Question Answering” extends AQA to answerability assessment (Kuan et al., 18 Jan 2026). It evaluates whether a model can abstain when an answer is absent, the candidate set is semantically incompatible, or the question is not supported by the audio.

The benchmark defines three scenarios:

1. Absent Answer Detection: the correct answer is removed from the choices.
2. Incompatible Answer Set Detection: the choices belong to the wrong semantic category.
3. Incompatible Audio Question Detection: the question asks for information unavailable from the audio.

AQUA-Bench supplements ordinary accuracy with conditional accuracy, which grants credit for an unanswerable item only when the model also correctly answers its corresponding answerable version. This prevents a model from receiving credit for abstention merely because it failed to understand the original question.

The benchmark reveals a strong gap between answerable and unanswerable performance. For example, Qwen2.5-Omni obtains 96.4% on animal-sound questions but 20.5% on animal-sound absent-answer detection. Explicit instructions and chain-of-thought prompting improve results, suggesting that many failures involve forced-choice behavior and poor answerability policy rather than an inability to perceive the sound.

Tutorial-video question answering

“AQuA: Automated Question-Answering in Software Tutorial Videos with Visual Anchors” addresses questions about screen-recorded software tutorials (Yang et al., 2024). It treats questions as multimodal acts of reference: the textual question is augmented by user-drawn visual anchors, transcript context, tutorial metadata, software documentation, and visual recognition.

The system was demonstrated for Fusion 360. Its Visual Recognition Module combines BLIP-2 image captioning, UIED region detection, feature matching against a database of 1,286 UI images, template matching, and Google Cloud Vision OCR. A Retrieval Module searches 2,937 documentation files and 2,375 tutorial videos, segmented into 5,635 embedded chunks. GPT-4 then generates an answer using the retrieved material, transcript context, question, and anchor description.

In a 69-question evaluation with 16 participants, the full AQuA pipeline was selected as the preferred answer in 55.4% of cases, compared with 26.8% for question-plus-video context and 17.8% for question-only input. Its limitations include incomplete UI databases, retrieval errors, verbosity, hallucination, and approximately 30 seconds of answer-generation latency.

Ambiguous visual question answering

“AQuA: Toward Strategic Response Generation for Ambiguous Visual Questions” introduces a four-level ambiguity taxonomy for VQA (Jang et al., 8 Mar 2026). The associated response strategies are direct answering, contextual inference, listing plausible alternatives, and requesting clarification.

Level 0 contains unambiguous questions. Level 1 contains low-level referential ambiguity with a dominant referent. Level 2 contains two or three plausible interpretations, for which alternatives should be enumerated. Level 3 contains many similarly salient referents and requires clarification.

The dataset contains approximately 7.2K MS COCO-based samples balanced across four levels. Qwen2.5-VL-3B-Instruct and InternVL3-2B-Instruct are fine-tuned with supervised learning and Group Relative Policy Optimization. The reward favors the correct strategy and factual grounding, with a penalty parameter of $D-S$0 for strategically correct but factually distorted responses.

AQUA-tuned Qwen2.5-VL-3B reaches 86.28% overall strategic accuracy, compared with 32.83% in its zero-shot configuration and 42.25% for GPT-5 with strategy prompting. The findings indicate that factual grounding and strategic appropriateness are distinct capabilities: a model may identify a real object while answering about the wrong referent.

3. Quality assessment, annotation, and discourse

Analytical video quality

“AQuA: Analytical Quality Assessment for Optimizing Video Analytics Systems” defines AQuA as a pre-inference quality filter for edge video analytics (Paul et al., 2021). It assesses whether a frame is useful to a downstream model rather than whether it appears visually pleasing to a human.

The framework introduces the Classifier Opinion Score and Mean Classifier Opinion Score, which measure degradation in classifier confidence and correct-class rank between original and distorted images. A semi-supervised variant, $D-S$1, compares complete softmax distributions and does not require labels.

AQuA uses a truncated Inception-v3 feature extractor through the first Inception module, followed by pooling and fully connected regression layers. The feature extractor is frozen during training. The model has approximately 45 MB and 1.5 GFLOPs, with reported latency of 14.2 ms per frame. A MobileNetV3 implementation reduces the model to approximately 10 MB.

Across face-recognition experiments, AQuA reduces false positives by 17.62% while reducing true positives by 0.55%. In continuous-video experiments, filtering reduces average computation time and bandwidth usage by approximately 25%, with the exact trade-off controlled by thresholds and temporal sampling. The method is not universal: analytical quality depends on the downstream task, distortion distribution, model architecture, and operating point.

Label-quality benchmarking

“AQuA: A Benchmarking Tool for Label Quality Assessment” introduces a benchmark and toolkit for detecting erroneous labels (Goswami et al., 2023). It covers 17 real-world public datasets across image, text, tabular, and time-series modalities and supports uniform, asymmetric, class-dependent, instance-dependent, dissenting-label, dissenting-worker, and crowd-majority noise.

The benchmark evaluates two complementary hypotheses: whether a method identifies corrupted examples and whether removing those examples improves downstream classification. Its principal baseline methods are AUM, Confident Learning, CINCER, and SimiFeat, together with a no-cleaning baseline.

SimiFeat generally ranks first for direct detection of synthetic errors, with CINCER often close behind. AUM is competitive in several settings, while Confident Learning performs poorly in the reported benchmark partly because it removes approximately 57.4% of training examples on average. Downstream differences are smaller because deep classifiers can tolerate moderate noise.

AQuA’s central methodological contribution is standardization: identical corrupted examples, fixed downstream models, multiple modalities, multiple noise mechanisms, multiple metrics, data cards, and statistical comparisons. Its results do not establish a universal best detector, and synthetic noise, limited hyperparameter tuning, and single-label classification constrain interpretation.

Deliberative-quality scoring

“AQuA: Combining Experts’ and Non-Experts’ Views To Assess Deliberation Quality in Online Discussions Using LLMs” defines AQuA as an Additive deliberative Quality score with Adapters (Behrendt et al., 2024). It combines 20 adapter-predicted deliberative criteria into an interpretable score.

The criteria cover rationality, reciprocity, civility, and storytelling. Examples include relevance, facts, justification, solution proposals, additional knowledge, references to users and content, respect, vulgarity, insults, sarcasm, discrimination, and storytelling.

For comment $D-S$2, the unnormalized score is

$D-S$3

where $D-S$4 is a criterion-specific adapter prediction and $D-S$5 is the Pearson correlation between that expert-coded criterion and non-expert perceptions of an enriching comment. The result is affinely normalized to the interval $D-S$6.

Twenty adapters are trained on frozen multilingual-BERT representations using KODIE annotations. The selected system achieves weighted-average $D-S$7. On the unseen SOCC dataset, the normalized score reaches 81.73 $D-S$8 for binary constructiveness. The additive design preserves a decomposition of the score into individual deliberative aspects, but its coefficients are dataset-specific associations rather than universal normative weights.

4. Scientific models and experimental facilities

H₂O equation of state

“AQUA: A Collection of H$D-S$9O Equations of State for Planetary Models” provides a wide-domain, tabulated collection of water equations of state for planetary modeling (Haldemann et al., 2020). It spans approximately 0.1 Pa to 400 TPa and 150 K to $r$0 K, covering gas, liquid, supercritical fluid, ordinary ice, high-pressure ice, superionic water, and hot dissociated or ionized regimes.

AQUA combines seven component regions: ice-Ih; intermediate-pressure ices; high-pressure ice; low-pressure liquid, gas, and supercritical water; higher-pressure liquid and supercritical water; hot dilute gas; and high-temperature high-pressure fluid and superionic water. Component models include IAPWS, SeaFreeze, French and Redmer, Brown, CEA, and Mazevet et al.

The tables provide density, entropy, internal energy, bulk sound speed, adiabatic temperature gradient, phase identifiers, and, in the gas regime, molecular weight, dissociation fraction, and ionization fraction. The products include $r$1–$r$2, $r$3–$r$4, and $r$5–$r$6 grids.

AQUA is not globally derivable from one thermodynamic potential. It uses physical phase switching, derivative interpolation, overlap matching, and reference-state shifts. Consequently, it is thermodynamically consistent within component regions but not exactly consistent everywhere, particularly near phase boundaries, the critical point, and interpolated transitions.

Above approximately 10 GPa, AQUA predicts systematically higher densities than ANEOS and QEOS. For pure-water spheres, this produces radii typically 3–8% smaller than ANEOS and 7–14% smaller than QEOS. The paper emphasizes that thermal profiles can matter even more: for a 0.5 Earth-mass pure-water sphere, the adiabatic radius exceeds the isothermal radius by approximately 0.27% at 200 K, 103% at 1000 K, and 1254% at 2000 K.

AQUA free-electron-laser beamline

“FEL performance and tolerance studies of the EuPRAXIA@SPARC_LAB beamline AQUA” describes a proposed soft-X-ray SASE free-electron-laser beamline (Nguyen et al., 7 Aug 2025). Its target is the 3–4 nm water window, corresponding approximately to 310–410 eV photon energy.

The beamline uses an electron beam accelerated to approximately 1–1.2 GeV by an X-band normal-conducting linac followed by plasma wakefield acceleration. The radiator contains ten two-meter APPLE-X permanent-magnet undulator modules with an 18 mm period and selectable linear or circular polarization.

At a 1 GeV, 1.5 kA working point and a 4 nm target wavelength, semi-analytical models predict saturation lengths of approximately 25–28 m for linear polarization and 15–20 m for circular polarization. The expected FEL output power exceeds $r$7 W.

The resistive-wall study identifies a 2.5 mm chamber radius as a practical compromise. For the short-bunch example, longitudinal wakefields do not noticeably alter FEL power growth. Transverse effects are more sensitive to offsets. To retain at least 60% of ideal peak power in the coupled offset–tilt scan, the approximate requirements are

$r$8

These are simulation-based tolerances rather than a complete start-to-end or Monte Carlo reliability budget.

5. Quantum annealing and quantitative research

Agile quantum-annealing development

“AQUA: an Agile Process to Develop Quantum Annealing Applications” defines AQUA as Agile QUantum Annealing, a software lifecycle for QUBO and quantum-annealing applications (Marchesi et al., 20 Jan 2026). It combines Design Science Research with Scrum-like iteration and explicit decision gates.

The four stages are:

1. Initial assessment and formal modeling.
2. Prototype-driven computational strategy and algorithm selection.
3. Agile implementation.
4. Deployment, validation, and ongoing maintenance.

The gates are the Lifecycle Objective Milestone, Algorithms Selection Milestone, Operational Capability Milestone, and Lifecycle Assessment Milestone. Each can result in continuation, iteration, redesign, or abandonment.

AQUA treats mathematical modeling, solver selection, embedding, decomposition, hardware access, classical baselines, and quantum experiments as first-class engineering activities. It supports direct QA, hybrid QA, QAOA, digital quantum annealers, and classical HPC alternatives.

In a credit-scoring feature-selection case, direct QA, hybrid QA, and QAOA prototypes were compared. Hybrid QA was selected because it was easier to implement and considered more suitable for larger QUBOs, despite higher latency and vendor dependence. On a larger dataset, selecting four of ten features increased classification accuracy from 0.75 to 0.83.

The case demonstrates process feasibility but not quantum advantage. The principal unresolved issue is that AQUA defines decision gates without universal quantitative thresholds for quality, runtime, cost, scalability, or operational value.

Recursive quantitative-trading research

“AQuA: Recursively Self-Improving Quantitative Trading Research Agents” studies recursive improvement at the level of research workflows rather than self-modification of the underlying LLM (Guo et al., 13 Aug 2026). It contains two independent systems: symbolic factor discovery for five-minute cryptocurrency data and trainable model development for intraday U.S. equities.

Both systems maintain persistent research state, evaluate candidates under fixed protocols, retain validated evidence, and use that evidence to guide subsequent hypotheses. Their sandboxes fix data splits, features, labels, and evaluators while restricting agent actions to causal factor expressions or configuration diffs.

Part I is a manager-mediated six-agent pipeline involving a Data Steward, Visual Analyst, Idea Miner, Factor Evaluator, Backtest Engineer, and Research Librarian. Mechanism-grounded factors are evaluated using information coefficients, regime stability, event windows, turnover, complexity, baseline comparisons, and sign reversal tests. Combined factor signals reach approximately 0.190 Spearman information coefficient on the crypto universe.

Part II searches configuration diffs over hybrid time-series architectures. Training uses 2010–2019 data, 2020 as an embargo period, and an untouched 2021–2025 test period. The hybrid model reaches a per-stock Pearson information coefficient of $r$9, compared with $F$0 for the strongest listed baseline. A threshold long/short strategy reaches held-out Sharpe up to $F$1 at a two-leg 2-basis-point cost and remains positive in each year from 2021 through 2025.

The authors qualify recursive improvement as bounded by human-authored objectives, fixed evaluators, restricted candidate spaces, and governance. The results do not establish unconstrained self-improvement or live-trading performance. Test isolation reduces but does not eliminate adaptive overfitting, and the reported factor and model details are insufficient for complete independent reproduction.

6. Copyright, cybersecurity, authorship, and cross-domain terminology

Multimodal RAG copyright protection

“Safeguarding Multimodal Knowledge Copyright in the RAG-as-a-Service Environment” introduces AQUA as a watermarking framework for image knowledge in text-to-text multimodal RAG (Chen et al., 10 Jun 2025). It addresses indirect watermark propagation from an image retriever through a vision-LLM into textual output.

AQUA uses two semantic watermark mechanisms. $F$2 embeds rare acronym–expansion pairs in synthetic images, while $F$3 embeds unusual object relationships such as a red apple on the head of a reading dog. Probe queries retrieve the image and ask the generator to reproduce an expected signature.

Verification uses retrieval rank, conditional generation success rate, verification success rate, normalized substring matching, and Welch statistical tests. Across MMQA and WebQA with four VLMs, acronym and spatial watermarks achieve high retrieval and generation success, substantially exceeding naive and optimization-based baselines. Normal-query retrieval and CGSR are reported as zero in the principal benign-query experiment.

The method is limited to image-retrieval-to-text-generation RAG architectures. Its security depends on watermark secrecy, image persistence in the retrieval index, model OCR or spatial reasoning, and resistance to filtering, query detection, deduplication, and paraphrasing. The experiments do not report PSNR, SSIM, LPIPS, comprehensive adversarial attacks, or direct storage overhead.

Cybersecurity QA evaluation

“AQUA-LLM: Evaluating Accuracy, Quantization, and Adversarial Robustness Trade-offs in LLMs for Cybersecurity Question Answering” evaluates small LLMs under base, quantized-only, fine-tuned, and fine-tuned-plus-quantized configurations (Gungor et al., 16 Sep 2025).

The study uses six models, CyberBench and CyberMetric datasets, 4-bit NF4 quantization, LoRA fine-tuning, and direct prompt-injection attacks measured by Attack Success Rate. Quantization reduces GPU memory by approximately 75% and improves latency by 1.10–1.28 times, but quantized-only models average approximately 67% accuracy and increase average ASR from 52.25% to 77.92%.

Fine-tuning raises average accuracy to 97.17% on CyberBench and 98.75% on CyberMetric, but full-precision fine-tuning can retain high attack success rates. Fine-tuning the quantized model provides the reported compromise: FTQ accuracy averages 94.92% on CyberBench and 98.25% on CyberMetric, while ASR is lower than full-precision fine-tuning in the reported averages.

The evaluation is limited by small datasets, multiple-choice formulation, direct single-turn prompt injection, automated harmful-content judging, high-end A100 hardware, incomplete per-model reporting, and the absence of comprehensive uncertainty analysis.

Computational authorship verification

“The Questio de aqua et terra: A Computational Authorship Verification Study” applies computational authorship verification to the Latin cosmological treatise traditionally attributed to Dante Alighieri (Leocata et al., 7 Jan 2025). The study uses 330 medieval Latin texts, including 16 Dantean and 314 non-Dantean texts, and evaluates logistic-regression systems with stylometric features.

The best system combines token lengths, function words, sentence lengths, part-of-speech $F$4-grams, character $F$5-grams, and Distributional Random Oversampling. It achieves $F$6, correctly classifying all 16 Dantean texts and 313 of 314 non-Dantean texts. The sole false positive is an epistle by Giovanni Boccaccio.

Applied to the Questio, the system classifies it as Dantean in all ten stochastic DRO runs, with median posterior

$F$7

Dante is also the leading author in 15-author and 38-author attribution experiments, and Monarchia is the text most similar to the Questio under cosine similarity. The findings provide strong endogenous stylistic evidence but do not logically exclude an unrepresented author, a highly successful imitator, edition effects, or residual genre and topic confounding. Model posterior confidence is not equivalent to an assumption-free historical probability.

Across these domains, the shared name AQuA is primarily a naming coincidence. Some systems use the capitalization AQuA, others AQUA, and the expansions differ substantially. The appropriate interpretation therefore depends on the surrounding field: network architecture, question answering, quality assessment, thermodynamics, quantum software engineering, copyright protection, authorship verification, accelerator physics, or quantitative finance.