Zeus: A Polysemous Research Label

• Zeus is a polysemous research label used in domains such as high-energy physics, cybersecurity, and machine learning, each with its own system design and focus.
• In high-energy physics, ZEUS refers to the HERA electron–proton collider experiment, whose preserved datasets continue to support analyses of proton structure and QCD phenomena.
• In cybersecurity and ML systems, Zeus denotes both a notorious banking Trojan and advanced frameworks for GPU optimization, video analytics, and tuning-free modeling.

In the cited research literature, Zeus denotes several unrelated scientific and technical entities rather than a single subject. The name appears as ZEUS, the HERA electron–proton collider experiment and its physics program; as Zeus or Zbot, a banking Trojan; and as the name of systems for GPU-efficient DNN training, zero-shot clustering of tabular data, tuning-free time-series modeling, diffusion acceleration, distributed transactions, Bayesian inference, numerical optimization, video analytics, and autonomous driving (Verbytskyi, 2016, Mohaisen et al., 2013, You et al., 2022, Marszałek et al., 15 May 2025, Fu et al., 2 Jul 2026).

1. Scope of the term in research usage

A common source of ambiguity is that the same label refers both to a major high-energy-physics experiment and to multiple acronymic or project names in computer science and engineering. In the cited literature, the term is therefore best understood as a polysemous research label whose meaning is domain-specific.

Domain	Zeus denotes	Representative reference
High-energy physics	HERA $ep$ experiment and data products	(Verbytskyi, 2016)
QCD and PDF analysis	ZEUS diffractive, jet, and high-$x$ studies	(Wing, 2010)
Beyond-the-Standard-Model searches	ZEUS and H1/ZEUS searches in $ep$ collisions	(Hüttmann, 2010)
Cybersecurity	Zeus/Zbot banking Trojan and its detection/mitigation	(Mohaisen et al., 2013)
ML systems	GPU-energy optimization and action-query processing systems	(You et al., 2022)
Foundation models	Zero-shot tabular and tuning-free time-series models	(Marszałek et al., 15 May 2025)
Systems and engineering	MCMC, optimization, distributed datastore, autonomous vehicle	(Karamanis et al., 2021)

This distribution of meanings suggests that “Zeus” functions less as a stable concept than as a recurrent project name attached to systems that emphasize generality, efficiency, or broad applicability. A plausible implication is that interpretation should begin from disciplinary context, not from the name alone.

2. ZEUS at HERA: experiment, detector, and preserved scientific infrastructure

In high-energy physics, ZEUS was one of the major experiments at HERA, the Hadron-Elektron Ring Anlage at DESY in Hamburg, and HERA was the world’s only electron(positron)-proton collider (Verbytskyi, 2016). HERA ran from 1991 to 2007, with a shutdown in 2001–2002 for a major upgrade; the electron or positron beam energy was about 27.5 GeV, while the proton beam energy was 820 GeV and 920 GeV in HERA-I and 920 GeV, 575 GeV, and 460 GeV in HERA-II (Verbytskyi, 2016). ZEUS itself was a general-purpose detector of about $$12\,\mathrm{m} \times 11\,\mathrm{m} \times 20\,\mathrm{m}$$ and roughly 3600 tons, with charged-particle tracking over $7^\circ < \theta < 170^\circ$, a depleted-uranium calorimeter, muon tracking, forward detectors, and a luminosity monitor (Verbytskyi, 2016).

The long-term value of ZEUS is tied to the uniqueness of the collision system and to the scale of the dataset. The experiment recorded more than 360 billion $ep$ collision events, corresponding to close to $0.5\,\mathrm{fb}^{-1}$ (Verbytskyi, 2016). Because no later facility reproduced the same $ep$ environment, the collaboration undertook a dedicated preservation program centered on continued access to data, documentation, and executable analysis workflows rather than bit preservation alone (Verbytskyi, 2016).

The preservation strategy replaced dependence on Mini Data Summary Tapes (MDST) and the ADAMO format with Common Ntuples (CN) stored as ROOT files, with PAW files with identical information also produced (Verbytskyi, 2016). These CN files were designed to support ongoing and future analyses while reducing reliance on a large experiment-specific software stack. The preserved holdings include approximately $250\,\mathrm{TB}$ in about 1.1 million files, stored at both DESY and MPCDF with disk and tape replicas (Verbytskyi, 2016). ZEUS also preserved the ZeVis event display, the CNINFO metadata catalogue in SQLite3 and plain text, and the ZEUS MC Standalone Package (ZMCSP) for detector simulation and reconstruction from external event records in HEPMC2, HEPMC3, and HEPEVT (Verbytskyi, 2016).

The scientific significance of this preservation effort lies in continued reuse of ZEUS data for proton structure, diffraction, jets, event shapes, photon structure, electroweak measurements, and searches for instantons, as well as reinterpretation with improved N(N)LO theory and modern Monte Carlo tools (Verbytskyi, 2016). In that sense, ZEUS persists not only as a historical detector but as a maintained analysis ecosystem.

3. QCD, PDFs, and new-physics searches using ZEUS data

A substantial part of the ZEUS scientific legacy is methodological: the data were repeatedly used to stress-test perturbative QCD, extract parton densities, and constrain short-distance new-physics scenarios. In the diffractive sector, ZEUS performed a next-to-leading-order QCD extraction of diffractive parton distribution functions (DPDFs) from inclusive diffractive DIS and diffractive DIS dijet data, using the factorized form

$$d\sigma^{ep\to eXp} = \sum_i f_i^D \otimes d\hat{\sigma}^{\gamma^* i}.$$

The analysis used a starting scale $x$0, fixed $x$1, and found that the perturbative framework describes the ZEUS diffractive data only for $x$2 (Wing, 2010). Inclusive diffractive measurements constrained the quark sector well but left the gluon density ambiguous at high $x$3; adding diffractive DIS dijets in fit SJ sharply constrained the gluon and disfavoured the high-$x$4 behavior allowed by fit S (Wing, 2010). The resulting ZEUS DPDFs were broadly compatible with H1, differed by about 10% for $x$5, and successfully described diffractive dijet photoproduction and diffractive charm DIS, even though they did not remove the much larger factorization discrepancy with Tevatron diffractive data (Wing, 2010).

ZEUS jet measurements in neutral-current DIS served a complementary role by testing pQCD precision and determining $x$6. Dijet cross sections were measured with $x$7 in $x$8 and $x$9, using the longitudinally invariant inclusive $ep$0 algorithm in the Breit frame, with $ep$1 and $ep$2 (Glasman, 2010). Inclusive-jet cross sections used $ep$3 for $ep$4, and NLO calculations with DISENT and NLOJET++ described the data very well (Glasman, 2010). ZEUS extracted

$ep$5

with total uncertainty about 3.5%, and found consistent values with anti-$ep$6, SIScone, and $ep$7 algorithms (Glasman, 2010). The measured algorithm-to-algorithm differences stayed below about 3.2% as a function of $ep$8 and below about 3.6% as a function of $ep$9, while the dijet phase space retained substantial gluon sensitivity, with predicted gluon fractions of about 75% at low $$12\,\mathrm{m} \times 11\,\mathrm{m} \times 20\,\mathrm{m}$$0 and about 60% by $$12\,\mathrm{m} \times 11\,\mathrm{m} \times 20\,\mathrm{m}$$1 (Glasman, 2010).

At large Bjorken $$12\,\mathrm{m} \times 11\,\mathrm{m} \times 20\,\mathrm{m}$$2, ZEUS revisited published neutral-current DIS event counts that had not been used in PDF fits. Using $$12\,\mathrm{m} \times 11\,\mathrm{m} \times 20\,\mathrm{m}$$3 of $$12\,\mathrm{m} \times 11\,\mathrm{m} \times 20\,\mathrm{m}$$4 data and $$12\,\mathrm{m} \times 11\,\mathrm{m} \times 20\,\mathrm{m}$$5 of $$12\,\mathrm{m} \times 11\,\mathrm{m} \times 20\,\mathrm{m}$$6 data over $$12\,\mathrm{m} \times 11\,\mathrm{m} \times 20\,\mathrm{m}$$7 and $$12\,\mathrm{m} \times 11\,\mathrm{m} \times 20\,\mathrm{m}$$8, the collaboration argued that low-count bins require a Poisson likelihood rather than a Gaussian $$12\,\mathrm{m} \times 11\,\mathrm{m} \times 20\,\mathrm{m}$$9 treatment (Collaboration, 2020). Modern NNLO PDF sets differed by more than a factor of 2 at the largest $7^\circ < \theta < 170^\circ$0, with spread larger than the quoted one-sigma uncertainties of HERAPDF2.0 and NNPDF3.1, and the paper proposed explicit schemes for incorporating the ZEUS high-$7^\circ < \theta < 170^\circ$1 data into future extractions (Collaboration, 2020).

ZEUS data also underpinned direct searches for new physics in $7^\circ < \theta < 170^\circ$2 collisions. In contact-interaction analyses, ZEUS observed no deviations from the SM prediction for NC DIS and set 95% CL lower limits on compositeness scales $7^\circ < \theta < 170^\circ$3 from 3.8 TeV to 8.9 TeV for 19 different helicity structures (Hüttmann, 2010). In ADD large-extra-dimension searches, ZEUS set $7^\circ < \theta < 170^\circ$4 for both signs of the effective coupling (Hüttmann, 2010). In combined H1+ZEUS exclusive channels, the isolated-lepton-plus-missing-$7^\circ < \theta < 170^\circ$5 analysis at $7^\circ < \theta < 170^\circ$6 observed 23 events versus $7^\circ < \theta < 170^\circ$7 expected, and the multi-lepton channel at $7^\circ < \theta < 170^\circ$8 observed 7 events versus $7^\circ < \theta < 170^\circ$9 expected, but the overall interpretation remained one of broad consistency with the Standard Model (Hüttmann, 2010).

4. Zeus in cybersecurity: banking Trojan, behavioral classification, and network containment

In cybersecurity, Zeus or Zbot designates a banking Trojan and crimeware toolkit that became historically influential in the late 2000s and early 2010s. It is described as having caused 3.6 million infections in the United States in 2009, as bundling a Windows bot builder with PHP-based command-and-control components, and as becoming the progenitor of descendants including Citadel, Ice IX, KINS, and GameoverZeus after the public leak of the Zeus 2.0.8.9 source code in 2011 (Grammatikakis et al., 2021). The malware copies itself under a randomly named directory in %AppData%, creates a randomly named registry key under HKCU\Software\Microsoft, injects into modifiable processes such as taskhost.exe or explorer.exe, encrypts traffic and internal data with RC4 and XOR-based obfuscation, and communicates through a simple centralized HTTP pattern: GET /config.bin and POST /gate.php (Grammatikakis et al., 2021). The same paper frames Zeus as an earlier banking Trojan whose anti-analysis methods were comparatively simple when contrasted with later malware such as Emotet (Grammatikakis et al., 2021).

A separate line of work studied Zeus as a supervised malware-classification target using runtime behavior rather than static signatures. One dataset contained 1,980 Zeus samples executed in a VM-based sandbox called auto-mal for 1 minute, with artifacts logged to MySQL and summarized into 65 behavior features spanning file system, registry, and network activity (Mohaisen et al., 2013). The reported classification task was explicitly binary Zeus vs. non-Zeus, with a training set of 1001 Zeus + 1000 non-Zeus and a testing set of 979 Zeus + 1000 non-Zeus (Mohaisen et al., 2013). The evaluated classifiers were SVM, logistic regression with L1 and L2 regularization, decision trees, and KNN, implemented in mlpy (Mohaisen et al., 2013). In the train-on-A/test-on-B experiment, SVM yielded Zeus false positive/false negative rates of 6.84% / 4.29%, logistic L1 gave 11.03% / 1.43%, decision trees gave 4.70% / 22.98%, and the authors concluded that SVM offered the best balanced performance whereas L1 logistic regression minimized missed Zeus samples (Mohaisen et al., 2013). The paper also stressed an important caveat: this was not named-family-vs-named-family taxonomy, but Zeus against a heterogeneous “other malware” class (Mohaisen et al., 2013).

Network-level mitigation work treated Zeus less as a labeling problem than as a containment target. An intelligent intrusion response system, iIRS, combined a Bayesian attack graph derived from MulVAL with a discrete-time POMDP, Suricata IDS alerts, and gateway-enforced iptables rules to isolate infected hosts (Grammatikakis et al., 2021). In the Zeus testbed, the system successfully blocked Zeus binary transfer and gate.php command-and-control traffic by isolating host 192.168.0.17, thereby disrupting network communication and preventing data leakage in the reported scenarios (Grammatikakis et al., 2021). The same study also made clear that its evidence was operational rather than benchmark-style: it did not report confusion matrices, ROC curves, or long-horizon adversarial-evasion measurements (Grammatikakis et al., 2021).

5. Zeus in machine-learning systems: training efficiency, video analytics, and diffusion acceleration

Several recent systems use the name Zeus for runtime optimization in machine learning. One such Zeus addresses GPU energy consumption during recurring DNN training jobs by jointly optimizing batch size $ep$0 and GPU power limit $ep$1. It defines energy-to-accuracy

$ep$2

and a user-weighted cost

$ep$3

then combines a just-in-time power-limit profiler with a Thompson-sampling multi-armed bandit over batch sizes (You et al., 2022). The system reports energy reductions of 15.3%–75.8% over a default baseline in its main V100 evaluation, cluster-scale reductions of 7%–52% on an Alibaba GPU trace, and on four A40 GPUs achieves 12% more time but 21% less energy than Pollux for DeepSpeech2 on LibriSpeech (You et al., 2022). Its design premise is that ETA and TTA lie on a Pareto frontier rather than being jointly minimized by the usual “max batch size + max power limit” heuristic (You et al., 2022).

A different ZEUS is a video analytics system for action localization in long videos. It trains a reinforcement-learning agent to alter the input to an action classifier along three dimensions—sampling rate, segment length, and resolution—so as to meet a user-specified accuracy target while maximizing throughput (Chunduri et al., 2021). The policy is trained with an accuracy-aware aggregate reward, and the system uses proxy features from an action-recognition backbone such as R3D-18 rather than lightweight frame filters (Chunduri et al., 2021). Across six queries on three datasets, the paper reports up to 22.1× speedup over frame-based filtering and up to 4.7× over window-based static methods while staying close to requested accuracy targets (Chunduri et al., 2021).

In generative modeling, ZEUS denotes a training-free diffusion-sampling accelerator that predicts skipped denoiser outputs with a second-order rule

$ep$4

and stabilizes consecutive skipping with an interleaved scheme that alternates this extrapolation with direct reuse of $ep$5 (Wang et al., 2 Apr 2026). The method is designed around the claim that, beyond $ep$6 acceleration, the fresh denoiser output and its backward difference are the only causally grounded local information, making higher-order predictors unnecessarily fragile (Wang et al., 2 Apr 2026). It adds essentially zero overhead, no feature caches, and no architectural modifications, and reports up to 3.2× end-to-end speedup with maintained perceptual quality across image and video generation (Wang et al., 2 Apr 2026).

6. Zeus as pretrained or zero-shot model family

The name ZEUS also appears in pretrained or tuning-free representation models. In unsupervised tabular clustering, ZEUS stands for “Zero-shot Embeddings for Unsupervised Separation of tabular data” and is a transformer-based model trained over synthetic datasets drawn from a latent-variable prior (Marszałek et al., 15 May 2025). It processes whole datasets as inputs, maps each record $ep$7 to an embedding $ep$8, and optimizes a mixture-inspired loss

$ep$9

with $0.5\,\mathrm{fb}^{-1}$0 (Marszałek et al., 15 May 2025). The architecture uses 12 attention blocks, 6 attention heads per block, token dimension 512, and GeLU activations; inference requires a single forward pass followed by a standard clustering method such as k-means or a GMM in embedding space (Marszałek et al., 15 May 2025). On averaged ARI scores, ZEUS reports 57.43 on real/OpenML datasets, 89.03 on synthetic Gaussian datasets, and 86.33 on synthetic transformed datasets, with the strongest gains on the harder transformed regime (Marszałek et al., 15 May 2025). The paper is equally explicit about its limitations: datasets are restricted to about 2000 samples and 30 features in the reported implementation, and performance depends on the match between the synthetic prior and real target structure (Marszałek et al., 15 May 2025).

A more recent Zeus is a tuning-free Time Series Foundation Model (TSFM) intended to handle forecasting, probabilistic forecasting, imputation, anomaly detection, and classification without task-specific fine-tuning (Fu et al., 2 Jul 2026). Architecturally it is an encoder-only multi-scale Transformer with point-wise tokenization and a U-shaped hierarchy over scales $0.5\,\mathrm{fb}^{-1}$1, paired with Multi-Objective Temporal Masking (MOTM) to jointly teach extrapolation, interpolation, and global abstraction (Fu et al., 2 Jul 2026). The model predicts quantiles $0.5\,\mathrm{fb}^{-1}$2, contains about 100M parameters, is pretrained on roughly 300B observations, and uses both real corpora and the synthetic Aegis-Syn dataset (Fu et al., 2 Jul 2026). Its self-attention complexity is reported as

$0.5\,\mathrm{fb}^{-1}$3

corresponding to roughly a 3.8× reduction in self-attention FLOPs (Fu et al., 2 Jul 2026). Empirically it reports 19 wins on MSE and 14 wins on MAE in long-term forecasting tables, MASE = 0.693 and CRPS = 0.480 on GIFT-Eval, average anomaly-detection adjusted F1 of 0.900, and tuning-free classification accuracy of 0.675 with 1-NN on frozen features, rising to 0.728 with linear probing (Fu et al., 2 Jul 2026). This suggests a broader shift in the use of the name Zeus toward “foundation-style” models that foreground zero-shot or tuning-free deployment.

7. Other technical systems named Zeus: optimization, inference, distributed transactions, and autonomy

Beyond ML infrastructure and foundation models, the name attaches to several technically distinct systems. In numerical optimization, ZEUS combines particle swarm optimization (PSO), BFGS, automatic differentiation, and GPUs for high-dimensional non-convex minimization (Soos et al., 22 Jan 2026). The PSO phase updates particles by

$0.5\,\mathrm{fb}^{-1}$4

with $0.5\,\mathrm{fb}^{-1}$5, $0.5\,\mathrm{fb}^{-1}$6, and $0.5\,\mathrm{fb}^{-1}$7, after which each refined start launches an independent BFGS run with gradients from forward-mode AD (Soos et al., 22 Jan 2026). The paper reports that a handful of PSO iterations improves global convergence, especially on multimodal functions, and that the CUDA implementation yields roughly 10× to 100× speedup over the serial version on the tested functions (Soos et al., 22 Jan 2026).

In Bayesian computation, zeus is a Python implementation of Ensemble Slice Sampling (ESS) for MCMC (Karamanis et al., 2021). It targets posterior inference with strong correlations, multimodality, and many available CPUs, and emphasizes minimal hand-tuning, walker-parallel execution, and robustness relative to affine-invariant ensemble Metropolis variants (Karamanis et al., 2021). The package requires at least $0.5\,\mathrm{fb}^{-1}$8 walkers and generally recommends $0.5\,\mathrm{fb}^{-1}$9–$ep$0, with $ep$1–$ep$2 suggested for multimodal or strongly nonlinear targets (Karamanis et al., 2021). In the reported applications, Zeus performs about 9× better than emcee/AIES in a cosmological inference problem and about 29× better in an exoplanet problem when efficiency is measured as independent samples per likelihood evaluation (Karamanis et al., 2021).

In distributed systems, Zeus is a locality-aware distributed in-memory datastore that provides general transactions by acquiring all objects involved in the transaction to the same server and then executing a single-node transaction there (Katsarakis et al., 2021). Instead of paying the cost of a fully general distributed transaction protocol on every request, it relies on a dynamic ownership protocol, asynchronously pipelined replication, and a reliable commit layer to exploit workloads with slowly changing locality (Katsarakis et al., 2021). The system can move 250K objects per second per server, process millions of transactions per second, and outperform FaRM, FaSST, and DrTM on several locality-heavy workloads, while preserving strong consistency and fault tolerance (Katsarakis et al., 2021).

Finally, in autonomous driving, Zeus is a 2017 Chevrolet Bolt converted by aUToronto into a self-driving research platform for the SAE AutoDrive Challenge (Burnett et al., 2020). The vehicle uses a Velodyne HDL-64, four Blackfly S monocular cameras, a Novatel PwrPak7 GPS/IMU, a Continental ARS430 radar, 2 × Intel Xeon E5-2699 v4 CPUs, and an Intel Arria 10 FPGA, reflecting a design constrained by the absence of onboard NVIDIA GPUs (Burnett et al., 2020). The software stack was built on ROS Kinetic, used SqueezeDet accelerated through OpenVINO and a custom Zeus DLA, and combined map-centric planning with NMPC control (Burnett et al., 2020). In the Year 2 competition, Zeus placed first in each dynamic challenge and achieved 885 / 1000, with the next-best team at 523, illustrating a version of “Zeus” whose significance derived not from abstraction alone but from reliable full-system integration (Burnett et al., 2020).

Across these uses, the name marks neither a shared architecture nor a shared theory. It instead indexes a recurring pattern in contemporary research naming: “Zeus” is repeatedly attached to systems that claim breadth, efficiency, or operational decisiveness, but the technical substance of each Zeus is entirely domain-dependent.