ZEUS: Diverse Applications in Science & Tech

• ZEUS is a recurring designation encompassing precise systems such as collider experiments, submillimeter spectrometers, banking trojans, and computational optimizers.
• The applications of ZEUS span high-energy physics, far-IR astronomy, cybersecurity, Bayesian inference, DNN training, distributed transactions, and autonomous driving.
• ZEUS emphasizes operational effectiveness through advanced methodologies like real-time adaptation, ensemble sampling, and locality-aware data management across domains.

ZEUS is a recurring designation in modern research literature rather than a single technical object. It refers, in different contexts, to the HERA collider experiment and collaboration in high-energy physics, the ZEUS-2 submillimeter spectrometer for far-infrared line astronomy, the Zeus or Zbot banking Trojan, and a wide family of computational systems for Bayesian inference, GPU-energy optimization, distributed transactions, video analytics, tabular clustering, numerical optimization, diffusion acceleration, time-series foundation modeling, and autonomous driving (Glasman, 2010, Ferkinhoff et al., 2010, Grammatikakis et al., 2021, You et al., 2022, Karamanis et al., 2021, Burnett et al., 2020). As used across these works, the name is typically attached to systems intended for precise measurement, efficient inference, or robust real-world operation.

1. ZEUS as a high-energy physics experiment at HERA

ZEUS was one of the two major collider experiments at HERA, where electrons or positrons of $27.5~\mathrm{GeV}/c$ collided with protons of $920~\mathrm{GeV}/c$, giving $\sqrt{s}=319~\mathrm{GeV}/c$; each of ZEUS and H1 collected about $0.5~\mathrm{fb}^{-1}$ of data (Hüttmann, 2010). In the jet-physics program, ZEUS used neutral-current deep inelastic $ep$ scattering in the Breit frame to test perturbative QCD, extract $\alpha_s$, and constrain proton PDFs. The measurements included dijet cross sections from $374~\mathrm{pb}^{-1}$ and inclusive-jet cross sections from $300~\mathrm{pb}^{-1}$, with comparisons based on NLO QCD and with modern jet algorithms such as $k_T$, anti-$k_T$, and SIScone (Glasman, 2010).

A central precision result of that program was the inclusive-jet determination

$920~\mathrm{GeV}/c$0

obtained from $920~\mathrm{GeV}/c$1 data, with a total uncertainty stated to be about $920~\mathrm{GeV}/c$2 and direct consistency with the QCD running of $920~\mathrm{GeV}/c$3 (Glasman, 2010). ZEUS also emphasized PDF sensitivity, especially to the gluon density at mid- to high-$920~\mathrm{GeV}/c$4, including in phase space where the predicted gluon fraction remained about $920~\mathrm{GeV}/c$5 at low $920~\mathrm{GeV}/c$6 and about $920~\mathrm{GeV}/c$7 around $920~\mathrm{GeV}/c$8 (Glasman, 2010).

In diffraction, ZEUS performed an NLO QCD extraction of diffractive PDFs using inclusive diffractive DIS data together with diffractive dijet DIS data. The analysis used $920~\mathrm{GeV}/c$9, fixed $\sqrt{s}=319~\mathrm{GeV}/c$0, and a proton-vertex-factorized ansatz

$\sqrt{s}=319~\mathrm{GeV}/c$1

showing that inclusive data constrain the quark density well while the dijet data are essential for fixing the diffractive gluon at high $\sqrt{s}=319~\mathrm{GeV}/c$2 (Wing, 2010). The preferred ZEUS fit also described dijet photoproduction across $\sqrt{s}=319~\mathrm{GeV}/c$3, with no discernible direct/resolved discrepancy in that dataset (Wing, 2010).

ZEUS also contributed to indirect searches for new physics in $\sqrt{s}=319~\mathrm{GeV}/c$4 collisions. In high-$\sqrt{s}=319~\mathrm{GeV}/c$5 neutral-current DIS, it found no deviation from the Standard Model and set 95% CL contact-interaction limits of $\sqrt{s}=319~\mathrm{GeV}/c$6 to $\sqrt{s}=319~\mathrm{GeV}/c$7 across 19 helicity models, as well as an ADD large-extra-dimensions limit $\sqrt{s}=319~\mathrm{GeV}/c$8 (Hüttmann, 2010). Later studies of ZEUS high-$\sqrt{s}=319~\mathrm{GeV}/c$9 data used exact Poisson-likelihood comparisons of predicted and observed event counts up to $0.5~\mathrm{fb}^{-1}$0, showing that modern PDF sets differ substantially in the large-$0.5~\mathrm{fb}^{-1}$1 region and arguing that ZEUS high-$0.5~\mathrm{fb}^{-1}$2 data should enter future PDF fits through forward-folded event-count likelihoods rather than standard Gaussian treatments (Collaboration, 2020).

After data taking ended, the ZEUS long-term data preservation project converted the analysis environment to Common Ntuples in ROOT and PAW, preserved about $0.5~\mathrm{fb}^{-1}$3 across about $0.5~\mathrm{fb}^{-1}$4 million files, retained metadata via CNINFO, event display capability via ZeVis, and detector-simulation/reconstruction capability via ZMCSP (Verbytskyi, 2016). This preserved resource was motivated by the uniqueness of HERA as the world’s only electron–proton collider and by the continued utility of ZEUS data for proton structure, diffraction, jets, event shapes, electroweak measurements, and future theory validation (Verbytskyi, 2016).

2. ZEUS-2 as a submillimeter spectrometer

ZEUS-2, the second-generation z(Redshift) and Early Universe Spectrometer, is a ground-based submillimeter/far-IR long-slit echelle grating spectrometer designed to detect faint, broad emission lines from galaxies, especially redshifted far-infrared cooling lines such as [CII] $0.5~\mathrm{fb}^{-1}$5 (Ferkinhoff et al., 2010). It is a moderate-resolution instrument with $0.5~\mathrm{fb}^{-1}$6, more precisely spanning roughly $0.5~\mathrm{fb}^{-1}$7 to $0.5~\mathrm{fb}^{-1}$8 by band, and covers seven telluric windows from $0.5~\mathrm{fb}^{-1}$9 to $ep$0 (Ferkinhoff et al., 2010).

Its detector system uses three transition-edge-sensed bolometer arrays built at NIST Boulder, totaling 555 pixels, with readout through a three-stage SQUID multiplexing system and Multi-Channel Electronics from the University of British Columbia (Ferkinhoff et al., 2010). The cryogenic chain replaces liquid cryogens with a Cryomech PT407 pulse-tube cooler and a dual-stage Janis ADR, reaching about $ep$1 and holding that temperature for about one day (Ferkinhoff et al., 2010). The instrument is designed for the CSO, APEX, and possibly JCMT, and the paper states improvements over the original ZEUS of about $ep$2 sensitivity at $ep$3, $ep$4 at $ep$5, more than a factor of 2 increase in accessible redshift range, and about a $ep$6 improvement in mapping speed (Ferkinhoff et al., 2010).

Scientifically, ZEUS-2 was optimized for the history of star formation across cosmic time, with particular emphasis on [CII], [NII], [OIII], [OI], [CI], and mid-$ep$7 CO diagnostics (Ferkinhoff et al., 2010). Its first-light results, obtained after commissioning on APEX in November 2012, included a $ep$8 detection of the [CII] $ep$9 line from H-ATLAS J091043.1-000322 (SDP11) at $\alpha_s$0, with flux

$\alpha_s$1

using 66 minutes on source (Ferkinhoff et al., 2013). Combined with a Herschel-PACS [OI] $\alpha_s$2 flux of $\alpha_s$3, the line ratios were interpreted with PDR models as implying $\alpha_s$4, gas density of order a few $\alpha_s$5, and a compact emitting diameter of roughly $\alpha_s$6–$\alpha_s$7, consistent with a high-redshift analogue of a local ULIRG and likely a compact merger-driven starburst (Ferkinhoff et al., 2013).

A common misconception is that ZEUS-2 is related to the ZEUS HERA experiment. The literature does not support that linkage: ZEUS-2 is an astronomical spectrometer with a distinct expansion of the name and a wholly different scientific program (Ferkinhoff et al., 2010, Ferkinhoff et al., 2013).

3. Zeus as a banking Trojan

In cybersecurity literature, Zeus, also known as Zbot, denotes a banking Trojan toolkit rather than a single immutable binary. It is described as a crimeware toolkit that allowed operators to build tailored Windows banking Trojans and manage botnets through a conventional client-server architecture; the paper cites 3.6 million infections in the United States in 2009 and notes that the 2011 source-code leak seeded later variants such as Citadel, GameoverZeus, Ice IX, and KINS (Grammatikakis et al., 2021).

Its core behavior combined credential theft and centralized command-and-control. The paper attributes to Zeus HTML injection against predefined banking sites, transparent redirects, keylogging, screenshot capture, interception of FTP and POP3 credentials, and certificate installation (Grammatikakis et al., 2021). Communications occurred over HTTP, with /config.bin retrieved by GET and status reports sent to /gate.php by POST, all RC4-encrypted with a botnet-specific key (Grammatikakis et al., 2021). Zeus version 2.0.8.9 is described as unpacking itself, copying into a randomly named folder under %AppData%, creating a unique mutex, writing a randomly named registry key under HKCU\Software\Microsoft, injecting into processes, hooking APIs, and leaving resident threads in processes such as taskhost.exe or explorer.exe to handle network communication and monitoring (Grammatikakis et al., 2021).

The malware literature also treats Zeus as a behaviorally coherent family that can be recognized from execution traces. A dynamic-analysis study used 1,980 Zeus samples, split across training and testing with 1,000 non-Zeus malware samples in each partition, and represented each run by 65 behavioral features derived from file system, registry, and network activity (Mohaisen et al., 2013). Those features included created, modified, and deleted files; file-size quartiles; registry-key creation and modification; unique destination IPs; request and reply size quartiles; TCP/UDP/RAW usage; HTTP request types; and DNS record types (Mohaisen et al., 2013). In that study, SVM gave the best overall balance, while L1-regularized logistic regression produced the lowest Zeus false-negative rate in the first train/test split; the abstract summarizes the general outcome as classification accuracy “in some cases as high as 95%” (Mohaisen et al., 2013).

A second line of work studied network-based mitigation rather than classification alone. A BAG/POMDP-based intrusion-response system tested on Zeus version 2.0.8.9 observed seven alerts when the bot executable traversed the gateway in one scenario and seventeen alerts when an already infected host attempted C&C communication in another, then isolated the Windows host by installing general iptables drop rules for 192.168.0.17, thereby blocking Zeus communications and preventing data leakage in the testbed (Grammatikakis et al., 2021).

4. ZEUS in statistical inference and optimization

Several computational works use ZEUS for systems that reduce tuning effort or improve efficiency in inference and optimization. In Bayesian computation, the Python package zeus implements Ensemble Slice Sampling, a gradient-free MCMC method intended for correlated, non-Gaussian, and multimodal posteriors common in astronomy and cosmology (Karamanis et al., 2021). The method uses slice sampling along ensemble-defined directions rather than Metropolis accept/reject proposals, is described as requiring only 1–2 hyperparameters that are often trivial to set, and is reported to scale to 1000s of CPUs (Karamanis et al., 2021). In the paper’s two principal scientific applications, ZEUS achieved about $\alpha_s$8 better efficiency than emcee/AIES in a 22-parameter cosmology problem and about $\alpha_s$9 better efficiency in a 14-parameter exoplanet radial-velocity problem (Karamanis et al., 2021).

In systems optimization for machine learning, Zeus is an online framework for recurring DNN training jobs that jointly tunes mini-batch size and GPU power limit (You et al., 2022). It defines energy-to-accuracy as

$374~\mathrm{pb}^{-1}$0

and optimizes a hybrid energy/time objective through just-in-time power profiling and Gaussian Thompson sampling over batch sizes (You et al., 2022). The reported gains relative to the default configuration are reductions in training energy consumption of $374~\mathrm{pb}^{-1}$1–$374~\mathrm{pb}^{-1}$2 across evaluated workloads, with the paper also emphasizing negligible profiling overhead and adaptation to drift through a sliding-window posterior update (You et al., 2022).

A different 2026 work introduces ZEUS as a hybrid GPU optimization method combining PSO, BFGS, automatic differentiation, and GPUs for high-dimensional non-convex minimization (Soos et al., 22 Jan 2026). Its workflow uses a small number of PSO iterations to improve starting points, then launches many independent BFGS refinements in parallel with forward-mode automatic differentiation for exact machine-level gradients (Soos et al., 22 Jan 2026). The reported performance is roughly $374~\mathrm{pb}^{-1}$3 to $374~\mathrm{pb}^{-1}$4 faster than the serial implementation, with the paper arguing that a handful of PSO iterations can materially improve global convergence on functions such as Rastrigin while GPUs make multistart optimization practical at scale (Soos et al., 22 Jan 2026).

These ZEUS systems are unrelated in implementation or domain. What they share is a design preference for online adaptation, reduced manual tuning, and efficient exploitation of limited high-value information—ensemble geometry in MCMC, just-in-time profiling in DNN training, or swarm-informed multistart search in nonlinear optimization.

5. ZEUS in data systems and AI pipelines

The name ZEUS also appears in several systems papers on data management and machine learning deployment. In distributed systems, Zeus: Locality-aware Distributed Transactions is an in-memory datastore that avoids complex fully distributed transaction protocols whenever possible by moving or acquiring all objects touched by a transaction onto one server, then executing a single-node transaction locally (Katsarakis et al., 2021). The system remains strongly consistent and fault-tolerant, with a reliable dynamic object-sharding protocol that can move 250K objects per second per server and, according to the paper, process millions of transactions per second on locality-heavy workloads (Katsarakis et al., 2021).

In video analytics, ZEUS is a reinforcement-learning-based system for action localization in long untrimmed videos (Chunduri et al., 2021). It adapts the next segment’s sampling rate, segment length, and resolution using a DQN policy over proxy features produced by an R3D-18-based action module, and trains that policy with an accuracy-aware aggregate reward tied to a user-specified target (Chunduri et al., 2021). Across three datasets, the paper reports speedups of up to $374~\mathrm{pb}^{-1}$5 over frame-based filtering methods and $374~\mathrm{pb}^{-1}$6 over window-based alternatives while consistently meeting the requested accuracy target (Chunduri et al., 2021).

For unsupervised learning on tables, ZEUS: Zero-shot Embeddings for Unsupervised Separation of Tabular Data is a transformer-based encoder pretrained on synthetic latent-variable mixture datasets and then applied to unseen tabular datasets without additional training or fine-tuning (Marszałek et al., 15 May 2025). It outputs embeddings $374~\mathrm{pb}^{-1}$7 shaped to support downstream clustering through a GMM-like latent geometry with centroid compactness and separation regularizers (Marszałek et al., 15 May 2025). On the 34 real datasets in its evaluation, ZEUS attained an average ARI of 57.43, higher than k-means, GMM, DEC, IDEC, and G-CEALS in the reported table, while remaining much faster than deep clustering baselines because inference requires only a forward pass plus a simple clustering step (Marszałek et al., 15 May 2025).

In generative modeling, ZEUS: Accelerating Diffusion Models with Only Second-Order Predictor is a training-free acceleration method for ODE-based denoising models (Wang et al., 2 Apr 2026). It predicts skipped denoiser outputs using the second-order extrapolator

$374~\mathrm{pb}^{-1}$8

then stabilizes longer skip runs with an interleaved schedule that alternates extrapolated and reused values (Wang et al., 2 Apr 2026). The paper argues that beyond $374~\mathrm{pb}^{-1}$9 acceleration, aggressive step skipping leaves at most one fresh evaluation per local window, so higher-order predictors amplify error without access to genuinely new information (Wang et al., 2 Apr 2026). Across image and video generation, it reports up to $300~\mathrm{pb}^{-1}$0 end-to-end speedup while maintaining perceptual quality (Wang et al., 2 Apr 2026).

In time-series modeling, Zeus: Towards Tuning-Free Foundation Model for Time Series Analysis is a 100M-parameter encoder-only TSFM with context length 4096, point-wise tokenization, a U-shaped multi-scale Transformer, and Multi-Objective Temporal Masking (Fu et al., 2 Jul 2026). Pretrained on about 300B observations from real and synthetic corpora, it supports point forecasting, probabilistic forecasting, imputation, anomaly detection, and classification in a tuning-free setting without task-specific parameter updates (Fu et al., 2 Jul 2026). The paper reports state-of-the-art or near-state-of-the-art zero-shot results across these tasks, including strong gains in imputation and anomaly detection relative to prior TSFMs and several task-specific baselines (Fu et al., 2 Jul 2026).

6. ZEUS as an autonomous vehicle system

In autonomous driving, Zeus denotes aUToronto’s Chevrolet Bolt platform for the SAE AutoDrive Challenge (Burnett et al., 2020). The Year 2 event was held in June 2019 at MCity, and the paper reports that ZEUS placed first in each dynamic challenge and first overall with 885 points (Burnett et al., 2020). The system was notable not only for competition performance but also for the team context: aUToronto had close to 100 students, mostly undergraduates, and the platform was developed under strict competition constraints, limited onsite practice, and Intel-only computing sponsorship (Burnett et al., 2020).

The vehicle carried a Velodyne HDL-64, four Blackfly S monocular cameras, a Novatel PwrPak7 GPS/IMU with Terrastar subscription, and a Continental ARS430 radar, although the radar was not used during the Year 2 competition (Burnett et al., 2020). Compute was provided by two Intel Xeon E5-2699 v4 processors and an Intel Arria 10 FPGA, with software built in C++ on ROS Kinetic and perception accelerated through OpenVINO (Burnett et al., 2020). This hardware choice drove a lightweight perception stack centered on SqueezeDet rather than more expensive CNN detectors (Burnett et al., 2020).

ZEUS’s perception and planning modules were explicitly modular. The system used a custom C++ inference library called Zeus DLA, a multi-object pipeline called aUToTrack combining 2D detections, LiDAR point projection, Euclidean clustering, and Kalman filtering, a semantic-map-driven global planner, a local path optimizer over centerlines and turn arcs, and a velocity planner that inserted zero-speed control points or pseudo-obstacles for red lights, stop signs, pedestrians, and railroad crossings (Burnett et al., 2020). Traffic-light handling relied on semantic-map association and a conservative default-red logic, while parking behavior used map-defined spots, occupancy estimates from point counts, and dedicated pull-in planning (Burnett et al., 2020).

Control used nonlinear MPC with steering and torque as outputs, a feed-forward acceleration-to-torque term, and an integral correction for acceleration error (Burnett et al., 2020). The paper reports that ZEUS generally stayed within 15 cm of the desired centerline and drove smoothly even at 40 km/h (Burnett et al., 2020). Its competition postmortem also documents failure modes, including temporary deadlock at an intersection due to missed green-light detections and a dynamic-deer collision in the MCity Challenge, illustrating that ZEUS was a robust but still competition-bounded autonomous-driving system rather than a fully general urban autonomy stack (Burnett et al., 2020).

Across these disparate literatures, ZEUS is best understood as a recurrent project name rather than a unified technical lineage. The HERA experiment, the ZEUS-2 spectrometer, the Zeus banking Trojan, and the many computational and autonomous systems that bear the name share neither common architecture nor common domain. What they do share, as reflected in the papers, is an emphasis on operational effectiveness under concrete constraints: collider precision, line-sensitivity in the submillimeter, credential theft and command-and-control in malware, and algorithmic efficiency or deployment practicality in modern computing systems.