Mega: Multi-Domain Terminology in Research

• Mega is a polysemous term that functions as both an acronym for specialized methods (e.g., MEGA in cosmology and ML) and a descriptor for large-scale systems like mega-constellations and events.
• In computational research, Mega methods enhance performance by addressing limitations in transformer attention, improving speed and memory efficiency through novel algorithmic approaches.
• In observational and infrastructural studies, 'mega' labels massive datasets, constellations, or events, emphasizing challenges in scalability and whole-system optimization.

Searching arXiv for papers using the term “Mega” across domains to ground the article in current research usage. Mega is a polysemous research term used across astrophysics, machine learning, systems, hardware, security, and network science. In contemporary arXiv literature, it most often functions either as an acronym naming a specific method or survey, or as a scale descriptor denoting exceptionally large systems such as mega-constellations, mega-categories, mega-events, mega-kernels, or Mega-Hertz signals. The term therefore does not denote a single concept; rather, it indexes a family of domain-specific constructs whose common feature is either explicit acronymic branding or emphasis on scale. Representative uses include “Merger graphs of structure formation” for cosmological halo assembly (Roper et al., 2020), “Moving Average Equipped Gated Attention” for long-sequence modeling (Ma et al., 2022), “Merging Multiple Independently Trained Neural Networks Based on Genetic Algorithm” (Yun, 2024), medium-band and mid-infrared astronomical surveys named MEGA (Suess et al., 2024, Backhaus et al., 24 Mar 2025), and multiple studies of LEO mega-constellations (Cen et al., 2024, Wang et al., 2024, Reiland et al., 2020).

1. Mega as an acronymic naming pattern

In the literature surveyed here, “Mega” appears in several stylized forms—MEGA, Mega, and MeGA—each tied to a distinct expansion and technical agenda. In cosmology, MEGA denotes “Merger graphs of structure formation,” a framework that replaces merger trees with directed graphs in which haloes can both merge and split (Roper et al., 2020). In sequence modeling, Mega denotes “Moving Average Equipped Gated Attention,” a single-head gated attention mechanism equipped with an exponential moving average (Ma et al., 2022). In neural network fusion, MeGA denotes “Merging Multiple Independently Trained Neural Networks Based on Genetic Algorithm” (Yun, 2024). In cinematic video understanding, MEGA expands to “Multimodal alignmEnt aGgregation and distillAtion” (Sadoughi et al., 2023). In graph continual learning, MEGA denotes “Model-Agnostic Meta Graph Continual Learning” (Pang et al., 18 Apr 2025).

The same naming pattern also appears in astronomy. “Mega-Archive” and “Mega-Precovery” designate EURONEAR data-mining tools for astronomical image archives (Vaduvescu et al., 2019). “MEGA” also names the “MIRI EGS Galaxy and AGN” survey in the Extended Groth Strip (Backhaus et al., 24 Mar 2025). “MegaScience” is the shorthand used for “Medium Bands, Mega Science,” a JWST/NIRCam medium-band survey of Abell 2744 (Suess et al., 2024).

A second usage is scalar rather than acronymic. “Mega-constellations” refers to LEO satellite systems with hundreds to thousands of satellites (Cen et al., 2024, Wang et al., 2024, Reiland et al., 2020). “Mega-category” refers to OCR with label spaces on the order of tens of thousands up to 100K+ categories, exemplified by MegaHan97K with 97,455 categories (Zhang et al., 5 Jun 2025). “Mega-kernel” denotes a single persistent GPU kernel executing an entire tensor program end to end (Cheng et al., 22 Dec 2025). “Mega-Hertz” refers to the MHz gravitational-wave component predicted from first-order QCD phase transitions during neutron-star mergers (Casalderrey-Solana et al., 2022). “Mega events” refers to short-duration, extremely high-footfall gatherings operating as temporary, digitally enabled cities (Negi et al., 21 Jul 2025).

This distribution suggests that “Mega” functions in two stable ways across arXiv research: as an acronym for a concrete method or dataset, and as a marker of exceptional operational scale.

2. Computational and machine-learning uses

In sequence modeling, Mega is introduced as a response to two limitations of standard Transformer attention: weak inductive bias and quadratic time and memory complexity with sequence length (Ma et al., 2022). The mechanism injects a position-aware inductive bias through a multi-dimensional damped exponential moving average, then integrates that signal into a single-head gated attention unit. The paper defines the EMA recursion as

$$\mathbf{y}_t = \boldsymbol{\alpha} \odot \mathbf{x}_t + (1 - \boldsymbol{\alpha} \odot \boldsymbol{\delta}) \odot \mathbf{y}_{t-1},$$

and reports that Mega achieves average accuracy 88.21 on the Long Range Arena, with task-wise scores including ListOps 63.14, Text 90.43, Retrieval 91.25, Image 90.44, Pathfinder 96.01, and Path-X 97.98 (Ma et al., 2022). On 4K-input Text, Mega is reported as 2.9× faster and using 31% of Transformer peak memory, while Mega-chunk is 5.5× faster and uses 13% of memory (Ma et al., 2022). On WikiText-103, the reported perplexity is 18.07, and inference speed is 48k tokens/s versus 5.6k tokens/s for Transformer (Ma et al., 2022).

A different computational use appears in network fusion. MeGA searches over element-wise weight combinations of pre-trained models with a genetic algorithm using tournament selection, crossover, mutation, and elitism (Yun, 2024). The initialization and crossover are both element-wise linear combinations,

$$\theta_{i} = \alpha \theta_{1} + (1-\alpha)\theta_{2}, \qquad \theta_{\text{child}} = \beta \theta_{a} + (1-\beta)\theta_{b},$$

while mutation adds Gaussian noise with per-parameter probability $p_{\text{mut}} = 0.02$ (Yun, 2024). On CIFAR-10, the paper reports that naive weight averaging collapses to 0.010 test accuracy across multiple architectures, whereas MeGA reaches 0.822 on ResNet-56, 0.816 on ResNet-110, 0.819 on ResNet-152, 0.754 on Xception, 0.742 on DenseNet-121, and 0.753 on DenseNet-169 (Yun, 2024).

MEGA in cinematic long-video segmentation addresses multimodal temporal alignment and fusion for movies longer than 60 minutes (Sadoughi et al., 2023). Its key device is Alignment Positional Encoding with index

$$i_{\text{align}} = \left\lfloor \frac{L_n}{L} i \right\rfloor,$$

which normalizes positions to a coarse shared temporal grid across modalities (Sadoughi et al., 2023). The paper reports an Average Precision improvement of +1.19% on MovieNet scene segmentation and a Total Agreement improvement of +5.51% on TRIPOD act segmentation (Sadoughi et al., 2023).

MEGA in graph continual learning is likewise not about scale but about alignment across learning stages. “Model-Agnostic Meta Graph Continual Learning” introduces second-order gradient alignment for Graph Few-Shot Class-Incremental Learning under a support-only incremental protocol (Pang et al., 18 Apr 2025). The meta-update is

$$\theta \leftarrow \theta - \beta \nabla_\theta \sum_i L^{\text{outer}}_i,$$

with second-order dependence through the inner-loop updates (Pang et al., 18 Apr 2025). Reported gains include 74.56±1.79% at Task 1 on Amazon-Clothing, 43.53±1.88% at Task 1 on DBLP, and 52.74±3.89% at Task 1 on Cora-Full, with strong improvements over baselines under the paper’s rigorous protocol (Pang et al., 18 Apr 2025).

This body of work shows that “Mega” in machine learning is usually method branding rather than a shared architecture family. The commonality is not algorithmic but nominal: each paper attaches “Mega” to a specific intervention—EMA-equipped attention, genetic weight merging, multimodal temporal bottlenecks, or second-order meta-continual alignment.

3. Astronomical and cosmological uses

In cosmology, MEGA redefines halo assembly records as graphs rather than trees (Roper et al., 2020). A merger graph is described as a directed, time-ordered graph $G=(V,E)$ in which nodes are halos at specific snapshots and edges encode continuation, merger, or split relations. The progenitor and descendant sets are written as

$$P(v_{j,s+1}) = \{v_{i,s} : (v_{i,s}\to v_{j,s+1}) \in E\},$$

$$D(v_{i,s}) = \{v_{j,s+1} : (v_{i,s}\to v_{j,s+1}) \in E\}.$$

The central result is that allowing splits as well as merges yields smoother inferred mass growth histories and eliminates catastrophic failures in which massive haloes have no progenitors or descendants (Roper et al., 2020). The same work also states that two different density thresholds can be used to distinguish host haloes from higher-density subhaloes, interpreted as sites of galaxy formation (Roper et al., 2020).

A very different astronomical use is Mega-Archive and its associated EURONEAR tools (Vaduvescu et al., 2019). Mega-Archive is a metadata index of raw science images designed for mining Solar System objects and time-domain phenomena. By 23 February 2019 it indexed about 15 million images from 111 instrument archives across six major collections, with daily updates running since 2014 (Vaduvescu et al., 2019). Mega-Precovery allows three input modes—designation, orbit, or observations—and two ephemeris engines, Miriade and OrbFit (Vaduvescu et al., 2019). The survey also introduced FindCCD, FindCCD for Fixed Objects, MASFO, and MASDS, with supported mosaic cameras including Subaru-SuprimeCam, VST-OmegaCam, INT-WFC, VISTA-VIRCAM, CFHT-MegaCam, Blanco-DECam, and Subaru-HSC (Vaduvescu et al., 2019).

MEGA is also the name of a JWST/MIRI mid-infrared survey. The “MIRI EGS Galaxy and AGN” survey comprises 25 pointings over 70 arcmin$^2$ in the Extended Groth Strip, using F770W, F1000W, F1500W, and F2100W, with three pointings lacking F770W (Backhaus et al., 24 Mar 2025). The paper reports 5$\sigma$ point-source limits of 0.18, 0.41, 1.26, and 4.10 $$\theta_{i} = \alpha \theta_{1} + (1-\alpha)\theta_{2}, \qquad \theta_{\text{child}} = \beta \theta_{a} + (1-\beta)\theta_{b},$$0Jy in F770W, F1000W, F1500W, and F2100W respectively, and a final catalog of 4444 sources with S/N $$\theta_{i} = \alpha \theta_{1} + (1-\alpha)\theta_{2}, \qquad \theta_{\text{child}} = \beta \theta_{a} + (1-\beta)\theta_{b},$$1 3 in F770W (Backhaus et al., 24 Mar 2025). The survey emphasizes obscured star formation and AGN activity at “Cosmic Noon,” and preliminary results suggest $$\theta_{i} = \alpha \theta_{1} + (1-\alpha)\theta_{2}, \qquad \theta_{\text{child}} = \beta \theta_{a} + (1-\beta)\theta_{b},$$2 of MEGA galaxies host low-luminosity AGN at that epoch (Backhaus et al., 24 Mar 2025).

Related but distinct is “Medium Bands, Mega Science,” a JWST/NIRCam survey of Abell 2744 (Suess et al., 2024). MegaScience obtained about 30 arcmin$$\theta_{i} = \alpha \theta_{1} + (1-\alpha)\theta_{2}, \qquad \theta_{\text{child}} = \beta \theta_{a} + (1-\beta)\theta_{b},$$3 of NIRCam imaging and about 17 arcmin$$\theta_{i} = \alpha \theta_{1} + (1-\alpha)\theta_{2}, \qquad \theta_{\text{child}} = \beta \theta_{a} + (1-\beta)\theta_{b},$$4 of NIRISS parallel imaging, completing deep coverage in all NIRCam medium- and broad-band filters when combined with UNCOVER (Suess et al., 2024). The paper states that medium bands improve both the scatter and catastrophic outlier rate of photometric redshifts by factors of 2–3, and demonstrates spatially resolved [OIII] and continuum mapping in three spectroscopically confirmed $$\theta_{i} = \alpha \theta_{1} + (1-\alpha)\theta_{2}, \qquad \theta_{\text{child}} = \beta \theta_{a} + (1-\beta)\theta_{b},$$5 galaxies (Suess et al., 2024).

These uses share only the name. In one case, MEGA is a graph formalism for nonlinear structure growth; in the others, it denotes data infrastructure and survey programs in observational astronomy.

4. Mega as a scale descriptor in infrastructure, networks, and systems

In satellite networking, “mega-constellation” is a literal scale label. SatFlow studies LEO networks with hundreds to thousands of satellites interconnected with inter-satellite links (Cen et al., 2024). Its objective is joint planning of topology, traffic allocation, and fine-grained ISL terminal power allocation through a two-tier architecture: an upper-level multi-agent reinforcement learning module and a lower-level distributed alternating-step optimizer (Cen et al., 2024). The paper reports reductions of the flow violation ratio by up to 21.0% and total costs by up to 89.4% relative to state-of-the-art benchmarks, as well as throughput values of 177.2, 339.3, and 227.8 Mbps for three evaluated constellations (Cen et al., 2024).

A complementary work proposes a KPI framework for LEO mega-constellation satellite networks (Wang et al., 2024). The reference system comprises 1800 satellites in 60 orbital planes with 30 satellites per plane at altitude 508 km and inclination 55° (Wang et al., 2024). The paper introduces “interfering area” and “spherical geographic cells” as abstractions for tractable system-level evaluation, and reports achieved area traffic capacity around 4 Kbps/km$$\theta_{i} = \alpha \theta_{1} + (1-\alpha)\theta_{2}, \qquad \theta_{\text{child}} = \beta \theta_{a} + (1-\beta)\theta_{b},$$6, service availability ranging from 0.36 to 0.39, average access success probability approximate to 96%, and handover failure rate approximate to 10% under nearest-satellite association (Wang et al., 2024).

The scale meaning of “mega” also appears in collision analysis for LEO constellations (Reiland et al., 2020). Using FCC-filed OneWeb and Starlink designs, the paper reports 2,522 close approaches below 1 km for a OneWeb target-plane endogenous 90-day analysis in the nominal configuration, versus 232 under a MiSO configuration (Reiland et al., 2020). For OneWeb this corresponds to a 90.8% reduction in sub-kilometer endogenous close approaches, while minimum separation improves from 0.0064 km to 0.5502 km (Reiland et al., 2020). This usage is descriptive rather than acronymic: “mega-constellation” names a large orbital system whose scale creates new collision-risk regimes.

In GPU systems, “mega-kernel” has yet another meaning. “Mirage Persistent Kernel” defines a mega-kernel as a single persistent GPU kernel that executes an entire tensor program end to end, including inter-GPU communication (Cheng et al., 22 Dec 2025). The system introduces an SM-level task graph, a compiler lowering programs into SM-granularity tasks and events, and an in-kernel decentralized runtime (Cheng et al., 22 Dec 2025). Empirically, the paper reports up to 1.7× lower single-batch latency versus kernel-per-operator serving systems and 1.1–1.4× improvements on 8×H100 multi-GPU scale-out (Cheng et al., 22 Dec 2025). Here “mega” denotes fusion granularity rather than system size.

A plausible implication is that “mega” in systems papers often marks a shift from local optimization to whole-system optimization: whole constellations rather than single links, whole programs rather than individual kernels, and whole operational environments rather than isolated services.

5. Security, sensing, hardware, and scientific frontiers

In cybersecurity, “mega events” are defined as short-duration, extremely high-footfall gatherings functioning as temporary, digitally enabled cities (Negi et al., 21 Jul 2025). The MahaKumbh 2025 case involved 45 consecutive days and more than 600 million footfalls, ultimately about 660 million, with six special days exceeding 20 million visitors each (Negi et al., 21 Jul 2025). The oversight program covered 12 web and 5 mobile applications and tracked 130 technical findings: 14 Critical, 36 High, 48 Medium, 31 Low, and 1 Informational (Negi et al., 21 Jul 2025). The paper reports that none of the cyber attacks during the 45-day event was successful (Negi et al., 21 Jul 2025). In this setting, “mega” denotes extreme operational exposure, temporary infrastructure, and unusual dependence on OT/IT convergence.

In OCR, “mega-category” denotes classification with extremely large label spaces (Zhang et al., 5 Jun 2025). MegaHan97K contains exactly 97,455 categories and 4,614,675 character images, fully covering GB18030-2022 and extending to rare and variant characters (Zhang et al., 5 Jun 2025). The paper reports that synthetic training data improves all evaluated methods strongly, with ResNet50 rising from 34.89% to 88.76%, CCR-CLIP from 82.04% to 89.56%, and HierCode from 66.58% to 92.32% (Zhang et al., 5 Jun 2025). In the zero-shot setting over 69,922 unseen classes, CCR-CLIP reaches 79.04% (Zhang et al., 5 Jun 2025). This use of “mega” is purely categorical scale.

In neuromorphic hardware, Mega names a 22 nm convolutional spiking neural network accelerator (Luiken et al., 29 Jun 2026). Fabricated in GlobalFoundries 22 nm FDSOI, the chip achieves 0.375 pJ/SOP at 0.55 V and 155 MHz, and up to 148.7 GSOP/s at 1.1 V and 600 MHz (Luiken et al., 29 Jun 2026). The design centers on 9 compute clusters for 3×3 kernel-offset parallelism, 32 convolution units per cluster, unified memory for spikes, neuron states, and weights, and a spike streamer with 96-bit dense spike vectors and dual leading-zero counters (Luiken et al., 29 Jun 2026). Unlike the scale-denoting cases above, this is again acronymic branding attached to a concrete architecture.

In gravitational-wave astrophysics, “Mega-Hertz” is literal. A first-order QCD phase transition during neutron-star mergers could nucleate bubbles whose sound-wave dynamics generate a GW signal peaking around 0.6 MHz (Casalderrey-Solana et al., 2022). Using benchmark parameters, the paper derives a mean bubble separation of about 530 m and an observed characteristic strain

$$\theta_{i} = \alpha \theta_{1} + (1-\alpha)\theta_{2}, \qquad \theta_{\text{child}} = \beta \theta_{a} + (1-\beta)\theta_{b},$$7

with projected reach up to tens of Mpc for future superconducting radio-frequency cavity detectors in strong-transition cases (Casalderrey-Solana et al., 2022). Here “Mega” refers both to Mega-Hertz frequency and Mega-parsec distance reach.

Even planetary geomorphology uses the term descriptively. “Mega-yardangs” are large, wind-abraded linear ridges that can be confused with dunes in SAR imagery (Paillou et al., 2015). The paper reports that yardangs are on average about 10 dB brighter than dunes in X-band terrestrial analogs and uses this contrast to interpret Titan radar signatures (Paillou et al., 2015).

6. Conceptual synthesis and recurring misconceptions

The main misconception is that “Mega” names a unified framework across fields. The literature shows the opposite. MEGA in cosmology (Roper et al., 2020), Mega in attention mechanisms (Ma et al., 2022), MeGA in genetic network merging (Yun, 2024), MEGA in graph continual learning (Pang et al., 18 Apr 2025), and MEGA in mid-infrared survey science (Backhaus et al., 24 Mar 2025) are unrelated constructs that merely share a mnemonic label.

A second misconception is that “mega” always indicates large physical scale. This is true for mega-constellations (Cen et al., 2024, Wang et al., 2024, Reiland et al., 2020), mega-events (Negi et al., 21 Jul 2025), mega-category OCR (Zhang et al., 5 Jun 2025), mega-kernels (Cheng et al., 22 Dec 2025), Mega-Hertz GWs (Casalderrey-Solana et al., 2022), and mega-yardangs (Paillou et al., 2015), but not for several acronymic methods where the technical contribution is architectural or algorithmic rather than scalar (Ma et al., 2022, Sadoughi et al., 2023, Pang et al., 18 Apr 2025).

A third misconception is that acronymic capitalization signals shared lineage. In fact, stylization varies by author choice. “MEGA,” “Mega,” and “MeGA” correspond to unrelated phrases and communities. This suggests that disambiguation should always be domain-first: one must identify whether the term appears in cosmological structure formation, transformer alternatives, OCR datasets, JWST surveys, satellite networking, or cyber-resilience engineering before inferring meaning.

Taken together, the research record presents “Mega” as a cross-domain lexical hub rather than a concept. Its encyclopedic significance lies precisely in that multiplicity: it names graph-based halo assembly, EMA-equipped attention, multi-model weight evolution, massive OCR label spaces, medium-band and mid-infrared surveys, mega-constellation planning, persistent GPU kernels, temporary digital cities, MHz gravitational waves, and geomorphic macroforms. The term therefore functions less as a stable scientific object than as a recurring convention for labeling either extreme scale or a compact, memorable acronymic program.