MUTE: A Cross-Domain Disambiguation

• MUTE is a polysemous term describing a family of domain-specific designations ranging from datasets and neural methods to muography instruments and privacy controls.
• In multimodal benchmarks, MUTE encapsulates datasets and architectures for hateful meme detection and multi-unit transformer models that show measurable performance gains.
• Across disciplines like machine learning, muography, and privacy research, MUTE methods demonstrate innovative encoding, routing, and suppression techniques for enhanced system robustness.

Searching arXiv for the listed MUTE-related papers to ground the article in cited sources. Searching for MUTE on arXiv. MUTE is a recurrent but non-unified term in arXiv literature. Depending on domain, it denotes a multilingual or multimodal benchmark, a target-encoding method, a Transformer variant, a neural SLAM system, a digital-twin decision support system, a muography instrument, an underground muon-transport code, or the operational notion of muting in privacy and adversarial-audio settings. The term therefore functions less as a single concept than as a family of domain-specific designations whose meanings are fixed by local technical context.

1. Acronymic landscape and disambiguation

Across the cited literature, MUTE and related forms such as mute and MuTe refer to distinct research objects rather than a shared framework. In machine learning, MUTE can denote a dataset, an encoding scheme, or an architecture. In audio and HCI, mute denotes a privacy control or a suppression objective. In particle physics and muography, MuTe denotes a telescope, whereas MUTE denotes a computational code for underground muon intensities. This dispersion makes disambiguation mandatory in technical writing.

Designation	Field	Identifier
MUTE	Multimodal hateful meme detection	Bengali and code-mixed Bengali-English dataset (Hossain et al., 2024)
MUTE	Neural network design	Data-similarity-driven multi-hot target encoding (Jaiswal et al., 2019)
MUTE	Neural machine translation	Multi-Unit Transformers (Yan et al., 2020)
MUTE-SLAM	Neural RGB-D SLAM	Multiple tri-plane hash representations (Yan et al., 2024)
MUTE-DSS	Maritime acoustics	Decision support for minimizing underwater radiated noise (Venkateshwaran et al., 3 Aug 2025)
MuTe	Muography instrumentation	Hybrid Muon Telescope for volcano imaging (Peña-Rodríguez et al., 2020)
MUTE	Underground muon physics	MUon inTensity codE (Woodley et al., 2024)

A plausible implication is that the term has become attractive as a compact mnemonic for suppression, multiplicity, or muon-related computation, but the cited works do not define any common cross-domain ontology.

2. MUTE in multimodal benchmarks and multimodal reasoning

In multimodal hateful-content research, MUTE stands for “A Multimodal Dataset for Detecting Hateful Memes.” It is explicitly characterized as a Bengali and code-mixed Bengali-English hateful meme dataset containing 4158 memes with hateful/non-hateful labels, introduced to address the English-centric bias of prior hateful meme resources and the underrepresentation of code-mixed, low-resource settings (Hossain et al., 2024). The task is binary hateful meme classification over paired image and caption inputs, and the paper emphasizes that visual-only, text-only, and multimodal configurations should be evaluated separately because hateful meaning is often distributed across both modalities.

The associated modeling paper argues that multimodal hateful content detection requires feature alignment before fusion. Its proposed context-aware attention framework is evaluated on MUTE and MultiOFF, reporting F1-scores of $69.7\%$ and $70.3\%$, respectively, with approximately $2.5\%$ and $3.2\%$ improvement over state-of-the-art systems on those datasets (Hossain et al., 2024). The text also states that joint evaluation of visual and textual features improves classification by approximately 3%, reinforcing the claim that MUTE is not reducible to OCR-enhanced text classification.

This use of MUTE is notable because the dataset is simultaneously linguistic, visual, and sociotechnical. Code-mixing, low-resource NLP, sarcasm, and cross-modal inconsistency all enter the benchmark definition. The detailed annotation guideline mentioned in the source material further positions MUTE as a template for comparable datasets in other resource-constrained languages.

3. MUTE in machine learning methods and computational systems

In neural network design, MUTE denotes “Data-Similarity Driven Multi-hot Target Encoding.” The method replaces one-hot labels with $K$-hot codes of length $N$ for an $N$-class problem, using a confusion-derived weighting scheme to optimize Hamming distances so that more confusable classes are assigned greater separation in code space (Jaiswal et al., 2019). The stated objective is to improve generalizability and robustness without increasing model size, and the paper reports gains on MNIST, CIFAR-10, and ICON-50, including average MNIST accuracy improvements of 2.8% with LeNet and 7.1% with ConvNet over one-hot encoding (Jaiswal et al., 2019). The method is presented as a drop-in replacement based on sigmoid outputs, binary cross-entropy, and nearest-code inference.

In neural machine translation, MUTE denotes “Multi-Unit Transformers.” Here the central modification is the replacement of the standard single-unit layer with several parallel Transformer units, combined with a bias module and sequential dependency to encourage diversity and complementarity (Yan et al., 2020). The reported improvements over Transformer-Base are up to +1.52, +1.90, and +1.10 BLEU on NIST Chinese-to-English, WMT’14 English-to-German, and WMT’18 Chinese-to-English, with only a mild inference-speed drop of about 3.1% for the four-unit setting; the paper also states that the method surpasses Transformer-Big with only 54% of its parameters (Yan et al., 2020).

In neural SLAM, MUTE appears in MUTE-SLAM, a real-time neural RGB-D SLAM system built on multiple tri-plane hash-encodings and dynamic sub-map allocation (Yan et al., 2024). The method is explicitly designed to avoid pre-defined scene boundaries, to optimize all sub-maps intersecting the current camera frustum concurrently, and to perform periodic global bundle adjustment. The paper reports an average ATE RMSE of 8.00 cm on ScanNet and frame processing times of 0.21 s/FPT on Replica, 0.28 s/FPT on ScanNet, and 0.22 s/FPT on Apartment, with parameter counts that remain comparatively stable across scene sizes (Yan et al., 2024).

In maritime acoustics, MUTE denotes “Mitigating Underwater Noise Transmission and Effects” in MUTE-DSS, a ROS2-centric digital-twin-based decision support system for voyage planning (Venkateshwaran et al., 3 Aug 2025). Its pipeline combines a semi-empirical reference spectrum for near-field ship noise, 3D Bellhop ray tracing for transmission loss, Batch Informed Trees for routing, and a genetic algorithm for adaptive speed profiling. The reported case studies show reductions in cumulative noise exposure level of up to 7.14 dB, corresponding to approximately 80.68% reduction in a simplified scenario, and an average 4.90 dB reduction, corresponding to approximately 67.6% reduction in a more realistic dynamic setting (Venkateshwaran et al., 3 Aug 2025).

Taken together, these works use MUTE to label methods that restructure representation spaces, computation paths, or optimization pipelines. This suggests a recurring association between the acronym and controlled multiplicity: multiple hot bits, multiple parallel units, multiple sub-maps, or multi-stage route–speed optimization.

4. Mute as privacy control, interface state, and adversarial suppression

In smart-speaker privacy research, mute denotes a built-in privacy control whose underuse motivates alternative interaction design. The paper on the Privacy Hat argues that smart speakers can misinterpret hot words and record voice data without consent, while existing mute buttons are rarely used (Tiefenau et al., 2019). The proposed artifact is a tangible object that can be placed on top of the speaker so that “covering the device with the Privacy Hat will mute it,” simultaneously preventing listening and making the muted state visually salient. The reported prototype uses a Raspberry Pi, a modified Amazon Echo Dot, wiring to the red status LED and mute-button contact, and a distance sensor in a 3D-printed docking case; the planned study design is a 4-week field study with two 2-week phases (Tiefenau et al., 2019).

In video-conferencing privacy analysis, mute is examined as an implementation-dependent and often misleading interface state rather than a guaranteed microphone cutoff. A user study with 223 valid participants found that people often treat mute as a privacy boundary, especially for concealing background activities, but runtime binary analysis showed fragmented behaviors across native desktop clients (Yang et al., 2022). The paper reports that all studied apps could actively query the microphone while muted, and that Cisco Webex continuously sampled raw audio across mute state in the tested Windows and macOS configurations, transmitting telemetry to https://tsa3.webex.com once per minute while muted. Using intercepted telemetry packets, the authors implemented a proof-of-concept background activity classifier achieving 81.9% macro accuracy over six activity classes (Yang et al., 2022). Browser-based VCAs using WebRTC are contrasted with this behavior because the browser enforces software mute.

In adversarial ASR, muting becomes a targeted attack objective. “Muting Whisper” learns a universal acoustic realization of Whisper’s $\texttt{<|endoftext|>}$ token so that a prepended 0.64-second segment causes immediate decoding termination (Raina et al., 2024). The paper reports muting success above 97% across eight Whisper models, including 99.7% on tiny.en and 97.8% on medium, with transfer across datasets and partial transfer across tasks (Raina et al., 2024). The attack is universal across utterances but generally model-specific across Whisper sizes.

In polyphonic sound event detection, Mute is the deletion component of the Mirage and Mute Attack framework. The attack sets the target region to absence while preserving non-target outputs via a dedicated preservation loss, and the paper introduces Editing Precision (EP) to jointly measure target success and collateral stability (Su et al., 2 Oct 2025). Reported single-target results include 94.56% EP on CRNN and 99.11% EP on ATST-SED (Su et al., 2 Oct 2025). Here, mute no longer means user privacy or interface muting; it means adversarially forcing a detection model not to report an actually present event.

5. MuTe and MUTE in muography and underground muon physics

In volcanic muography, MuTe denotes a hybrid Muon Telescope designed for imaging Colombian volcanoes. The telescope combines a scintillator hodoscope for trajectory reconstruction with a Water Cherenkov Detector (WCD) for deposited-energy measurement and background rejection, plus a picosecond Time-of-Flight (ToF) system for rejecting backward and scattered low-momentum muons (Peña-Rodríguez et al., 2020). The design paper reports reconstruction of 3481 discrete directions, a maximum acceptance of about $3.69\ \text{cm}^2\ \text{sr}$, angular resolution of about 32 mrad at 250 cm panel spacing, estimated ToF resolution of about 138 ps, a momentum threshold of approximately $0.4 \pm 0.1\ \text{GeV}/c$ for rejecting soft muons, and total power consumption of about 41.4 W (Peña-Rodríguez et al., 2020).

Commissioning and calibration studies refine this picture. Layered trigger logic defines increasingly selective event classes, with the full T5 condition requiring both hodoscope coincidence and WCD trigger (Peña-Rodríguez et al., 2019). The paper divides WCD-hodoscope events into three deposited-energy regions: below 144 MeV for mostly $70.3\%$0, a muon window of $70.3\%$1, and above 400 MeV for mostly multiparticle events; the muon window represents about 40% of the acquired WCD-hodoscope events, while the remaining 60% is background (Peña-Rodríguez et al., 2019). It also reports that low-momentum scattered muons satisfy approximately $70.3\%$2 for traversing one meter, and that T5 events are only about 0.2% of total WCD events (Peña-Rodríguez et al., 2019).

A detailed Geant4 simulation models materials, geometry, and photosensitivity for both the hodoscope and WCD (Vásquez-Ramírez et al., 2019). For the WCD, 10$70.3\%$3 injected vertical 3 GeV muons produce 46,857 Cherenkov photons, 1,617 photons at the PMT surface, and 203.2 photoelectrons, whereas vertical 20 MeV electrons yield 16.7 photoelectrons (Vásquez-Ramírez et al., 2019). These results underwrite the hybrid trigger logic and the use of the vertical muon equivalent for calibration.

The SiPM characterization study shows why MuTe requires explicit thermal management and thresholding. For the Hamamatsu S13360-1350CS, the breakdown voltage at 25 °C is approximately 52.3 V, with $70.3\%$4; the authors conclude that the discrimination threshold must be above 5 pe to suppress dark count, crosstalk, and afterpulsing (Sánchez-Villafrades et al., 2021). The on-field analysis also reports an empirical temperature dependence of the detected flux of $70.3\%$5 (Sánchez-Villafrades et al., 2021).

A distinct use of MUTE appears in underground muon transport. In “Comparing Calculations of Seasonal Variations of Atmospheric Muons in Deep Underground Detectors,” MUTE is the MUon inTensity codE, combining MCEq with PROPOSAL to compute realistic survival probabilities for muons propagated to 2000 m.w.e. (Alves et al., 25 Aug 2025). The paper states that using MUTE survival probabilities gives underground rates about 15% lower than a simple minimum-energy threshold approximation and changes angular distributions, especially at large zenith angles (Alves et al., 25 Aug 2025). In MUTE v3, the same code is extended to flat overburdens, mountains, and underwater sites, integrating daemonflux, site-specific rock densities, chemical compositions, and topographic profiles; the paper reports excellent agreement with available data for most underground sites and provides examples such as a prediction of $70.3\%$6 for Soudan versus a measured $70.3\%$7 (Woodley et al., 2024).

6. Incomplete or textually unavailable usages

One additional usage is attached to arXiv entry (Li et al., 26 Feb 2026), titled “Layer-Targeted Multilingual Knowledge Erasure in LLMs.” Its abstract presents MUTE as “Multilingual Unlearning via Targeted Erasure,” a framework that uses Centered Kernel Alignment (CKA) and Linguistic Regions Development Score (LRDS) to identify intermediate, language-agnostic layers for multilingual knowledge erasure (Li et al., 26 Feb 2026). However, the supplied material explicitly states that the provided document is not the research paper but a NeurIPS 2025 formatting template and checklist, and that it contains no actual research content on multilingual unlearning, CKA, LRDS, experimental setup, or mechanistic analysis (Li et al., 26 Feb 2026).

This discrepancy is important because MUTE is unusually polysemous: acronym reuse can create false continuity across unrelated literatures, while incomplete or mismatched source texts can make one nominal usage appear commensurable with better documented ones. In the present record, the multilingual-unlearning MUTE can only be described at the level of its abstract; no further factual technical account is available from the supplied material (Li et al., 26 Feb 2026).

In aggregate, MUTE names a heterogeneous set of technical objects rather than a unified concept. Its meanings range from dataset construction and representation learning to privacy controls, adversarial suppression, muography instrumentation, and underground muon transport. For arXiv-oriented scholarship, the term therefore functions primarily as a disambiguation problem: interpretation depends entirely on expansion, capitalization, and disciplinary setting.