W-av: Polysemous Term in Multiple Domains

• W-av is a polysemous term representing diverse technical constructs across algebraic modules, machine learning for AV tasks, human-state modeling, and high-energy physics.
• It encompasses differentiable holonomic AV-modules, weakly-supervised video parsing, and Whisper-based audiovisual ASR, each employing distinct methodologies and measurable metrics.
• In autonomous vehicle studies and collider experiments, W-av refers to a user well-being variable and anomalous W-boson gauge vertices, highlighting the need for domain-aware disambiguation.

W-av is not a single standardized technical term. In recent arXiv literature, it denotes several distinct objects that share only the juxtaposition of “$W$” and “AV.” These include the finite-dimensional $\mathfrak{gl}_n$-module $W$ that appears in differentiable holonomic $AV$-modules, weakly-supervised audio-visual video parsing, Whisper-based audiovisual automatic speech recognition, the autonomous-vehicle user’s well-being variable $w_k$, and anomalous $W$-boson gauge vertices in electroweak production. The term is therefore best understood as a polysemous label whose meaning is fixed entirely by disciplinary context (Billig et al., 18 Nov 2025, Lai et al., 14 May 2025, Li et al., 26 Jan 2026, Zahedi et al., 21 May 2025, Collaboration, 2017, Yang et al., 2012).

1. Terminological range

In the supplied literature, the string “W-av” spans mathematics, machine learning, human-centered autonomy, and high-energy physics. The following map summarizes the principal uses.

Usage	Meaning	Representative source
$W$ in $AV$-modules	Finite-dimensional $\mathfrak{gl}_n$-module controlling the fiber factor in a differentiable holonomic $AV$-module	(Billig et al., 18 Nov 2025)
Weakly-supervised AV parsing	Audio-visual video parsing trained from video-level labels via pseudo-labels	(Lai et al., 14 May 2025)
Whisper-AV	Whisper-based audiovisual ASR with visual fusion in encoder and decoder	(Li et al., 26 Jan 2026)
$\mathfrak{gl}_n$0 for AV users	AV user’s well-being in a Dynamic Bayesian Network and causal decision model	(Zahedi et al., 21 May 2025)
Anomalous $\mathfrak{gl}_n$1 vertex	Triple or quartic gauge couplings involving $\mathfrak{gl}_n$2 bosons in collider phenomenology	(Collaboration, 2017, Yang et al., 2012)

This distribution of meanings is substantive rather than merely lexical. In each field, the notation anchors a different formal object: a module factor, a supervision regime, a multimodal ASR architecture, a latent cognitive state, or an EFT deformation of gauge interactions. A plausible implication is that cross-domain indexing of “W-av” requires domain-aware disambiguation rather than string matching alone.

2. $\mathfrak{gl}_n$3 as the fiber factor in differentiable holonomic $\mathfrak{gl}_n$4-modules

In algebraic geometry and representation theory, $\mathfrak{gl}_n$5 denotes the pair $\mathfrak{gl}_n$6, where $\mathfrak{gl}_n$7 is a commutative algebra of functions on a variety and $\mathfrak{gl}_n$8 is a Lie algebra of derivations of $\mathfrak{gl}_n$9. An $W$0-module is a left module over $W$1, equivalently an $W$2-module $W$3 with a $W$4-action satisfying

$W$5

with no assumption that the dependence on $W$6 is $W$7-linear. On a smooth irreducible quasi-projective variety $W$8, the paper studies sheaf-theoretic $W$9-modules, their differentiability, and their holonomicity (Billig et al., 18 Nov 2025).

The key structural point is that differentiability truncates the jet action. For $AV$0-differentiable modules, the quotient

$AV$1

acts on the module, and in the basic $AV$2 case one has $AV$3. This is the source of the paper’s $AV$4: a finite-dimensional $AV$5-module encoding the fiberwise first-jet contribution of vector fields. The principal theorem states that if $AV$6 is a simple differentiable holonomic sheaf of $AV$7-modules on a smooth irreducible quasi-projective variety of dimension $AV$8, then on every étale chart $AV$9,

$w_k$0

where $w_k$1 is a simple holonomic $w_k$2-module and $w_k$3 is a simple finite-dimensional $w_k$4-module. In this local model, the $w_k$5-action is through $w_k$6, while the vector-field action combines the $w_k$7-module action with the $w_k$8-action on $w_k$9.

For the 2-differentiable case, the local action is

$W$0

This formula records the precise failure of naive $W$1-linearity: the product rule is preserved in the function argument, but the coordinate-change term lands in the $W$2-fiber. When $W$3 is integrable, meaning that it integrates to a rational algebraic representation $W$4, the local decomposition globalizes to

$W$5

with $W$6 a simple holonomic $W$7-module and $W$8 the tensor module associated to $W$9. When $W$0 is non-integrable, the local factorization need not glue; the paper resolves the obstruction via “charged $W$1-modules.” In this context, “W-av” denotes the coupling of a fixed $W$2-type $W$3 to the differentiable $W$4-module structure.

3. W-av as weakly-supervised audio-visual video parsing

In computer vision and multimodal learning, “W-av” is used for weakly-supervised audio-visual video parsing. Audio-Visual Video Parsing (AVVP) aims to recognize and temporally localize unimodal events, which occur exclusively in either the visual or acoustic modality, and multimodal events, which occur in both modalities concurrently. The weakly-supervised setting provides only modality-agnostic video-level labels $W$5 during training, while segment-level labels are unavailable for supervision. UWAV addresses this setting by combining pseudo-label generation, uncertainty weighting, feature mixup, and class-balanced re-weighting (Lai et al., 14 May 2025).

The inference backbone is the Hybrid Attention Network (HAN). Videos are split into one-second segments; for LLP and AVE, $W$6. Segment features are formed as $W$7 and $W$8, and HAN applies self- and cross-attention: $W$9

$AV$0

Segment probabilities are pooled to video-level predictions by MMIL pooling. UWAV’s distinctive addition is a pseudo-label generator that uses CLIP and CLAP segment encodings followed by transformer temporal modeling, pre-trained on UnAV. This directly addresses the limitation of prior per-segment pseudo-label methods, which ignore inter-segment dependencies.

Pseudo-labels are thresholded and masked by the video-level label $AV$1, and uncertainty is represented by the sigmoid margin to the threshold: $AV$2 These soft pseudo-labels drive the uncertainty-weighted loss, while a class-balanced variant compensates for the strong dominance of absent classes. UWAV also applies feature mixup across segments,

$AV$3

with an analogous audio branch and $AV$4.

Empirically, UWAV reports on LLP, with HAN backbone, segment-level Visual F1 $AV$5, Type@AV $AV$6, and Event@AV $AV$7, and event-level Visual F1 $AV$8, Type@AV $AV$9, and Event@AV $\mathfrak{gl}_n$0. On AVE with CLIP/CLAP features, the reported segment-level accuracy is $\mathfrak{gl}_n$1. Pseudo-label quality also improves, with LLP test Type@AV pseudo-label F1 of $\mathfrak{gl}_n$2 for UWAV versus $\mathfrak{gl}_n$3 for VALOR. In this usage, W-av is therefore a supervision regime and training methodology for AV temporal parsing rather than a symbolic variable.

4. W-av as Whisper-based audiovisual speech recognition

A second machine-learning usage identifies W-av with Whisper-based audiovisual ASR, or Whisper-AV. Here the central problem is noise-robust speech recognition through fusion of lip movements with a pre-trained Whisper encoder-decoder. The dual-use method injects visual features both into the Whisper encoder and into the decoder, with the stated motivations that encoder-side fusion learns audiovisual interactions and decoder-side fusion lets decoding dynamically weigh modalities (Li et al., 26 Jan 2026).

The system combines pre-trained Whisper ASR and AV-HuBERT large. On the encoder side, AV-HuBERT produces visual latents $\mathfrak{gl}_n$4, which are upsampled by frame-wise repetition to length $\mathfrak{gl}_n$5, projected to model dimension $\mathfrak{gl}_n$6, and gated by a zero-initialized scalar $\mathfrak{gl}_n$7: $\mathfrak{gl}_n$8 Because $\mathfrak{gl}_n$9 is initialized to zero, the audio path is initially unchanged and visual influence ramps up during fine-tuning. This additive early fusion was found to outperform concatenation at the encoder input.

On the decoder side, one Flamingo block is inserted before each original Whisper decoder block. Within a Flamingo block, the decoder state attends both to the audiovisual encoder outputs and to the visual memory, with gated residual fusion

$AV$0

where $AV$1 and $AV$2 are zero-initialized learned scalars. The full system is trained with sequence-to-sequence cross-entropy on 1929 hours of audiovisual data from LRS3, LRS2, and the VoxCeleb2 English subset.

The reported gains are concentrated in noisy conditions. At $AV$3 dB babble noise on LRS3 test, dual-use fusion improves Whisper small from $AV$4 WER to $AV$5, a $AV$6 relative improvement, and Whisper medium from $AV$7 to $AV$8, a $AV$9 relative improvement. Fine-tuned on 1929 hours, Whisper medium dual-use achieves average WER $\mathfrak{gl}_n$00 on MUSAN babble noise and $\mathfrak{gl}_n$01 on NoiseX babble noise across various SNRs. The paper also reports that encoder-only early fusion degrades for larger Whisper models in clean conditions, whereas dual-use fusion stabilizes training through zero-initialized encoder and decoder gates. In this literature, W-av denotes an architectural family centered on multimodal adaptation of Whisper.

5. W-av as the AV user’s well-being variable

In autonomous-vehicle human-state modeling, W-av denotes the AV user’s well-being variable $\mathfrak{gl}_n$02. The variable is part of a Dynamic Bayesian Network that models the AV user’s cognitive state $\mathfrak{gl}_n$03, where $\mathfrak{gl}_n$04 is well-being, $\mathfrak{gl}_n$05 is trust in the AV, and $\mathfrak{gl}_n$06 is the user’s intention toward another road user. The other road user’s state is $\mathfrak{gl}_n$07. Observed interaction variables are the AV’s accommodative action $\mathfrak{gl}_n$08, the other user’s accommodative action $\mathfrak{gl}_n$09, and the alignment indicator $\mathfrak{gl}_n$10 (Zahedi et al., 21 May 2025).

The learned model discretizes $\mathfrak{gl}_n$11, $\mathfrak{gl}_n$12, and $\mathfrak{gl}_n$13 into 6 bins and performs Bayesian filtering with tabular CPDs. The AV user’s well-being is defined as a multidimensional mobility construct including positive social interactions, satisfaction with travel, trust in other road users, and general well-being; it is measured by averaging 7 Likert-scale items and scaling to $\mathfrak{gl}_n$14. Trust is measured separately by a distinct Likert item. The paper reports that $\mathfrak{gl}_n$15 increases when the other road user yields rather than behaves unyieldingly, that both $\mathfrak{gl}_n$16 and $\mathfrak{gl}_n$17 are higher when the AV’s action aligns with the user’s intention, and that $\mathfrak{gl}_n$18 is positively correlated with $\mathfrak{gl}_n$19, with $\mathfrak{gl}_n$20.

The DBN is extended to a causal inference model for AV decision-making. Choosing $\mathfrak{gl}_n$21 is treated as a do-intervention, and the policy is

$\mathfrak{gl}_n$22

The paper analyzes utilities $\mathfrak{gl}_n$23, $\mathfrak{gl}_n$24, and a trade-off

$\mathfrak{gl}_n$25

Without evidence, the optimal policy for maximizing $\mathfrak{gl}_n$26 or $\mathfrak{gl}_n$27 is “always yield.” If the user’s intention $\mathfrak{gl}_n$28 is observed, the optimal policy is to align with it: yield for $\mathfrak{gl}_n$29 and unyield for $\mathfrak{gl}_n$30. Under the cost-sensitive trade-off, the paper reports that if $\mathfrak{gl}_n$31, the AV should choose $\mathfrak{gl}_n$32, whereas if $\mathfrak{gl}_n$33, it increasingly chooses $\mathfrak{gl}_n$34 except in very high well-being contexts. The reported state inference accuracies are $\mathfrak{gl}_n$35 for well-being, $\mathfrak{gl}_n$36 for trust, and $\mathfrak{gl}_n$37 for intention. In this usage, W-av is a latent human-centered state variable embedded in sequential probabilistic inference and decision optimization.

6. W-av as anomalous $\mathfrak{gl}_n$38-boson gauge interactions

In collider phenomenology, W-av refers to anomalous $\mathfrak{gl}_n$39-boson gauge vertices. Two distinct but related settings appear. The ATLAS $\mathfrak{gl}_n$40 analysis studies electroweak production of a $\mathfrak{gl}_n$41 boson with two jets at high dijet invariant mass, a topology sensitive to the $\mathfrak{gl}_n$42 vertex with $\mathfrak{gl}_n$43. The $\mathfrak{gl}_n$44 study examines $\mathfrak{gl}_n$45 as a probe of anomalous quartic gauge couplings in the $\mathfrak{gl}_n$46 sector (Collaboration, 2017, Yang et al., 2012).

For $\mathfrak{gl}_n$47, the analysis uses the Hagiwara–Peccei–Zeppenfeld–Hikasa parameterization,

$\mathfrak{gl}_n$48

together with CP-violating terms and HISZ relations. ATLAS measures fiducial electroweak $\mathfrak{gl}_n$49 cross sections of $\mathfrak{gl}_n$50 fb at $\mathfrak{gl}_n$51 TeV and $\mathfrak{gl}_n$52 fb at $\mathfrak{gl}_n$53 TeV, using data corresponding to $\mathfrak{gl}_n$54 and $\mathfrak{gl}_n$55 fb$\mathfrak{gl}_n$56. In a high-$\mathfrak{gl}_n$57 region defined by $\mathfrak{gl}_n$58 TeV and leading-jet $\mathfrak{gl}_n$59 GeV, the paper sets $\mathfrak{gl}_n$60 CL limits. With $\mathfrak{gl}_n$61 TeV, these include

$\mathfrak{gl}_n$62

No significant deviation from the Standard Model is observed in inclusive or differential observables.

For $\mathfrak{gl}_n$63, the effective Lagrangian is written in terms of the two photonic quartic Lorentz structures

$\mathfrak{gl}_n$64

or equivalently by coefficients $\mathfrak{gl}_n$65 and $\mathfrak{gl}_n$66. The study uses MadGraph/MadEvent 5, PYTHIA 6.4, and DELPHES, and finds that at $\mathfrak{gl}_n$67 TeV the Standard Model $\mathfrak{gl}_n$68 process can be observed with approximately $\mathfrak{gl}_n$69 significance at $\mathfrak{gl}_n$70 fb$\mathfrak{gl}_n$71 and approximately $\mathfrak{gl}_n$72 at $\mathfrak{gl}_n$73 fb$\mathfrak{gl}_n$74. The same channel constrains $\mathfrak{gl}_n$75 and $\mathfrak{gl}_n$76 at the $\mathfrak{gl}_n$77 scale. Without form factors, for example, the $\mathfrak{gl}_n$78 CL interval at $\mathfrak{gl}_n$79 fb$\mathfrak{gl}_n$80 for $\mathfrak{gl}_n$81 is

$\mathfrak{gl}_n$82

with analogous bounds for $\mathfrak{gl}_n$83. In this domain, W-av denotes deviations from Standard Model $\mathfrak{gl}_n$84-boson gauge interactions, probed through high-energy tails in VBF-like $\mathfrak{gl}_n$85 and associated $\mathfrak{gl}_n$86 production.