WLZZ Models: AI, Physics & Phenomenology

• WLZZ models are a multifaceted topic spanning language models enhanced with embodied experience, matrix models in mathematical physics, flavor constructions, and collider shorthand usage.
• In embodied AI, WLZZ models leverage simulator-generated traces and techniques like EWC-LoRA to significantly boost performance in tasks such as plan generation and object tracking.
• In mathematical physics, WLZZ models reveal deep integrability structures via W-representations and beta-deformations, connecting partition functions to generalized Ward identities.

The designation WLZZ models refers to several distinct objects in current research. In one usage, it denotes LLMs enhanced with world models via embodied experience, where simulator-generated interaction traces are used to finetune autoregressive transformers for physical reasoning and planning (Xiang et al., 2023). In mathematical physics, it denotes Wang–Liu–Zhang–Zhao matrix models, a family of partition functions defined by $W$-representations, Schur or Jack expansions, and associated commutative subalgebras of $W_{1+\infty}$ (Mironov et al., 7 Jul 2025). A separate phenomenological usage applies the label to models in which the $W$ mass and a heavy $Z^\prime$ are jointly used to constrain $b \rightarrow s \ell \ell$ anomalies (Allanach et al., 2022). In collider phenomenology, WLZZ is also used as shorthand for $W^\pm ZZ$ production with leptonic decays at the LHC (Yong-Bai et al., 2015).

1. Terminological scope

The principal uses of the term are summarized below.

Usage	Defining object	Representative source
Embodied AI	LLMs finetuned with world-model experiences	(Xiang et al., 2023)
Mathematical physics	Wang–Liu–Zhang–Zhao matrix-model partition functions	(Mironov et al., 7 Jul 2025)
Flavor phenomenology	$M_W$- and $Z^\prime$-selected models for $b \rightarrow s \ell \ell$ anomalies	(Allanach et al., 2022)
Collider shorthand	$W^\pm ZZ$ production with leptonic decays	(Yong-Bai et al., 2015)

In the mathematical-physics literature, WLZZ stands for Wang–Liu–Zhang–Zhao; the corresponding models were introduced as new partition functions or $W_{1+\infty}$0-functions defined by $W_{1+\infty}$1-representations, organized into positive and negative branches, and later generalized to arbitrary integer rays, rational rays, cones, and $W_{1+\infty}$2-deformations (Mironov et al., 2023). In the embodied-AI literature, the same label is used for LLMs that internalize embodied knowledge from a simulator rather than relying only on written text (Xiang et al., 2023). In phenomenological usage, the label is attached to constructions where $W_{1+\infty}$3 and a heavy $W_{1+\infty}$4 are treated jointly in flavor fits (Allanach et al., 2022). A plausible implication is that any encyclopedia treatment must be explicitly disambiguating, because the term does not denote a single universally shared model class across disciplines.

2. Embodied-experience WLZZ models in machine learning

In the embodied-AI usage, WLZZ models are LLMs enhanced with world models via embodied experience (Xiang et al., 2023). The motivation is that LLMs trained only on written text lack embodied experiences needed for robust physical reasoning and planning, and therefore fail at tasks such as counting objects after a sequence of actions, tracking items manipulated by multiple agents, or writing multi-step plans that respect affordances and object state changes. The proposed remedy is to deploy an embodied agent in VirtualHome, a multi-agent 3D household simulator supporting atomic actions such as Walk, Grab, Put, Open, and SwitchOn, collect diverse embodied experiences, convert those traces into supervised tasks, and finetune the LLM itself rather than prompting it task-by-task with a simulator (Xiang et al., 2023).

The data pipeline has two sources. First, goal-oriented planning uses a Monte Carlo Tree Search planner over goal predicates such as $W_{1+\infty}$5 or $W_{1+\infty}$6, with reward shaping $W_{1+\infty}$7 when a goal predicate is satisfied and $W_{1+\infty}$8 per step. Second, random exploration logs single-agent or multi-agent trajectories containing occlusions, handoffs, distractors, and irrelevant actions. These traces are converted into supervised tasks: plan generation, activity recognition, counting, object path tracking, and object location QA. The multi-task supervised objective is written as

$W_{1+\infty}$9

with task weights chosen on held-out data as 1.0, 0.7, 1.0, 1.0 for plan generation, activity recognition, counting, and object tracking, respectively (Xiang et al., 2023).

To preserve general language capability, the method combines EWC and LoRA. The EWC term is

$W$0

while LoRA uses

$W$1

The paper further uses an EWC-LoRA form in which the Fisher-weighted regularizer acts on the low-rank delta $W$2 (Xiang et al., 2023). The implementation details given are AdamW, Int8 inference/training optimization, learning rate $W$3, batch size 20, LoRA rank 8, and LoRA scaling coefficient 32, with experiments on GPT-Neo-1.3B, GPT-J-6B, OPT-13B, and LLaMA-13B (Xiang et al., 2023).

The empirical results are unusually strong on embodied tasks. The scope is 18 downstream tasks in total, with 64.28% average improvement over base LMs. Representative numbers include GPT-J plan generation Rouge-L 34.31 $W$4 51.23, GPT-J counting accuracy 30.41% $W$5 67.01%, GPT-J object path tracking LCS 33.86 $W$6 98.67, surpassing ChatGPT 59.53, LLaMA-13B object location QA 79.0% vs ChatGPT 67.5% vs base 28.5%, and LLaMA-13B counting 79.38% vs ChatGPT 66.49% vs base 29.38% (Xiang et al., 2023). General language ability is largely preserved: Pile perplexity changes only from 3.443 to 3.537 for GPT-J, 4.120 to 4.193 for GPT-Neo, 4.077 to 4.358 for OPT-13B, and 3.036 to 3.069 for LLaMA-13B (Xiang et al., 2023). This suggests that the proposal is not merely simulator-assisted prompting, but a form of embodied finetuning intended to internalize object permanence, tracking, and physically grounded planning.

3. Wang–Liu–Zhang–Zhao matrix models

In mathematical physics, WLZZ models are a class of matrix-model partition functions originally defined via $W$7-representations—that is, as $W$8-functions annihilated by families of $W$9/Virasoro-type constraint operators (Mironov et al., 7 Jul 2025). In the simplest instances, they are generated by commutative subalgebras of the $Z^\prime$0 algebra, and admit $Z^\prime$1- and $Z^\prime$2-deformations realized by commutative subalgebras of the affine Yangian of $Z^\prime$3 and of the Ding–Iohara–Miki or elliptic Hall algebra, respectively (Mironov et al., 7 Jul 2025). The operator formulation distinguishes three branches,

$Z^\prime$4

$Z^\prime$5

$Z^\prime$6

with $Z^\prime$7 the cut-and-join generator and higher $Z^\prime$8 produced recursively by commutators (Mironov et al., 7 Jul 2025).

A central realization is the interpolating two-matrix integral

$Z^\prime$9

with $b \rightarrow s \ell \ell$0 and normalization $b \rightarrow s \ell \ell$1 (Mironov et al., 7 Jul 2025). A closely related formulation presents the same object as the interpolating two-matrix model

$b \rightarrow s \ell \ell$2

with $b \rightarrow s \ell \ell$3 Hermitian, $b \rightarrow s \ell \ell$4 anti-Hermitian, and $b \rightarrow s \ell \ell$5, understood as a formal power series in the times (Mironov et al., 2023). Specializations recover the positive branch, the negative branch, and the $b \rightarrow s \ell \ell$6 one-matrix reduction after integrating out the auxiliary matrix (Mironov et al., 2023).

The matrix-model realization is accompanied by a pronounced superintegrability structure. If all $b \rightarrow s \ell \ell$7- and $b \rightarrow s \ell \ell$8-contours pass through the origin, the partition function equals the Schur expansion

$b \rightarrow s \ell \ell$9

with

$W^\pm ZZ$0

while the interpolating WLZZ models more generally appear as skew hypergeometric $W^\pm ZZ$1-functions built from skew Schur functions $W^\pm ZZ$2 and content factors $W^\pm ZZ$3 (Mironov et al., 7 Jul 2025). This places WLZZ matrix models at the intersection of two-matrix integrals, cut-and-join operators, Schur expansions, and 2D Toda integrability.

4. Spectral curves, commutative rays, and generalized $W^\pm ZZ$4 algebras

A major strand of the WLZZ literature studies the hidden integrable structure behind these partition functions. The guiding statement is that the relevant Hamiltonians lie on integer slope rays in $W^\pm ZZ$5, and that each ray gives a commuting family. The resulting many-body systems include the rational Calogero model as the simplest example, while higher rays produce additional integrable systems that had “escaped attention in the past” (Mironov et al., 2023). The same framework was then extended to rational rays $W^\pm ZZ$6, cones, one-body differential operators on a circle, matrix and eigenvalue realizations, bosonic time variables, and $W^\pm ZZ$7-deformations, with integer rays surviving the $W^\pm ZZ$8-deformation while rational rays generally do not (Mironov et al., 2023).

On the spectral side, the key observation is that the spectral curve can be extracted directly from the part of the $W^\pm ZZ$9-operator that is linear in time variables, denoted $M_W$0 (Mironov et al., 2022). In the negative branch one obtains the family

$M_W$1

while the boundary case $M_W$2 yields Lambert or higher Lambert curves, and positive-branch examples produce small-$M_W$3 algebraic curves such as

$M_W$4

or, with multiple couplings,

$M_W$5

The paper emphasizes that for $M_W$6 the relation between topological and $M_W$7 expansions is broken, and that positive-branch WLZZ models are naturally small-$M_W$8 rather than large-$M_W$9 objects (Mironov et al., 2022).

The Ward-identity side is organized by generalized $Z^\prime$0 algebras. Each integer ray is associated with a family $Z^\prime$1, and the WLZZ partition functions $Z^\prime$2 satisfy generalized Ward identities

$Z^\prime$3

or equivalently

$Z^\prime$4

(Drachov et al., 2023). A later formulation expresses the ray Hamiltonians as

$Z^\prime$5

with partition functions

$Z^\prime$6

and conjectural full Ward identities

$Z^\prime$7

for $Z^\prime$8 (Drachov, 2024). The vertical ray acts diagonally on Schur functions and yields hypergeometric KP/Toda $Z^\prime$9-functions, while the $b \rightarrow s \ell \ell$0 and vertical families coincide with the rational and trigonometric Calogero–Sutherland Hamiltonians, respectively (Drachov et al., 2023). This suggests that the phrase “WLZZ models” in this literature denotes not just a family of partition functions, but an entire operator-theoretic infrastructure linking $b \rightarrow s \ell \ell$1, cut-and-join recursions, generalized Ward identities, and Calogero-type commuting Hamiltonians.

5. $b \rightarrow s \ell \ell$2-deformations, $b \rightarrow s \ell \ell$3-ensembles, and phase structure

The $b \rightarrow s \ell \ell$4-deformed WLZZ models replace Schur technology by Jack-polynomial technology and admit a two-$b \rightarrow s \ell \ell$5-ensemble realization (Mironov et al., 2024). The central partition function is

$b \rightarrow s \ell \ell$6

with $b \rightarrow s \ell \ell$7, $b \rightarrow s \ell \ell$8 the $b \rightarrow s \ell \ell$9-deformed HCIZ kernel, and normalization $W^\pm ZZ$0 (Mironov et al., 2024). The same object equals the Jack expansion

$W^\pm ZZ$1

with Jack norms and $W^\pm ZZ$2 determined explicitly by partition data (Mironov et al., 2024). The $W^\pm ZZ$3-deformed HCIZ kernel is governed by Dunkl operators

$W^\pm ZZ$4

and satisfies the key identity

$W^\pm ZZ$5

which underlies the $W^\pm ZZ$6-deformed Ward identities (Mironov et al., 2024).

Direct evaluation of these $W^\pm ZZ$7-ensemble integrals shows that more than one contour choice is possible (Mironov et al., 2024). In particular, the scalar product with $W^\pm ZZ$8, $W^\pm ZZ$9 is equal to the scalar product with $W_{1+\infty}$00, and the proof uses a Macdonald-conjecture-type integral transform together with Jack orthogonality (Mironov et al., 2024). This contour ambiguity becomes still richer in the cubic two-matrix theory. For the unshifted cubic WLZZ model, the solution space of the Ward identities is a single point in the graded power-series ansatz, and the integral is nonzero only if all $W_{1+\infty}$01- and $W_{1+\infty}$02-contours pass through the origin; otherwise it vanishes (Mironov et al., 7 Jul 2025). For shifted models, obtained for example by $W_{1+\infty}$03, the single master equation acquires extra lower-grading operators, the number of free parameters increases, and nonperturbative phases are encoded by Lefschetz thimbles (Mironov et al., 7 Jul 2025).

The explicit $W_{1+\infty}$04, $W_{1+\infty}$05 example makes the phase structure concrete. The model

$W_{1+\infty}$06

has a two-dimensional solution space parameterized by contour weights $W_{1+\infty}$07, with two saddle contributions $W_{1+\infty}$08 and $W_{1+\infty}$09; the second saddle is exponentially suppressed as $W_{1+\infty}$10, exhibiting a Stokes phenomenon (Mironov et al., 7 Jul 2025). A plausible implication is that $W_{1+\infty}$11-deformation and contour choice are not auxiliary technicalities but part of the nonperturbative definition of the WLZZ matrix-model series.

6. $W_{1+\infty}$12- and $W_{1+\infty}$13-selected WLZZ models in flavor phenomenology

A separate usage of the term appears in flavor phenomenology, where WLZZ denotes models in which the $W_{1+\infty}$14 mass and a heavy $W_{1+\infty}$15 are jointly used to explain the $W_{1+\infty}$16 anomalies (Allanach et al., 2022). The gauge extension is

$W_{1+\infty}$17

with $W_{1+\infty}$18, $W_{1+\infty}$19, and anomaly freedom ensured by the inclusion of a right-handed neutrino $W_{1+\infty}$20 (Allanach et al., 2022). After the standard permutation that assigns the non-zero left-handed lepton charge to the second family, the charges are

$W_{1+\infty}$21

with all other fermions uncharged (Allanach et al., 2022). The Higgs charge $W_{1+\infty}$22 induces tree-level $W_{1+\infty}$23–$W_{1+\infty}$24 mass mixing,

$W_{1+\infty}$25

and therefore a positive shift in

$W_{1+\infty}$26

which raises $W_{1+\infty}$27 (Allanach et al., 2022).

Flavor violation is introduced through a single left-handed down-quark mixing angle $W_{1+\infty}$28, giving

$W_{1+\infty}$29

and generating the Wilson coefficients $W_{1+\infty}$30, $W_{1+\infty}$31, and $W_{1+\infty}$32 required by the global $W_{1+\infty}$33 fits (Allanach et al., 2022). Including the CDF II $W_{1+\infty}$34 value

$W_{1+\infty}$35

the paper performs a two-parameter global fit to 277 observables using smelli, flavio, and wilson, and finds that the original $W_{1+\infty}$36 model $W_{1+\infty}$37 is somewhat disfavoured, while the generalized models prefer

$W_{1+\infty}$38

with global $W_{1+\infty}$39-values above 0.05 (Allanach et al., 2022). A concrete example is $W_{1+\infty}$40, $W_{1+\infty}$41, which yields a global $W_{1+\infty}$42-value of 0.12, compared with the Standard Model value $W_{1+\infty}$43 (Allanach et al., 2022). In this usage, WLZZ does not denote a matrix model or a language-model architecture, but a flavor-gauge construction selected simultaneously by electroweak precision data and semileptonic $W_{1+\infty}$44-anomaly observables.

7. $W_{1+\infty}$45 collider usage

In collider phenomenology, WLZZ is also used as shorthand for $W_{1+\infty}$46 production, not for a standalone model class (Yong-Bai et al., 2015). The process

$W_{1+\infty}$47

with subsequent leptonic decays probes both the $W_{1+\infty}$48 triple gauge coupling and the $W_{1+\infty}$49 quartic gauge coupling. The cited calculation evaluates NLO QCD + NLO EW corrections at the 14 TeV LHC, using an improved narrow width approximation that retains off-shell contributions and spin correlations through MadSpin, with NNPDF2.3QED PDFs, a mixed $W_{1+\infty}$50 electroweak scheme, and the scale choice

$W_{1+\infty}$51

(Yong-Bai et al., 2015).

The numerical results are sizable. In the jet-veto event selection scheme with $W_{1+\infty}$52, the NLO QCD+EW relative corrections to the integrated cross section are 20.5% for $W_{1+\infty}$53 and 31.1% for $W_{1+\infty}$54, while the genuine NLO EW relative corrections are $W_{1+\infty}$55 and $W_{1+\infty}$56, respectively (Yong-Bai et al., 2015). The inclusive corrections are much larger, with $W_{1+\infty}$57-factors of approximately 2.56 for $W_{1+\infty}$58 and 2.87 for $W_{1+\infty}$59, driven largely by real light-quark emission and $W_{1+\infty}$60 channels (Yong-Bai et al., 2015). The study also emphasizes that the LO scale uncertainty, about 1.6%, is artificially small because $W_{1+\infty}$61 does not enter the LO matrix elements, whereas NLO results provide more realistic uncertainty bands (Yong-Bai et al., 2015).

This collider usage is terminologically distinct from the AI, matrix-model, and flavor-phenomenology usages. It nevertheless illustrates how the same four-letter label can denote either a model family or a physical production channel depending on disciplinary context.