WPgen: Constrained Generation in Structured Systems

• WPgen is a term for domain-specific generators that produce constrained, structured outputs rather than unconstrained synthesis.
• In maritime autonomy, WPgen employs multiobjective search to perturb waypoints, exposing self-adaptation failures and aiding safety validation.
• In web rendering and collider physics, WPgen predicts standardized CSS properties and generates NLO-accurate events, ensuring technical correctness.

Searching arXiv for the WPgen-related papers and primary sources. Searching arXiv for "WPgen". WPgen is a label applied in recent arXiv literature to several distinct generation systems that operate on structured technical representations rather than unconstrained natural-language outputs. In one line of work, it denotes a multiobjective search-based method for generating waypoint perturbations that expose failures in the self-adaptation logic of maritime autonomous vessels (Nylænder et al., 22 Jul 2025). In another, the term can be used as an interpretation of Web Rendering Parameters Generation, a task that predicts standardized element-level rendering parameters from HTML in order to produce coherent web presentation without direct CSS program synthesis (Shao et al., 2024). A further usage appears in collider phenomenology, where the term is associated with an event-generation workflow for same-sign $W$-boson scattering implemented in POWHEG-BOX-RES with RECOLA2, providing unweighted events with NLO electroweak accuracy matched to a QED parton shower and interfaced to a QCD parton shower (chiesa et al., 2019). Across these uses, the common conceptual thread is controlled generation over structured spaces under domain-specific constraints.

1. Terminological scope

The term “WPgen” is not a single standardized method across the literature. Instead, it appears as a name or interpretive label for systems in maritime autonomy, automated web presentation, and high-energy physics event generation. These systems differ in objective functions, representations, and evaluation protocols, but each replaces open-ended synthesis with a more constrained generation problem tied to an existing formal structure such as a waypoint sequence, a DOM tree, or a gauge-invariant scattering process (Nylænder et al., 22 Jul 2025).

This multiplicity of usage is important because it precludes treating WPgen as a unified algorithmic family in the narrow sense. A more accurate characterization is that WPgen denotes domain-specific generators whose outputs are validated against explicit technical criteria: dependability-triggering behavior for autonomous vessels, element-wise rendering fidelity for web pages, or NLO electroweak correctness with shower matching for collider events. This suggests that the name is best understood contextually rather than taxonomically.

2. Maritime autonomous vessels: waypoint generation for self-adaptation triggers

In maritime autonomous vessels, WPgen is a multiobjective, search-based approach for validating self-adaptation triggers in navigation software operating under uncertain and rapidly changing conditions (Nylænder et al., 22 Jul 2025). Autonomous vessels typically navigate through a predefined waypoint set $W=\{w_1,\dots,w_n\}$. WPgen perturbs this set within bounded neighborhoods to produce modified routes $W'=\{w_1',\dots,w_n'\}$ that remain close to the original trajectory while inducing inappropriate navigation. The purpose is not route optimization in the conventional sense, but the systematic discovery of conditions under which adaptation should be triggered.

The design context is the MAPLE-K loop, an extension of MAPE-K in which Monitor and Analyze detect issues, Plan and Execute enact changes, and Legitimate verifies that adaptations meet safety requirements. WPgen targets the navigation component because it is one of the most critical software elements in autonomous vessels. The method is intended to help developers validate both the logic of adaptation triggers and the correctness of their implementation.

“Inappropriate navigation” has a precise operational meaning in this setting. It consists of either unstable motion or failure to reach one or more waypoints. Instability is characterized through autocorrelation analysis of angle measurements such as roll, pitch, and yaw. An angle is unstable if its autocorrelation exhibits more than one peak above a $0.1$ threshold. A sub-path is unstable if at least one axis is unstable, stable if all are stable, and missing if the vessel does not reach the next waypoint. The optimization loop itself uses a proxy for this behavior: the normalized inflation of executed sub-path length relative to the straight-line distance between consecutive original waypoints. Larger inflation indicates more meandering motion and therefore a greater tendency toward instability.

The formal problem is bi-objective. The first objective minimizes deviation from the original route, while the second maximizes the instability proxy after simulation in the Marine Systems Simulator. For overwater vessels, waypoints lie in $\mathbb{R}^2$ with North-East coordinates; for underwater vessels they lie in $\mathbb{R}^3$ with North-East-Down coordinates. The first waypoint is fixed, each coordinate is bounded within a hypercube of half-side $\beta$ around the original, and a vessel-specific minimum separation $\delta_{\min}$ is enforced between consecutive waypoints through a death penalty. WPgen therefore searches for Pareto-optimal trade-offs between closeness and inappropriateness rather than collapsing the problem to a single scalar score.

3. Search formulation, operators, and seeding strategies

WPgen for maritime navigation is implemented with PlatEMO’s NSGA-II and uses a compact but highly constrained search space (Nylænder et al., 22 Jul 2025). The decision variables are per-waypoint offsets $\Delta w_i=w_i'-w_i$ for $i=2,\dots,n$, subject to coordinate-wise bounds and minimum-separation feasibility. The first objective is the total Euclidean deviation from the original waypoints. The second is

$W=\{w_1,\dots,w_n\}$0

where $W=\{w_1,\dots,w_n\}$1 is the executed sub-path corresponding to the $W=\{w_1,\dots,w_n\}$2-th leg of the route. Hypervolume is used to assess Pareto-front quality, with larger values indicating better coverage and diversity.

The experimental NSGA-II configuration is deliberately small in population size and long in horizon: population size $W=\{w_1,\dots,w_n\}$3, $W=\{w_1,\dots,w_n\}$4 generations, and $W=\{w_1,\dots,w_n\}$5 fitness evaluations. Simulated Binary Crossover is used with $W=\{w_1,\dots,w_n\}$6 and $W=\{w_1,\dots,w_n\}$7, while polynomial mutation uses an expected number of mutated variables equal to $W=\{w_1,\dots,w_n\}$8 and $W=\{w_1,\dots,w_n\}$9. Two implementation details are central. First, crossover points are aligned with waypoint boundaries so that a waypoint is not split between parents. Second, mutation is coordinate-level but rejected if it violates the minimum-separation constraint. These customizations make the search operators waypoint-aware and constraint-aware.

The method introduces three seeding variants. WPgen—Closeness seed initializes solutions by starting from the original waypoint set and applying a small number of random waypoint mutations, with the probability of additional mutations decreasing as more mutations are added. WPgen—Combined assigns half of the initial population to this closeness scheme and half to uniform random initialization within the bounded neighborhoods. WPgen—Random initializes all individuals except the unmodified original route uniformly at random. The inclusion of the unmodified route in every initial population anchors the search near the intended mission trajectory while preserving room for exploration.

These seeding strategies encode different assumptions about the structure of adverse cases. Closeness seed privileges local exploration near the original route, whereas Random emphasizes diversity. Combined is explicitly a balance of exploitation and exploration. The paper’s results show that this choice is not merely a heuristic detail but a substantive determinant of effectiveness, especially for underwater vehicles.

4. Experimental behavior and practical role in autonomous-vessel engineering

The empirical evaluation covers three autonomous-vessel models in the Marine Systems Simulator: Mariner, an overwater tanker with $W'=\{w_1',\dots,w_n'\}$0 waypoints; Remus100, an underwater autonomous vessel with $W'=\{w_1',\dots,w_n'\}$1 waypoints; and Nspauv, another underwater autonomous vessel with $W'=\{w_1',\dots,w_n'\}$2 waypoints (Nylænder et al., 22 Jul 2025). Each algorithm is run $W'=\{w_1',\dots,w_n'\}$3 times. The waypoint bounds are $W'=\{w_1',\dots,w_n'\}$4 m for Mariner and $W'=\{w_1',\dots,w_n'\}$5 m for Remus100 and Nspauv. Statistical comparisons use Mann–Whitney $W'=\{w_1',\dots,w_n'\}$6-tests with $W'=\{w_1',\dots,w_n'\}$7 and Vargha–Delaney $W'=\{w_1',\dots,w_n'\}$8 effect sizes.

Against Random Search, all WPgen variants significantly outperform the baseline on Mariner with large effect sizes. On Remus100, WPgen—Combined and WPgen—Random significantly outperform Random Search, but WPgen—Closeness performs significantly worse than the baseline. On Nspauv, no significant differences are observed among the algorithms. In pairwise comparison among WPgen variants, Mariner shows no significant differences, Remus100 favors Combined and Random over Closeness, and Nspauv again shows no significant differences. These findings indicate that seeding diversity matters particularly for underwater vessels, where local perturbations near the original route may under-explore the adverse-behavior space.

The characterization of generated paths clarifies the type of failures being discovered. For Mariner, stable sub-paths account for approximately $W'=\{w_1',\dots,w_n'\}$9–$0.1$0, unstable sub-paths for approximately $0.1$1–$0.1$2, and missing sub-paths for $0.1$3. For Remus100, stable sub-paths are approximately $0.1$4–$0.1$5, unstable $0.1$6–$0.1$7, and missing approximately $0.1$8–$0.1$9. For Nspauv, stable sub-paths are approximately $\mathbb{R}^2$0–$\mathbb{R}^2$1, unstable $\mathbb{R}^2$2–$\mathbb{R}^2$3, and missing approximately $\mathbb{R}^2$4–$\mathbb{R}^2$5. Random seeding also yields the greatest diversity of full path classifications, with percentages of unique full-path classifications that exceed Combined and Closeness on all three vessels.

In engineering terms, WPgen is positioned as a simulation-based testing instrument. At design time it can populate regression suites with adaptation-triggering scenarios. In CI it can be rerun against updated controllers to verify that adaptation logic and verification mechanisms continue to detect and handle adverse paths near operational routes. During safety assessment it can flag waypoint configurations likely to induce instability or missed waypoints, thereby informing pre-mission adjustment and configuration of trigger thresholds and fallback plans. The paper explicitly notes threats to validity, including the limited set of autonomous vessels, the use of a single simulation environment, and the use of an autocorrelation threshold chosen after exploratory studies.

5. Web presentation generation from HTML

A separate line of work permits interpreting WPgen as Web Rendering Parameters Generation, or WebRPG, a task that automates the generation of visual presentation for web pages directly from HTML code (Shao et al., 2024). The problem is formulated over a DOM tree $\mathbb{R}^2$6 with elements $\mathbb{R}^2$7. The output is a set of rendering parameters for each element, $\mathbb{R}^2$8, and the learned function maps HTML to predicted rendering parameters, $\mathbb{R}^2$9. The method does not invent new HTML structure, widgets, or high-level templates, and it does not produce unconstrained CSS with arbitrary selectors, inheritance, or overrides. Instead, it predicts a fixed vocabulary of per-element visual properties recoverable to CSS.

The predicted rendering parameters comprise $\mathbb{R}^3$0 common CSS properties in three groups: layout parameters $\mathbb{R}^3$1; text parameters $\mathbb{R}^3$2; and color parameters $\mathbb{R}^3$3. Property values are standardized to a discrete vocabulary of size $\mathbb{R}^3$4. Positions and sizes use integer pixels over indices $\mathbb{R}^3$5–$\mathbb{R}^3$6 on a $\mathbb{R}^3$7 browser canvas. Colors are mapped to a palette of $\mathbb{R}^3$8 frequent RGBA colors. A PAD token occupies index $\mathbb{R}^3$9. This discretization converts the prediction problem into element-wise categorical generation over a fixed output space.

The benchmark is derived from the Klarna product page dataset, consisting of $\beta$0K English product pages from $\beta$1K e-commerce sites stored in MHTML for high-fidelity offline rendering. Headless Chrome and Selenium are used to render pages, extract visible HTML, capture computed CSS values from the browser, and traverse the DOM in pre-order. Large pages are segmented into sub-pages, uncommon tags and intricate widgets are removed, and sub-pages are relocated to the top-left of the canvas to standardize coordinates. After preprocessing and filtering by a Visual Complexity criterion that excludes samples with $\beta$2, the resulting dataset contains $\beta$3 sub-page samples with an $\beta$4 train/test split.

The model combines an HTML embedding with VAE-based rendering-parameter compression. Each element receives a $\beta$5 embedding

$\beta$6

where semantic information comes from frozen pre-trained MarkupLM_large, hierarchical information comes from XPath embeddings, and character-count information encodes text length. A VAE compresses each element’s rendering-parameter vector into a latent $\beta$7, avoiding the instability that would arise from flattening large numbers of categorical properties into a long one-dimensional sequence.

Two baselines are presented. WebRPG-AR uses a Transformer encoder-decoder with $\beta$8 layers each and approximately $\beta$9M parameters, together with masked latent augmentation inspired by BART and MaskGIT. WebRPG-DM uses a $\delta_{\min}$0-layer U-ViT latent diffusion model with a DDPM schedule over $\delta_{\min}$1 steps. Both are trained jointly with the VAE and auxiliary embeddings, using AdamW with learning rate $\delta_{\min}$2, $\delta_{\min}$3, $\delta_{\min}$4, batch size $\delta_{\min}$5, and $\delta_{\min}$6M training steps.

Evaluation uses Fréchet Inception Distance in the feature space of a binary classifier trained to distinguish real from corrupted rendering parameters, Elements Intersection over Union, Style Consistency Score, rendered-image FID, case studies, and a human pairwise preference evaluation with five annotators on $\delta_{\min}$7 random samples. On the test split, WebRPG-AR achieves $\delta_{\min}$8, layout FID $\delta_{\min}$9, $\Delta w_i=w_i'-w_i$0, style FID $\Delta w_i=w_i'-w_i$1, and $\Delta w_i=w_i'-w_i$2, whereas WebRPG-DM yields markedly worse values: $\Delta w_i=w_i'-w_i$3, layout FID $\Delta w_i=w_i'-w_i$4, $\Delta w_i=w_i'-w_i$5, style FID $\Delta w_i=w_i'-w_i$6, and $\Delta w_i=w_i'-w_i$7. Among ablations, removing the VAE and flattening rendering parameters to a one-dimensional sequence degrades performance to $\Delta w_i=w_i'-w_i$8, layout FID $\Delta w_i=w_i'-w_i$9, $i=2,\dots,n$0, style FID $i=2,\dots,n$1, and $i=2,\dots,n$2, indicating that latent compression is structurally important.

Within this interpretation, WPgen is a constrained visual-generation system: it predicts fixed, DOM-aligned rendering parameters rather than free-form stylesheets. This suggests a conception of automated web design in which semantic and hierarchical information extracted from HTML are sufficient to recover a substantial portion of coherent visual presentation without explicit template synthesis.

6. Event generation in same-sign $i=2,\dots,n$3-boson scattering

A further usage of WPgen appears in particle-physics software practice, where it denotes an event-generation workflow for same-sign $i=2,\dots,n$4-boson scattering at the LHC implemented in POWHEG-BOX-RES with RECOLA2 (chiesa et al., 2019). The process is

$i=2,\dots,n$5

with $i=2,\dots,n$6. The calculation is complete at order $i=2,\dots,n$7 for the Born amplitude and includes full off-shell effects, leptonic decays, the full gauge-invariant set of electroweak diagrams, and NLO electroweak corrections at order $i=2,\dots,n$8, consisting of virtual one-loop electroweak corrections and real photon emission. Photon-induced initial states and real heavy-boson emission are not included.

The implementation is based on the POWHEG formalism with FKS subtraction for QCD and QED singularities and hardest-emission generation through a positive-definite underlying-Born weight. In POWHEG-BOX-RES, Sudakov factors are built as products of resonance- and production-level terms, with resonance-aware mappings that preserve intermediate resonance masses. Virtual amplitudes are provided by RECOLA2 in the complex-mass scheme with on-shell renormalization of physical fields and the $i=2,\dots,n$9 scheme for the electromagnetic coupling. Infrared safety with massless leptons is ensured through dressed leptons and photon–jet recombination.

The tool generates unweighted Les Houches events with NLO electroweak accuracy matched to a QED parton shower and interfaced to a QCD parton shower. A native interface to PYTHIA 8, tested with version 8.235, provides QED and QCD showering, hadronisation, and hadron decays. Because NLO QCD corrections are not included in the module itself, the starting scale for showers off colored partons is set to $W=\{w_1,\dots,w_n\}$00 rather than to the POWHEG photon transverse momentum.

The phenomenological impact is substantial. In the fiducial phase space at $W=\{w_1,\dots,w_n\}$01 TeV, the paper reports for $W=\{w_1,\dots,w_n\}$02 a leading-order cross section of $W=\{w_1,\dots,w_n\}$03 fb and an NLO electroweak value of $W=\{w_1,\dots,w_n\}$04 fb, corresponding to $W=\{w_1,\dots,w_n\}$05 and $W=\{w_1,\dots,w_n\}$06. For $W=\{w_1,\dots,w_n\}$07, the corresponding numbers are $W=\{w_1,\dots,w_n\}$08 fb and $W=\{w_1,\dots,w_n\}$09 fb, with $W=\{w_1,\dots,w_n\}$10 and $W=\{w_1,\dots,w_n\}$11. In differential distributions, negative electroweak Sudakov effects reach approximately $W=\{w_1,\dots,w_n\}$12 to $W=\{w_1,\dots,w_n\}$13 in the high-energy tails of $W=\{w_1,\dots,w_n\}$14 and missing transverse momentum. Validation against MoCaNLO+RECOLA yields agreement within statistical uncertainties.

This use of WPgen differs fundamentally from the maritime and web-design cases. It is not an optimizer over perturbations or a predictor over DOM-aligned properties, but a Monte Carlo event generator that preserves gauge invariance, infrared safety, and shower consistency. The commonality lies instead in the controlled generation of technically valid structured outputs under stringent domain constraints.

7. Cross-domain comparison and conceptual significance

Across its distinct meanings, WPgen consistently refers to generation constrained by an existing formal object. In maritime autonomy, the generator perturbs a predefined route within bounded neighborhoods and minimum-separation constraints. In web presentation, it predicts a standardized fixed vocabulary of rendering parameters aligned to DOM elements. In collider phenomenology, it generates unweighted events for a specified scattering process while respecting perturbative accuracy, subtraction structure, and shower-matching requirements (Nylænder et al., 22 Jul 2025).

These systems therefore stand in contrast to unconstrained synthesis. The maritime version does not invent missions but searches nearby routes that expose adaptation failures. The web version does not synthesize arbitrary CSS selectors or new page structure but predicts per-element properties from HTML. The collider version does not generate arbitrary particle-level patterns but samples from a rigorously specified quantum-field-theoretic calculation. A plausible implication is that the recurring appeal of the WPgen label lies in domains where the chief technical challenge is not open-ended creativity but constrained generation under correctness, feasibility, or fidelity requirements.

The literature also reveals that success depends strongly on representation choice. Maritime WPgen benefits from waypoint-aware operators and seeding strategies tailored to the vehicle type. WebRPG succeeds when rendering parameters are compressed through a VAE and conditioned on semantic, hierarchical, and character-count embeddings. The collider generator succeeds through resonance-aware FKS subtraction, complex-mass renormalization, and carefully coordinated QED/QCD shower interfaces. In each case, the generative mechanism is effective precisely because it is tightly coupled to the structure of the underlying domain rather than abstracted away from it.