Preference-Integrated Design

• Preference-Integrated Design is a framework that systematically embeds stakeholder and human preferences into computational design to balance technical performance with subjective desirability.
• It employs methods like aggregated optimization, Bayesian modeling, and direct preference optimization to create unique, optimal solutions under complex constraints.
• This approach is applied in diverse fields—from UI and product design to molecular and infrastructure design—enhancing personalization, efficiency, and user trust.

Preference-Integrated Design refers to a class of rigorous, computational frameworks and methodologies for systematically incorporating human or stakeholder preferences into the automated or semi-automated design of systems, artifacts, or artifacts' configurations. These frameworks span engineering, product aesthetics, materials, interfaces, and infrastructure design, and aim to align design outputs with individual, group, or societal value functions while often subject to nontrivial constraints. The integration mechanism can be explicit (via preference models or direct query) or implicit (via learned signals from user actions, legacies, or physiological/behavioral measurements), and is central to advancing the efficacy, personalization, and adoption success of contemporary design pipelines.

1. Foundations and Motivation

Traditional design optimization and engineering workflows have relied on either single-objective or Pareto-based multi-objective optimization, frequently ignoring the explicit and quantitative modeling of user or stakeholder preferences, or else aggregating objectives without appropriate cardinal scaling. This omission leads to dissonance between what technical systems can deliver (“object capability”) and what stakeholders actually value (“subject desirability”) (Heukelum et al., 2023). Classical Pareto-front approaches provide sets of optimal trade-offs but do not yield a unique best-fit solution reflecting true preference integration, often leaving critical design decisions unresolved or requiring post-hoc intervention (Heukelum et al., 2023).

Preference-integrated design frameworks address these limitations by:

• Translating subjective objectives (e.g., satisfaction, aesthetic value) and technical performance (e.g., cost, energy, efficiency, structural integrity) into a unified preference domain.
• Modeling individual or collective stakeholder preference functions, possibly heterogeneously and contextually.
• Operating directly, or via surrogates, within this domain to achieve unique, defensible best-fit design solutions.

Applications range from user interface layout (Song et al., 23 Sep 2025), product form-function optimization (Kang et al., 2019), infrastructure and socio-technical systems (Heukelum et al., 2023, Teuber et al., 2024), and molecular/materials design (Astudillo et al., 2019, Xue et al., 30 May 2025, Sun et al., 24 Oct 2025, Cheng et al., 2024, Zhou et al., 2024), to collaborative fashion design (Shao et al., 21 Jan 2026).

2. Mathematical Formulations and Algorithms

Preference-integrated design methodologies instantiate across several mathematical and computational paradigms. Salient approaches include:

Multi-Attribute and Aggregated Preference Optimization

• Integrative Maximization of Aggregated Preferences (IMAP): Odesys/Preferendus formalizes the maximization of a group-aggregated preference function:

$$\max_{\mathbf{x}} U(\mathbf{x}) = T\bigl[ \{P_{k,i}(O_i(F_1(\mathbf{x},\mathbf{y}),...,F_J(\mathbf{x},\mathbf{y}))) ,\,w'_{k,i}\}_{k=1..K,\,i=1..I}\bigr]$$

where $P_{k,i}$ are stakeholder $k$'s preference functions for objective $i$, $w'_{k,i}$ are combined global and local weights, $T[\cdot]$ is often a least-squares aggregator, and $F_j$ are engineering performance functions (Heukelum et al., 2023, Teuber et al., 2024).

• Preference Modeling for Bayesian Optimization: Multi-attribute and multi-objective black-box optimization under unknown utility functions $U(\mathbf{y})$ is managed by explicit probabilistic modeling of the utility, frequently using a Bayesian or GP prior over $U$, with learning via pairwise comparisons and utility-uncertainty-aware acquisition functions such as EI-UU (Astudillo et al., 2019, Sun et al., 2019).

Direct Preference Optimization (DPO) and Variants

• Sequence and Structure-Generation Alignment: For design problems in molecular biology, generative policy models (e.g., protein sequence generators, RNA designers) are fine-tuned using DPO—training on preference pairs $(x_A, x_B)$ where $x_A$ is judged (via physical or learned surrogate) as superior—by minimizing:

$$L_{\rm DPO} = -\mathbb{E}_{(x_A, x_B)} \log \sigma (s_\theta(x_A) - s_\theta(x_B))$$

with $s_\theta(\cdot)$ the model score head and $\sigma$ the logistic function (Xue et al., 30 May 2025, Sun et al., 24 Oct 2025, Cheng et al., 2024, Zhou et al., 2024).

• Decomposition and Residue-Level Optimization: For decomposable objectives in biomacromolecule or drug design, DPO is extended to substructure or per-residue granularity (e.g., ResiDPO, DecompDPO) to provide fine-grained credit assignment (Xue et al., 30 May 2025, Cheng et al., 2024, Zhou et al., 2024).

Preference Model Learning and Hybrid Architectures

• Dimensionality-Reduced Utility Learning: High-dimensional aesthetic or functional design spaces can be embedded via multi-modal VAEs and Siamese networks, with per-individual and ensemble preference models over learned latent features to predict pairwise and absolute preferences (Shin et al., 2024).
• User-Conditioned Generative Models: In product or visual design, generative adversarial architectures can be conditioned on predicted preference (e.g., predicted popularity, attribute consensus scores, personalization vectors) to direct the generation toward user-aligned outputs (Li et al., 2024, Peng et al., 25 Nov 2025, Shao et al., 21 Jan 2026).

Metaheuristics with Preference Injection

• Preference-Based Multi-Objective Metaheuristics (PMOMH): Metaheuristic optimization frameworks (e.g., NSGA-II, MOEA/D) are modified by integrating explicit preference models (reference-point, utility, desirability functions, cones), injected into dominance relations, crowding distances, or set-sorting mechanisms, and categorized systematically in OWL ontologies (Li et al., 2016).

3. Preference Elicitation and Interaction Modalities

Eliciting and encoding preferences is central. Established protocols include:

• Choice-Based Conjoint Analysis (CBCA): Stakeholders select among design alternatives in controlled tasks; logistic regression or multinomial logit models estimate part-worths, converted into preference weights $w_{k,i}$ (Teuber et al., 2024).
• Bi-level Query Tasks: For coupled form-function problems, isolated queries for visual form versus overall system trade-off provide higher-fidelity estimation of preference structure (Kang et al., 2019).
• Interactive Adjustment: In UI or collaborative co-design, users directly manipulate design artifacts; their local adjustments are mapped into quantitative priority tiers or attribute-specific consensus scores, feeding back into probabilistic or lexicographic MOO (Song et al., 23 Sep 2025, Shao et al., 21 Jan 2026).
• Implicit Attention and Behavioral Signals: Eye-tracking in VR environments, dwell time, or gaze heatmaps are mapped to normalized preference weights and linked to explicit ratings for preference induction (Sankar et al., 27 Aug 2025).
• Legacy Data Fusion: Numerical data from disparate experimental conditions are merged by converting to within-dataset pairwise preferences; GP training over these comparisons yields calibration-free, information-preserving surrogate models (Sun et al., 2019).

4. Implementations, System Architectures, and Empirical Validation

Preference-integrated frameworks have been realized in the following system types:

• Socio-Technical Design Platforms: Tools such as Odesys/Preferendus implement the IMAP approach using interactive GA solvers, glass-box equation-sharing, and transparent preference aggregation, operationalized in infrastructure and wind farm design (Heukelum et al., 2023, Teuber et al., 2024).
• Generative Product Design Pipelines: Hybrid systems combine preference-predictive models with generators (GANs, diffusion models), leveraging both internal and external (UGC) data and user interaction; cold-start bootstrapping is supported by external preference signals (Li et al., 2024).
• Interactive Visual Design and Co-creation: DesignBridge fuses user region-level feedback and consensus scoring with designer-curated fine-tuning, leveraging LoRA for conditional diffusion and attribute-guided generation, with real-time explainability layers (e.g., SHAP) (Shao et al., 21 Jan 2026).
• Personalized Visual Design Evaluation: Datasets such as DesignPref establish the variability in human preference, showing the necessity (and gains) of per-individual judge models over aggregated scoring, with personalization yielding significant increases in predictive accuracy (Peng et al., 25 Nov 2025).
• Metaheuristic Infrastructure: Ontology-driven knowledge bases systematize the PMOMH landscape, supporting application-specific querying and method selection (Li et al., 2016).

Empirical studies report gains such as increased sampling efficiency in sequence design (pass@64 increase), higher satisfaction and reduced manual effort in UI adaptation, improved classification accuracy for personalized product recommendation, and higher expert-rated worthiness and effectiveness in form design (Sun et al., 24 Oct 2025, Song et al., 23 Sep 2025, Shin et al., 2024, Sankar et al., 27 Aug 2025, Shao et al., 21 Jan 2026).

5. Practical Trade-Offs, Challenges, and Future Directions

Key strengths of preference-integrated design include:

• Production of unique, a priori best-fit solutions, overcoming ambiguity inherent in Pareto or a posteriori selection (Heukelum et al., 2023).
• Integration of heterogeneous, possibly high-dimensional, and highly personalized or group-level preferences into a unified, cardinal preference domain (Teuber et al., 2024, Peng et al., 25 Nov 2025).
• Transparent, glass-box modeling supporting stakeholder trust, negotiation, and validation (Teuber et al., 2024).

However, limitations persist:

• Quality and fidelity of final solutions depend strongly on accurate and representative preference elicitation; dynamic, interactive protocols and real-time retraining are ongoing areas of research (Teuber et al., 2024, Peng et al., 25 Nov 2025).
• Calibration and meta-modeling challenges arise in combining external preference signals; careful within-dataset conversion is necessary (Sun et al., 2019).
• Scalability of preference learning to high-dimensional decision spaces and large user bases requires further innovation (e.g., trust-region approaches, sparse GP surrogates, or meta-learning) (Shin et al., 2024, Peng et al., 25 Nov 2025).

Future directions include probabilistic and continual learning for long-term preference drift, augmentation with behavioral and physiological feedback in real time, and integration of standardized performance indicators for benchmarking across preference-integrated frameworks (Sankar et al., 27 Aug 2025, Li et al., 2016).

6. Application Domains and Representative Systems

Domain	Paradigm(s) Employed	Notable Systems/Papers
Socio-technical infrastructure	IMAP, Odesys, CBCA	(Heukelum et al., 2023, Teuber et al., 2024)
Visual/UI Design	Personalized preference models, RAG	(Peng et al., 25 Nov 2025, Shao et al., 21 Jan 2026, Song et al., 23 Sep 2025)
Product Form-Function	Bi-level adaptive query, HB learning	(Kang et al., 2019, Shin et al., 2024)
Bio/Materials Design	DPO, ResiDPO, DecompDPO, RLPF	(Xue et al., 30 May 2025, Cheng et al., 2024, Sun et al., 24 Oct 2025, Zhou et al., 2024)
Data-driven Product Design	Preference-predictor + CcGAN	(Li et al., 2024)

Across these settings, preference-integrated design has moved from conceptual frameworks and simulation to operational deployment in commercial, public, and scientific settings.

7. Contributions to Theory and Practice

Preference-integrated design frameworks have demonstrated:

• Establishment of new standards for socio-technical synthesis, in both group and highly personalized settings.
• Advancement of methodologies that align generative models, simulators, and optimization solvers with multidimensional, multimodal, and sometimes uncertain or latent human preferences, often at superhuman or super-expert scale.
• Practical deployment in fields ranging from critical infrastructure to user interface personalization, molecular engineering, and co-creative tools.

The result is a substantial increase in the alignment of design outputs with stakeholder utility, system performance, and decision-maker satisfaction, as well as transparent, participatory, and scalable workflows for the next generation of intelligent design systems (Astudillo et al., 2019, Peng et al., 25 Nov 2025, Teuber et al., 2024).