Data-Centric Solution Preference

• Data-centric solution preference is a strategy that prioritizes enhancing data quality and selection to achieve system objectives without modifying model architectures.
• It employs methodologies such as self-generated preference data, model-guided selection, and automated curation to improve efficiency, interpretability, and robustness.
• This approach applies across domains like code optimization, recommender systems, federated learning, and governance, delivering measurable improvements in performance and cost efficiency.

A data-centric solution preference is the principle of prioritizing interventions on the data itself—such as the creation, selection, augmentation, curation, or annotation of data and preference signals—to achieve system or model objectives, rather than relying primarily on algorithmic or architectural modifications. This paradigm shift is found across diverse research domains, from code optimization and recommender systems to ontology-driven computing, federated graph learning, and language or vision model alignment. Data-centric solution preference is characterized by methodologies that synthesize, select, or enhance preferences directly within the data, often leveraging task-specific targets, quality assessments, or user-centric trade-offs. By reframing optimization as a data engineering problem, practitioners achieve improved generalization, efficiency, interpretability, and robustness without heavy dependence on model changes.

1. Principles and Motivations

Data-centric solution preference is motivated by several recurring limitations in model- or algorithm-centric workflows:

• Model improvements often plateau if the underlying data is noisy, sparse, unrepresentative, or poorly annotated.
• Real-world deployments expose models to out-of-distribution data or operational constraints (e.g., efficiency, security) that are difficult to encode purely decoratively.
• Domain shifts, user requirements, or governance mandates typically manifest at the data level, necessitating strategies that dynamically select, generate, or correct training data and preference signals.

The data-centric approach thus foregrounds questions such as:

• Which samples, preference pairs, or modalities are most informative or robust for a given task?
• How can preference data be generated to reflect multifaceted objectives or user privacy constraints?
• What quantitative criteria guide the selection, filtering, or augmentation of data?

Specific motivations in recent research include:

• Achieving joint correctness and efficiency in code generation (Gee et al., 2024).
• Systematically managing preference signal redundancy, label noise, and contrast in RLHF (Shen et al., 2024).
• Enabling scalable governance, auditability, and compliance through dataset-specification and evaluation (McGregor et al., 2023).
• Addressing heterogeneity, privacy, and collaboration constraints in federated graph learning (Wu et al., 22 Jul 2025).
• Bridging semantic integration and fine-grained access control in distributed ontologies (Johnson et al., 2024, Knowles et al., 2024).

2. Data-Centric Preference Generation and Selection

A core application of data-centric solution preference is the self-generation or selection of preference data that directly guides optimization objectives. This manifests in several general approaches:

2.1 Self-Generated Preference Data

Frameworks like Code-Optimise synthesize preference labels along axes matched to operational goals. For code LMs, every problem instance is associated with generated solutions, which are then labeled for:

• Functional correctness (passed vs. failed) via repeatable, statistically robust unit-test execution.
• Efficiency (quick vs. slow) based on strictly defined runtime statistics, e.g., stable mean runtime with low coefficient of variation.

Preference pairs are explicitly constructed—e.g., for each problem, (passed, failed) and (quickest, slowest passing)—and only instances admitting both kinds of discrimination are retained for optimization. No changes to model architecture are needed; the learning signals arise solely from the self-engineered data (Gee et al., 2024).

2.2 Model-Guided Data Selection

Recent work emphasizes the efficiency of learning from carefully selected preference examples. For instance, a model's uncertainty (or implicit reward gap) on a pair determines its informativeness: pairs with small DPO implicit reward gaps are selected preferentially, as they yield maximal gradient magnitude and information gain (Qi et al., 6 Aug 2025).

In the context of multi-aspect preference signals, selection is guided by the Preference Divergence (PD) term, which quantifies the degree of consensus or conflict across aspects for each sample. Selecting samples with the most negative PD values—i.e., high inter-aspect consensus—produces subsets that optimize overall alignment objectives, as proven by bounds on the DMPO loss (Zhang et al., 11 Aug 2025).

2.3 Automated Data Curation and Filtering

Analyses such as the Magpie framework systematically annotate open preference corpora for input/task quality, task category, and alignment with independent reward models. Datasets are then curated to maximize reward-aligned consensus, diversity, and coverage, removing redundant, noisy, or low-informative samples (Djuhera et al., 14 Nov 2025). This approach often yields smaller, higher-quality training sets with superior downstream performance and computational efficiency.

3. Data-Centric Preference Optimization Objectives

With preference data or signals in place, optimization proceeds using model-agnostic, data-driven objectives:

• Supervised Fine-Tuning (SFT): Cross-entropy over data selected for specific preference characteristics (e.g., fastest passing code solutions) (Gee et al., 2024).
• Direct Preference Optimization (DPO): Minimization of a loss that encourages the model to assign higher likelihood to the preferred sample in each pair or triplet, with reference to a fixed baseline model:

$$L_{DPO}(\pi_\theta; \pi_{ref}) = -\mathbb{E}_{(x, y_w, y_l)}\left[ \log \sigma\left( \beta \left[\log \frac{\pi_\theta(y_w|x)}{\pi_{ref}(y_w|x)} - \log \frac{\pi_\theta(y_l|x)}{\pi_{ref}(y_l|x)}\right]\right) \right]$$

where preference axes (e.g. correctness, efficiency, entity-centric contrast) can be directly encoded in pairwise data (Gee et al., 2024, Wu et al., 4 Jun 2025).

• Augmented Objectives: Preference learning can be made more data-centric by including rationales or aspect-specific annotations, with likelihood terms for both the preference and associated explanations, leading to improved data efficiency, shorter outputs, and higher reliability (Just et al., 2024).

4. Applications Across Domains

4.1 Code Optimization

Data-centric preference signals in code generation enable simultaneous improvements in correctness (pass@k), efficiency (runtime reduction), and code brevity by using self-generated, dual-axis (correct–incorrect, quick–slow) data (Gee et al., 2024). This is achieved purely through modification of training data, with no model architecture changes.

4.2 Recommender Systems

Data-centric approaches systematically identify and address limitations in recommendation data: incompleteness (through imputation or augmentation), noise (through reweighting or filtering), and bias (with propensity reweighting or causal adjustment). Explicit metrics and audit workflows are recommended to guide the selection and tuning of data-centric operations (Lai et al., 2024).

4.3 Federated and Graph Learning

A two-level taxonomy—data characteristics and data utilization—categorizes solutions for federated graph learning that align with varying graph formats, decentralization/visibility schemes, and challenges such as data heterogeneity, label imbalance, or communication constraints. The primary emphasis is on transformations, mending, and cross-silo harmonization of data shards, rather than on model modifications (Wu et al., 22 Jul 2025).

4.4 Governance and Ontology

Data-centric governance operationalizes requirements (fairness, safety, robustness) with evaluation datasets and metrics integrated throughout lifecycle pipelines (McGregor et al., 2023). Data-centric ontology engineering (quadrimodal: object, event, concept, action) serves to enforce auditability, interoperability, and role-based access consistently at the data layer, decoupling permission and provenance management from infrastructure or node identity (Johnson et al., 2024, Knowles et al., 2024).

4.5 Multimodal and RLHF Optimization

Entity-centric preference optimization in vision-LLMs, as well as rationale-enriched preference learning in RLHF, are achieved by constructing high-contrast, semantically rigorous negative examples, and by leveraging machine-generated rationales or multi-aspect consensus metrics. These strategies systematically reduce hallucination, boost instruction-following, and maximize information gain per data sample (Wu et al., 4 Jun 2025, Just et al., 2024).

5. Metrics, Empirical Findings, and Best Practices

Quantitative, model-agnostic criteria are central to data-centric solution preference:

• Scale: Empirical returns to increasing the data volume, with saturation and sometimes negative generalization beyond certain sample counts (Shen et al., 2024).
• Label Noise Invariance: Preference datasets can tolerate up to 30–40% label flipping with modest performance loss, suggesting efficient allocation of annotation budget (Shen et al., 2024).
• Information Content: High-contrast (low-similarity) preference pairs carry more training signal, especially for small models. For larger models, diversity and task coverage outweigh pairwise contrast (Shen et al., 2024, Djuhera et al., 14 Nov 2025).

Empirical results robustly demonstrate that:

• Subsets chosen by difficulty-based or consensus-based selection often match or surpass performance of full datasets with far less computation (Qi et al., 6 Aug 2025, Zhang et al., 11 Aug 2025).
• Curation that filters for reward-model-validated preference alignment, task diversity, and prompt quality yields both data and compute savings while improving benchmark scores (Djuhera et al., 14 Nov 2025).
• Data-centric governance reduces deployment time (30–50%), improves solution quality (20–40%), and cuts incidents by 60–80% within the first year (McGregor et al., 2023).

Recommended best practices include:

• Formal audit and annotation of data quality, coverage, and label consensus prior to model-centric tuning.
• Preference for iteratively constructing, selecting, or filtering preference data based on quantitative criteria directly linked to optimization objectives and deployment context.
• Integration of continuous evaluation, data stewardship, and, where applicable, active preference or rationale generation into the ML pipeline.

6. Limitations and Trade-Offs

• Initial investment in data-centric workflows (e.g., ontology definition, preference signal engineering, quality audits) is non-trivial compared to ad hoc model-centric experimentation (Johnson et al., 2024, Djuhera et al., 14 Nov 2025).
• The robustness and cost-effectiveness of these approaches are context-dependent; extremely noisy or adversarial domains may require hybrid model-data interventions for optimal performance.
• Automated or self-generated preference signals inherit possible failure modes of the underlying models used for selection/generation.
• Maintaining up-to-date data flow graphs, ontologies, and evaluation datasets necessitates dedicated governance and versioning infrastructure.

7. Outlook and Future Directions

Emerging frontiers in data-centric solution preference include:

• Systematic integration of modular, extensible ontologies at the core of computational ecosystems for seamless auditability, provenance tracking, and access control (Knowles et al., 2024).
• Hybrid approaches that combine preference signal curation with model-centric active learning and adaptive data selection (Shen et al., 2024, Qi et al., 6 Aug 2025).
• Cross-modal, federated, and privacy-preserving extensions, with strong focus on incentivizing collaboration while maintaining data integrity, interpretability, and regulatory compliance (Wu et al., 22 Jul 2025).
• Automated curation infrastructure leveraging independent judges or reward models for scalable, robust preference data construction and continual evaluation (Djuhera et al., 14 Nov 2025).

In summary, data-centric solution preference recasts optimization not merely as a modeling problem, but as an integrated engineering discipline focused on the generation, selection, and curation of data and preferences. This perspective is now substantiated by theory and large-scale empirical success in coding, recommendation, RLHF, multimodal AI, governance, and distributed computation, and provides foundational methodologies for future scalable, robust, and interpretable AI systems.