Model-Data Feature Matcher Agent

• Model-Data Feature Matcher Agent is a dynamic system that aligns data representations with model criteria by selecting, generating, and routing features for optimal task performance.
• It employs techniques such as cosine similarity, LLM-based generation, and multi-agent coordination to handle diverse settings like tabular data, schema matching, and clinical note extraction.
• Empirical evidence shows these agents can outperform traditional methods in accuracy and efficiency, though issues like computational overhead and privacy remain challenges.

A Model-Data Feature Matcher Agent denotes an agentic system that mediates between data representations and downstream models by selecting, generating, extracting, standardizing, or routing features so that the resulting representation better aligns with model requirements or improves task performance. In current arXiv literature, this role appears in agentic feature augmentation for tabular learning, latent-space expert routing in split learning, schema and header matching for data integration, model selection in clinical pipelines, and automated extraction of structured variables from unstructured records (Gong et al., 21 May 2025, Sharma et al., 2019, Shimgekar et al., 24 Jul 2025, Wang et al., 3 Aug 2025).

1. Conceptual scope

The defining property of a Model-Data Feature Matcher Agent is that it does not treat feature engineering as a purely static preprocessing stage. Instead, it explicitly links a representation on the data side to an operational criterion on the model side. In the tabular setting, that criterion can be a task score such as F1, Accuracy, $R^2$, or $1-\mathrm{MSE}$; in split learning it can be compatibility with a server-side expert; in schema matching it can be semantic correspondence between fields; and in clinical databases it can be successful extraction of model-required variables from heterogeneous schemas (Gong et al., 21 May 2025, Sharma et al., 2019, Khilji, 2023, Lee et al., 1 Oct 2025).

Across these works, the matched object is not uniform. It may be an evolving feature set represented as postfix token sequences, a hidden representation $x' \in \mathbb{R}^d$, a pair of feature names and descriptions, a set of required headers in a curated model repository, or a standardized clinical concept that must be localized in local code systems. This breadth distinguishes the topic from narrower notions such as classical feature selection or schema matching alone.

A common misunderstanding is to equate the topic with feature selection only. The literature is broader: some systems interleave feature generation and pruning, some route client representations to expert models, some score field correspondences using semantic similarity, and some convert raw notes or EMR schemas into model-ready structured variables. Taken together, this suggests that “matching” is best understood as a family of alignment mechanisms between model assumptions and data representations rather than a single algorithmic primitive.

2. Formal objectives and matching primitives

In agentic feature augmentation, the current feature set $\mathcal{F}_{i,j}$ is modified either by generation or by selection. Generation applies an operator $O_T \in \mathcal{O}$ to produce new columns,

$$\mathcal{F}_{i,j+1} = \mathcal{F}_{i,j} \cup O_T(\mathcal{F}_{i,j}),$$

whereas selection applies a binary mask $\mathbf{s}_{i,j}\in\{0,1\}^{|\mathcal{F}_{i,j}|}$,

$$\mathcal{F}_{i,j+1} = \mathcal{F}_{i,j} \odot \mathbf{s}_{i,j}.$$

The global objective is

$$\mathcal{F}^* = \arg\max_{i,j} \mathcal{S}(\mathcal{F}_{i,j}, \mathcal{Y}),$$

which turns matching into a search over feature sets under model-based feedback (Gong et al., 21 May 2025).

In latent-space expert routing, the primitive is not feature transformation but similarity-based assignment. A client produces a hidden representation $x'$, the server stores dataset-level centroids $1-\mathrm{MSE}$0 and class-level centroids $1-\mathrm{MSE}$1, and routing is performed by cosine similarity: $1-\mathrm{MSE}$2 The matcher therefore operates as a nearest-centroid router in representation space, with dataset-level coarse assignment and class-level fine-grained assignment (Sharma et al., 2019).

In schema and metadata matching, the primitive is pairwise similarity over text rather than over observed feature values. One hybrid formulation combines cosine similarity over pretrained BERT embeddings with Jaccard similarity over lemmatized token sets: $1-\mathrm{MSE}$3 with manual weighting options $1-\mathrm{MSE}$4, $1-\mathrm{MSE}$5, and $1-\mathrm{MSE}$6, and a default threshold $1-\mathrm{MSE}$7 for retaining candidate correspondences (Khilji, 2023).

In clinical model selection for tabular data, the matcher uses SapBERT embeddings of headers, greedy one-to-one assignment, and an empirical cosine threshold $1-\mathrm{MSE}$8. A model is eligible only if every required header can be matched to a dataset header above threshold; an LLM then selects among eligible models using textual metadata. For images, the same framework replaces headers with MedGemma-derived Modality and Disease Type and aligns them to model modality and caption fields in a curated repository (Shimgekar et al., 24 Jul 2025).

3. Agentic architectures

A representative architecture is MAGS, a Multi-Agent System with Long and Short-Term Memory. It contains a router agent that chooses between generation and selection, a generator agent that crosses features through unary and binary operators, a selector agent that prunes redundant or irrelevant features, and a memory pool storing full trajectories of actions, resulting feature sequences, and downstream scores. The router is optimized with offline PPO, while generator and selector are LLM-based agents using in-context learning with short-term memory for within-iteration adaptation and long-term memory for cross-iteration demonstrations (Gong et al., 21 May 2025).

Other systems decompose the same matching problem differently. SNOW uses a Feature Discovery Agent, Feature Extraction Agent, Feature Validation Agent, Post-Processing Agent, and Aggregation Code Generator to convert unstructured clinical notes into structured, interpretable features suitable for logistic regression; FELA uses Idea Agents, Code Agents, Critic Agents, and an Evaluation Agent to evolve feature ideas and implementations for industrial event logs; LLM-FS-Agent structures feature selection as a role-based debate among Initiator, Refiner, Challenger, and Judge agents, with the Judge producing a final score

$1-\mathrm{MSE}$9

(Wang et al., 3 Aug 2025, Ouyang et al., 29 Oct 2025, Bal-Ghaoui et al., 7 Oct 2025).

A further architectural line treats feature selection as sequential reasoning under explicit operational constraints. MoFA defines a feature universe $x' \in \mathbb{R}^d$0, represents each feature as $x' \in \mathbb{R}^d$1, and selects features greedily under a constraint predicate $x' \in \mathbb{R}^d$2, using a divide-and-conquer two-phase procedure to fit large feature spaces into the context window. This design is aimed at production settings with feature-group budgets, capacity limits, and maintainability constraints rather than purely statistical relevance (Zhou et al., 26 Mar 2026).

4. Operating regimes and representative systems

The same matcher abstraction appears in several technically distinct regimes.

Regime	Matching substrate	Representative mechanism
Tabular feature augmentation	Current feature set and score trajectory	Router-guided interleaving of generation and selection over postfix token sequences (Gong et al., 21 May 2025)
Split learning and expert routing	Client hidden representation and server centroids	Cosine-similarity coarse and fine assignment in latent space (Sharma et al., 2019)
Schema and repository alignment	Feature names, descriptions, headers, modality/disease metadata	BERT+Jaccard field matching or SapBERT-based model eligibility checking (Khilji, 2023, Shimgekar et al., 24 Jul 2025)
Clinical text and EMR standardization	Notes, code tables, schema documentation, SQL observations	Multi-agent feature discovery from notes or SQL-based code mapping across EMR schemas (Wang et al., 3 Aug 2025, Lee et al., 1 Oct 2025)
Industrial logs and constrained production systems	Idea-feature hierarchies, feature groups, cost constraints	UCB-guided feature evolution or sequential constraint-aware feature selection (Ouyang et al., 29 Oct 2025, Zhou et al., 26 Mar 2026)

These regimes differ primarily in what is observed and what is acted upon. In tabular augmentation, the agent acts directly on the feature space through operators and masks. In split learning, it acts on model routing while leaving the representation fixed. In schema and EMR settings, it acts on the correspondence between semantic requests and local database structure. In clinical note processing, it acts by creating new structured variables from unstructured text. In industrial systems, it often operates under extra-system constraints such as feature-group ownership, inference efficiency, or online experimentation budgets.

This heterogeneity is important because it shows that the topic is not tied to one modality or one optimization loop. A plausible implication is that any setting in which model behavior depends on how data are represented can be recast as a model-data matching problem, provided the system has a way to observe compatibility and act on the representation or routing policy.

5. Empirical evidence

In unified tabular feature engineering, MAGS was evaluated on six datasets from LibSVM, UCI, and OpenML. Table 1 shows that it outperforms all feature-generation and feature-selection baselines on most metrics and datasets. On openml_586, the best baseline, ELLM-FT, achieved $x' \in \mathbb{R}^d$3 and $x' \in \mathbb{R}^d$4, whereas MAGS achieved $x' \in \mathbb{R}^d$5 and $x' \in \mathbb{R}^d$6. On messidor, the best baseline F1 was about $x' \in \mathbb{R}^d$7, while MAGS reached $x' \in \mathbb{R}^d$8. In robustness tests on messidor with GB, KNN, LR, MLP, SVM, and RF, MAGS was best for all six downstream models; for KNN, the best baseline was about $x' \in \mathbb{R}^d$9 and MAGS reached $\mathcal{F}_{i,j}$0. In the openml_586 case study, MAGS added 4 new features and removed 20 redundant ones, whereas ELLM-FT generated 17 new features with no deletion and FSNS only removed 18 features (Gong et al., 21 May 2025).

In privacy-constrained expert routing, the hidden-representation version of ExpertMatcher achieved coarse assignment accuracy comparable to the original raw-data-based method. For Client A, average coarse assignment accuracy was $\mathcal{F}_{i,j}$1 versus $\mathcal{F}_{i,j}$2 for the original method; for Client B, it was $\mathcal{F}_{i,j}$3 versus $\mathcal{F}_{i,j}$4. Fine-grained assignment was harder: on MNIST, Client A dropped from $\mathcal{F}_{i,j}$5 to $\mathcal{F}_{i,j}$6, and Client B from $\mathcal{F}_{i,j}$7 to $\mathcal{F}_{i,j}$8; on NLOS, Client A dropped from $\mathcal{F}_{i,j}$9 to $O_T \in \mathcal{O}$0, and Client B from $O_T \in \mathcal{O}$1 to $O_T \in \mathcal{O}$2. On Diabetic Retinopathy, however, the hidden-representation method was slightly better, with Client A moving from $O_T \in \mathcal{O}$3 to $O_T \in \mathcal{O}$4 and Client B from $O_T \in \mathcal{O}$5 to $O_T \in \mathcal{O}$6 (Sharma et al., 2019).

In clinical note feature generation, SNOW was evaluated for 5-year prostate cancer recurrence prediction on 147 patients. Manual clinician feature generation achieved the highest AUC-ROC, $O_T \in \mathcal{O}$7, but SNOW reached $O_T \in \mathcal{O}$8 without requiring clinical expertise, exceeding both baseline structured features alone, $O_T \in \mathcal{O}$9, and all representational feature generation approaches. The clinician-guided LLM method reached $$\mathcal{F}_{i,j+1} = \mathcal{F}_{i,j} \cup O_T(\mathcal{F}_{i,j}),$$0 (Wang et al., 3 Aug 2025).

In production feature selection, MoFA improved or preserved performance under operational constraints. For True Interest prediction at $$\mathcal{F}_{i,j+1} = \mathcal{F}_{i,j} \cup O_T(\mathcal{F}_{i,j}),$$1, Lasso achieved test AUC $$\mathcal{F}_{i,j+1} = \mathcal{F}_{i,j} \cup O_T(\mathcal{F}_{i,j}),$$2 using 48 groups, while MoFA achieved $$\mathcal{F}_{i,j+1} = \mathcal{F}_{i,j} \cup O_T(\mathcal{F}_{i,j}),$$3 using 45 groups. At $$\mathcal{F}_{i,j+1} = \mathcal{F}_{i,j} \cup O_T(\mathcal{F}_{i,j}),$$4, Lasso achieved $$\mathcal{F}_{i,j+1} = \mathcal{F}_{i,j} \cup O_T(\mathcal{F}_{i,j}),$$5 with 187 groups, while MoFA achieved $$\mathcal{F}_{i,j+1} = \mathcal{F}_{i,j} \cup O_T(\mathcal{F}_{i,j}),$$6 with 85 groups. In the Value Model enhancement setting, the first three interaction terms selected by MoFA produced statistically significant relative lifts in total daily app sessions of $$\mathcal{F}_{i,j+1} = \mathcal{F}_{i,j} \cup O_T(\mathcal{F}_{i,j}),$$7, $$\mathcal{F}_{i,j+1} = \mathcal{F}_{i,j} \cup O_T(\mathcal{F}_{i,j}),$$8, and $$\mathcal{F}_{i,j+1} = \mathcal{F}_{i,j} \cup O_T(\mathcal{F}_{i,j}),$$9. In notification behavior prediction, selecting 4,000 features from 8,169 produced offline NE wins of $\mathbf{s}_{i,j}\in\{0,1\}^{|\mathcal{F}_{i,j}|}$0 for $\mathbf{s}_{i,j}\in\{0,1\}^{|\mathcal{F}_{i,j}|}$1 and $\mathbf{s}_{i,j}\in\{0,1\}^{|\mathcal{F}_{i,j}|}$2 for $\mathbf{s}_{i,j}\in\{0,1\}^{|\mathcal{F}_{i,j}|}$3 versus the production model, and $\mathbf{s}_{i,j}\in\{0,1\}^{|\mathcal{F}_{i,j}|}$4 and $\mathbf{s}_{i,j}\in\{0,1\}^{|\mathcal{F}_{i,j}|}$5 versus a random 4,000-feature baseline (Zhou et al., 26 Mar 2026).

In deliberative role-based selection, LLM-FS-Agent achieved superior or comparable classification performance while reducing downstream training time by an average of $\mathbf{s}_{i,j}\in\{0,1\}^{|\mathcal{F}_{i,j}|}$6; for XGBoost, the improvement was statistically significant with $\mathbf{s}_{i,j}\in\{0,1\}^{|\mathcal{F}_{i,j}|}$7. Table 6 reports average XGBoost training time of $\mathbf{s}_{i,j}\in\{0,1\}^{|\mathcal{F}_{i,j}|}$8 seconds for LLM-Select versus $\mathbf{s}_{i,j}\in\{0,1\}^{|\mathcal{F}_{i,j}|}$9 seconds for LLM-FS-Agent, corresponding to a $$\mathcal{F}_{i,j+1} = \mathcal{F}_{i,j} \odot \mathbf{s}_{i,j}.$$0 speedup (Bal-Ghaoui et al., 7 Oct 2025).

6. Limitations, misconceptions, and open directions

A persistent misconception is that a Model-Data Feature Matcher Agent is necessarily an end-to-end automatic substitute for expert pipeline design. The literature is more cautious. Many systems automate only one layer of the stack—feature generation, expert routing, code mapping, or feature selection—and several rely on curated repositories, external documentation, predefined operator sets, or strong assumptions about the feature space.

Unified tabular agents remain compute-intensive. MAGS explicitly notes high computational overhead from multi-agent coordination and LLM calls, growth of prompt length with larger feature sets and memory, and task-specificity of the optimized features. It also reports a search budget of 30 iterations $$\mathcal{F}_{i,j+1} = \mathcal{F}_{i,j} \odot \mathbf{s}_{i,j}.$$1 6 actions and router fine-tuning on 400 training samples for 5 epochs, which indicates nontrivial orchestration cost rather than zero-cost automation (Gong et al., 21 May 2025).

Privacy-preserving expert routing remains only weakly private in its current form. The hidden-representation ExpertMatcher sends only low-dimensional embeddings instead of raw data, but the work explicitly states that hidden representations can still leak information, provides no differential privacy or attack evaluation, depends on shared pseudorandom seeds to align latent spaces, and assumes the client distribution is reasonably covered by some server expert (Sharma et al., 2019).

Clinical model-repository matching is similarly constrained. The medical pipeline matcher depends heavily on header semantics, uses a fixed cosine threshold $$\mathcal{F}_{i,j+1} = \mathcal{F}_{i,j} \odot \mathbf{s}_{i,j}.$$2, and relies on a curated JSON model database. The paper identifies dependence on feature naming, a small model repository, lack of historical performance feedback, and possible failure when MedGemma misclassifies modality or disease type; matcher-specific quantitative ablations are not provided, since evaluation is system-level (Shimgekar et al., 24 Jul 2025).

Schema-agnostic EMR extraction is feasible but not equivalent to guaranteed correctness. EMR-AGENT shows strong generalization across MIMIC-III, eICU, and the unseen SICdb, yet still requires strong LLM reasoning, extensive use of database documentation and evaluation memos, iterative SQL observation, and expert validation of extracted cohorts and mappings. The paper is explicit that automation does not remove the need for clinical oversight (Lee et al., 1 Oct 2025).

Sequential LLM selectors also retain search-theoretic and systems limitations. MoFA identifies computational time, context-window constraints, and the greediness of one-by-one selection as open issues; LLM-FS-Agent notes that using a single backbone across all roles can propagate shared biases and that no in-loop quantitative tool use is yet integrated. This suggests that a next generation of model-data matchers will likely combine deliberative multi-agent reasoning with stronger tool feedback, hard constraint checking, and tighter coupling to downstream evaluation pipelines rather than relying on text-only planning alone (Zhou et al., 26 Mar 2026, Bal-Ghaoui et al., 7 Oct 2025).