Paper deep dive
Correspondences as Decisions: JevNexus for Decision-Centric Schema Matching
Runze Li, Hanchen Wang, Ying Zhang, Wenjie Zhang
Intelligence
Status: succeeded | Model: Gemma-4-26B-A4B | Prompt: intel-v1 | Confidence: 92%
Last extracted: 10/11/2026, 2:32:49 AM
Summary
The paper introduces JevNexus, a decision-centric schema matching architecture that improves upon generative language model approaches by using typed pairwise decisions combined with schema and instance evidence. It employs an evidence-aware gate to invoke listwise refinement only when necessary, significantly reducing latency while maintaining high accuracy (MRR 0.930, Hits@1 0.909) across six benchmark families.
Entities (14)
Relation Signals (10)
JevNexus → achievesmetric → MRR 0.930
confidence 95% · JevNexus obtains dataset-macro MRR... of 0.930
JevNexus → achievesmetric → Hits@1 0.909
confidence 95% · JevNexus obtains dataset-macro... Hits@1 of... 0.909
Hanchen Wang → affiliatedwith → University of Technology Sydney
confidence 95% · Hanchen Wang University of Technology Sydney
Runze Li → affiliatedwith → University of New South Wales
confidence 95% · Runze Li University of New South Wales
Ying Zhang → affiliatedwith → University of Technology Sydney
confidence 95% · Ying Zhang University of Technology Sydney
Wenjie Zhang → affiliatedwith → University of New South Wales
confidence 95% · Wenjie Zhang University of New South Wales
JevNexus → reduceslatencyby → 7.750x
confidence 95% · reducing mean latency from 123.452 to 15.929 seconds (7.750x)
JevNexus → evaluatedon → GDC-SM
confidence 90% · The evaluation covers 561 cases from six benchmark families... GDC-SM
JevNexus → evaluatedon →
Cypher Suggestions (0)
No Cypher suggestions yet.
Abstract
Abstract:Schema matching increasingly uses generative language models to rerank retrieved column candidates, although the underlying task is a bounded correspondence decision. We present JevNexus, which combines typed pairwise decisions with schema/instance evidence and invokes listwise refinement only when the evidence disagrees and the fused margin is small. The evaluation covers 561 cases from six benchmark families. JevNexus obtains dataset-macro MRR and Hits@1 of 0.930 and 0.909, compared with 0.926 and 0.903 for Magneto, while reducing mean latency from 123.452 to 15.929 seconds (7.750). Paired analysis finds no statistically significant difference in either MRR or Hits@1. The gate invokes listwise refinement for only 5.665% of source columns and avoids the degradation caused by unconditional refinement. Code and experimental artifacts are available at this https URL.
Tags
Links
- Source: https://arxiv.org/abs/2610.09487v1
- Canonical: https://arxiv.org/abs/2610.09487v1
Trouble viewing inline? Open PDF directly →
Full Text
65,553 characters extracted from source content.
Expand or collapse full text
Correspondences as Decisions: JevNexus for Decision-Centric Schema Matching Runze Li University of New South Wales Sydney, Australia runze.li1@unsw.edu.au Hanchen Wang University of Technology Sydney Sydney, Australia hanchen.wang@uts.edu.au Ying Zhang University of Technology Sydney Sydney, Australia ying.zhang@uts.edu.au Wenjie Zhang University of New South Wales Sydney, Australia wenjie.zhang@unsw.edu.au ABSTRACT Schema matching increasingly uses generative language models to rerank retrieved column candidates, although the underlying task is a bounded correspondence decision. We present JevNexus, which combines typed pairwise decisions with schema/instance evidence and invokes listwise refinement only when the evidence disagrees and the fused margin is small. The evaluation covers 561 cases from six benchmark families. JevNexus obtains dataset- macro MRR and Hits@1 of 0.930 and 0.909, compared with 0.926 and 0.903 for Magneto, while reducing mean latency from 123.452 to 15.929 seconds (7.750×). Paired analysis finds no statistically significant difference in either MRR or Hits@1. The gate invokes listwise refinement for only 5.665% of source columns and avoids the degradation caused by unconditional refinement. Code and experimental artifacts are available at https://github.com/RazeenLI/ JevNexus. 1 INTRODUCTION The continuing growth of structured data in scientific repositories, enterprise platforms, and public data ecosystems creates substantial opportunities for cross-source analytics, scientific discovery, and data-driven decision making [1,7,11,14,15]. Realizing these oppor- tunities, however, often requires independently developed datasets to be integrated before they can be queried or analyzed together. For example, cancer consortia harmonize genomic, proteomic, and clinical data to support pan-cancer studies [15]; enterprise discov- ery systems expose relationships among tables distributed across organizational data repositories [11]; and Web- and data-lake-scale systems identify tables that can be joined or unioned for down- stream analysis [1,7]. These workflows depend on schema match- ing, the task of identifying attributes in heterogeneous schemas that represent the same underlying concepts [8, 18, 25]. Formally, given a source schema푆and a target schema푇, schema matching seeks a set of valid correspondences푀 ⊆ 푆×푇and, in ranking-oriented settings, places the valid targets for each source attribute early in the resulting order. When푇is large, a localiza- tion stage can first restrict the search to a bounded candidate set 퐶(푠) ⊆ 푇for each푠 ∈ 푆. Once these candidates have been identified, the fundamental operation is a collection of constrained questions: for each푡 ∈ 퐶(푠), do푠and푡represent the same underlying schema attribute? This operation is naturally expressed as a bounded cor- respondence decision, with each pair assessed on its own merits COLUMN NAMEfigo_grade figo 1 SAMPLE VALUE figo 2figo 3 Source Table \(S \) ... more source columns Target Table \(T \) ... more target columns COLUMN NAMEtumor_grade grade 1 SAMPLE VALUE grade 2grade 3 COLUMN NAMEfigo_stage stage 1 SAMPLE VALUE stage 2stage 3 SOURCE COLUMNfigo_grade ❶ tumor_grade ... for each source column Candidates ❷ who_grade ❸ figo_stage❹ ... Figure 1: A schema-matching example with conflicting evi- dence. and multiple or no candidates allowed to match, rather than as an open-ended generation task. Recent LLM-based matchers instead approximate the operation by asking a generative model to emit candidate names together with textual or JSON-encoded similarity scores [17]. Although such generation can produce effective rank- ings, it represents the underlying judgments indirectly through autoregressive decoding, output parsing, and generated numeric values. We therefore investigate a decision-centric formulation in which localized candidates are primarily evaluated through sepa- rately scored, non-exclusive correspondence decisions. Figure 1 illustrates the distinction using a source column named figo_gradeand two plausible target columns,tumor_gradeand figo_stage, following the type of ambiguity found in the GDC schema-matching setting [17]. The tokenfigoprovides evidence forfigo_stage, whereas the attribute type expressed bygrade provides evidence fortumor_grade, and the sampled values may not fully eliminate this ambiguity. The correspondence therefore cannot be resolved reliably from a single surface-similarity signal. Evaluating the two pairs separately makes the plausibility of each correspondence explicit, while comparing the candidates remains useful when their evidence is closely balanced. The example sep- arates three operations that a matching system must coordinate: judging individual correspondences, integrating complementary evidence, and comparing unresolved candidates. Classical schema matchers combine lexical, structural, and instance- level signals, reflecting the long-standing observation that no single arXiv:2610.09487v1 [cs.DB] 7 Oct 2026 Runze Li, Hanchen Wang, Ying Zhang, and Wenjie Zhang form of evidence is reliable across all schema pairs [8,18,19,25]. Although these methods preserve complementary evidence, manu- ally designed similarities often struggle with correspondences that require broader semantic understanding. Pretrained language mod- els extend this semantic coverage by learning representations or matching functions for schema elements [9,34,38]. More recently, retrieve–rerank systems such as Magneto use an efficient semantic model to localize candidates and a generative LLM to produce their final scores or ordering [17]. This progression improves semantic matching and makes LLM reasoning feasible over wide schemas, but it leaves three gaps. First, the final correspondence judgment is expressed indirectly through generated output rather than through an explicit matching decision. Second, the move toward a single semantic reranker does not explicitly preserve the complementary roles of semantic, schema-level, and instance-level evidence. Third, uniformly reranking every localized candidate set does not distin- guish cases that require joint comparison from those for which the available evidence already supports a clear ordering. These gaps lead directly to three challenges for decision-centric schema matching. The first is to represent the plausibility of every localized pair directly, without autoregressive score generation or normalization across candidates. The second is to integrate seman- tic judgments with complementary schema and instance evidence when different signals provide different views of a candidate, as in Figure 1. The third is to retain joint candidate comparison for genuine residual ambiguities without allowing a listwise model to reconsider every otherwise confident ranking. Meeting these challenges requires pairwise judgment, evidence integration, and cross-candidate comparison to remain distinct but coordinated op- erations. We introduce JevNexus, a progressive schema-matching archi- tecture built around this separation. The pipeline begins by localiz- ing a bounded candidate set for each source attribute, separating candidate recall from final correspondence judgment. A typed de- cision operator assigns every localized pair a separately scored, non-exclusive probability that the two columns represent the same underlying attribute. These semantic decisions are combined with schema- and instance-level scores so that the default ranking re- flects complementary evidence rather than a single model output. When the evidence sources prefer different candidates and the high- est fused scores remain close, an evidence-aware gate sends only the leading alternatives to a compact listwise reranker. Otherwise, the fused ranking bypasses listwise refinement and remains un- changed. In this way, JevNexus uses pairwise decisions to establish correspondence plausibility, heterogeneous evidence to form the default order, and joint reasoning only to resolve remaining local ambiguities. We evaluate JevNexus on 561 schema-matching cases from GDC- SM and five Valentine benchmark families, covering 13,522 source columns and 8,944 ground-truth correspondences. Across the six benchmark families, JevNexus achieves the highest dataset-macro MRR and Hits@1 among the evaluated methods and is statistically comparable to the strongest generative method over all 561 cases. It obtains the highest or tied-highest MRR on GDC, Magellan, and WikiData and remains competitive on the other benchmark fam- ilies. More importantly, JevNexus reduces mean online runtime from 123.452 to 15.929 seconds per case, yielding a 7.750×speedup over generative reranking. Ablations show that complementary schema and instance evidence provides the largest effectiveness gain, while unconditional listwise refinement is less effective than the proposed selective strategy. The evidence-aware gate produces the strongest aggregate JevNexus variant while activating for only 5.665% of source columns, showing that most localized decisions do not require joint reranking. We make the following contributions: •Decision-centric formulation. We formulate localized schema matching as explicit, separately scored, and non- exclusive correspondence decisions rather than autoregres- sive similarity-score generation. •Evidence-aware progressive architecture. We develop JevNexus, which combines typed semantic decisions with complementary schema and instance evidence and invokes listwise comparison only under simultaneous evidence dis- agreement and fused-ranking uncertainty. •Controlled full-benchmark evaluation. We compare JevNexus with traditional, pretrained-neural, and genera- tive schema matchers across 561 cases and analyze ranking effectiveness, efficiency, component contributions, rout- ing behavior, dataset heterogeneity, and remaining failure modes. 2 PROBLEM FORMULATION AND MOTIVATION 2.1 Schema Matching and Ranked Output Let a source table퐴have schema푆= 푠 1 , . . .,푠 푚 and a target table퐵have schema푇= 푡 1 , . . .,푡 푛 . For a column푐 ∈ 푆 ∪푇, let 푁(푐)denote its name and푉(푐)the available sample of observed non-null values, which may be empty. We write the observable column evidence as 휙(푐)= 푁(푐),푉(푐) . A schema correspondence is a pair(푠,푡) ∈ 푆 ×푇whose columns represent the same underlying attribute according to the benchmark semantics [17,25]. The ground-truth correspondence relation is denoted by퐺 ⊆ 푆×푇, and the valid targets for a source column푠 are 퐺(푠)=푡 ∈ 푇 | (푠,푡) ∈ 퐺. This relation does not impose a one-to-one constraint: depending on the matching setting,퐺(푠)may contain zero, one, or multiple target columns. Following ranking-oriented schema matching [13,17], a matcher produces for every source column 푠 ∈ 푆 a complete ordering 휋 푠 =⟨푡 (1) ,푡 (2) , . . .,푡 (푛) ⟩ of the target columns. The ranking may later be thresholded or transformed into a discrete mapping, but our primary objective is to place the members of퐺(푠)as early as possible in휋 푠 . This ranking formulation separates the quality of correspondence evidence from any application-specific threshold used to accept or reject matches. 2.2 Candidate-Localized Matching When the target schema is wide, applying an expensive semantic operator to every pair in푆×푇is unnecessary. A localization operator Correspondences as Decisions: JevNexus for Decision-Centric Schema Matching Generative LLM autoregressive text generation <s>["column"...]</s> Retrieved Candidate Set tumor_gradefigo_stagewho_grade Source Column figo_grade [ "column": "figo_stage", "score": 0.85, "column": "tumor_grade", "score": 0.62, ... ] Parse + Validate Candidate Scores / Ranking Correspondence judgment via generated scores figo_grade + tumor_grade figo_grade + figo_stage figo_grade + who_grade ... Same underlying attribute? Same underlying attribute? Same underlying attribute? Same underlying attribute? ... \( P(match | s, t_1) \) Matching Evidence Candidate Ranking \( P(match | s, t_2) \) \( P(match | s, t_3) \) separately scored pairwise probabilities Direct correspondence decisions Retrieved Candidate Set tumor_gradefigo_stagewho_grade Source Column figo_grade Generative Reranking Generated Output Decision-Centric Matching Generate → Parse → Score Decide → Probability Figure 2: Generative and decision-centric inference over the same localized candidates. Ltherefore returns at most푘candidate targets for each source column: 퐶 푘 (푠)=L(푠,푇), |퐶 푘 (푠)| ≤ 푘. Localization and correspondence judgment nevertheless solve dif- ferent problems:Lidentifies a high-recall neighborhood, whereas the subsequent matcher must discriminate among the candidates in that neighborhood. If퐺(푠)∩ 퐶 푘 (푠)= ∅, no downstream deci- sion or reranking operator restricted to퐶 푘 (푠)can recover a valid target at the head of the ranking. Candidate localization therefore determines a recall ceiling, while the quality of the ordering within 퐶 푘 (푠)determines how effectively the available correspondences are surfaced. 2.3 Generated versus Decision-Based Evidence Once퐶 푘 (푠) has been obtained, generative reranking and decision- centric matching receive the same source column, candidate columns, and observable column evidence, and both ultimately support an or- dering over the candidates. They differ in how the correspondence evidence underlying that ordering is represented and obtained, as illustrated in Figure 2. In generative reranking, a model jointly inspects the source and candidate set and autoregressively emits candidate identifiers together with textual or JSON-encoded scores. The matching system then parses and validates the generated ar- tifact before converting it into candidate scores and a ranking. Although this procedure can produce effective rankings, its corre- spondence judgments are represented indirectly through generated tokens. A decision-centric formulation instead poses the bounded match- ing question directly for each localized pair: do푠and푡represent the same underlying attribute? The resulting answer is expressed as bounded pairwise evidence and can be consumed without gen- erating or parsing a textual score. These direct judgments provide semantic evidence for constructing the candidate ranking and, when useful, for combination with other matching signals. 2.4 Decision-Centric Matching Objective For each localized pair(푠,푡)with푡 ∈ 퐶 푘 (푠), define the correspon- dence label 푦 푠,푡 = 1 (푠,푡) ∈ 퐺 . A decision-centric matcher seeks bounded semantic evidence 푞(푠,푡) ∈ [0, 1] whose magnitude reflects the plausibility that푦 푠,푡 =1 given휙(푠) and휙(푡). Each푞(푠,푡)is obtained as a separately scored correspon- dence judgment rather than as one component of a multiclass dis- tribution over퐶 푘 (푠). Consequently, the values푞(푠,푡):푡 ∈ 퐶 푘 (푠) are not constrained to sum to one: multiple candidates may receive strong correspondence evidence, and all candidates may receive weak evidence. The values푞(푠,푡)represent semantic decision evidence rather than prescribing the complete matching algorithm or its final score. A matcher may combine them with complementary schema- or instance-level evidence before producing휋 푠 . The resulting problem is to rank valid targets early while bounding expensive inference to the localized candidate sets. Section 3 presents JevNexus, which realizes this formulation through typed correspondence decisions, complementary evidence integration, and selectively invoked cross- candidate refinement. 3 JEVNEXUS 3.1 Architecture Overview JevNexus is a progressive schema-matching architecture that sepa- rates candidate localization, direct correspondence judgment, com- plementary evidence integration, and residual cross-candidate com- parison. For each source column푠, the pipeline first localizes a bounded candidate neighborhood퐶 푘 (푠)and then assigns every localized pair two functionally distinct signals: typed semantic decision evidence푑(푠,푡)and complementary schema/instance ev- idenceℎ(푠,푡). The two signals are integrated into a default score 푓(푠,푡), and a routing criterion determines whether the fused head Runze Li, Hanchen Wang, Ying Zhang, and Wenjie Zhang figo_grade figo 1 figo 3 tumor_grade grade 1 grade 3 figo_stage stage 1 stage 3 MPNet eₛ eₜ COS Top-K candidates ❶ tumor_grade ❷ who_grade ❸ figo_stage ❹ ... (s, t1) (s, t2) (s, t3) ... Jev P name schema values ∑ figo 2 ... grade 2 ... stage 2 ... who_grade who 1 who 3 who 2 ... MPNet ❶ ❷ ❸ ❹ ❶ ❷ ❸ ❹ 0.86 0.56 0.22 0.10 0.68 0.89 0.37 0.15 ∑ ❶ ❷ ❸ ❹ 0.75 0.76 0.31 0.13 Top-1 t d ≠ t h Top-2 Δ < .02 AND Jina Completed rₛ independent binary decision for each (s, t) schema and instance matchers dₛ hₛ fₛ TRUE σₛ= (2, 1, 3) MUX FALSE πₛ Candidate localization A Complementary evidence B Fusion and conditional refinement C target schema T source column S Decide Match ❶ tumor_grade ❷ who_grade ❸ figo_stage ❹ ... Figure 3: Overview of the JevNexus matching pipeline. is sufficiently ambiguous to justify local listwise comparison. The resulting localized order is finally combined with the untouched target-schema tail to produce the complete ranking 휋 푠 . Figure 3 summarizes this information flow. Localization is re- sponsible for candidate recall, typed decisions express pairwise semantic correspondence, complementary evidence captures lexi- cal and value regularities, and listwise refinement is restricted to residual local ambiguities. This decomposition makes the contri- bution of each evidence source and the effect of selective routing directly observable. 3.2 Candidate Localization Candidate localization instantiates the operatorLintroduced in Section 2. Each column is represented using its normalized name and a deterministic sample of observed non-null values. The repre- sentation is serialized into a compact text sequence and encoded in a shared semantic space [27,32]. For a source column푠and target column 푡 , the retrieval score is cosine similarity, 푟(푠,푡)= cos 푒(푠),푒(푡) , where 푒(·) denotes the resulting column embedding. The localizer supplements semantic retrieval with exact normalized- name matches, removes pairs below the configured similarity thresh- old, orders the surviving targets by retrieval evidence, and retains at most푘candidates. Exact-name evidence prevents simple but im- portant correspondences from being lost because of representation noise. All ordering operations use deterministic tie breaking. The retrieval score푟(푠,푡)determines membership and initial order in퐶 푘 (푠)but is not treated as final correspondence evidence by the subsequent decision and fusion stages. The localizer also retains the target columns outside퐶 푘 (푠)in target-schema order so that the pipeline can later construct a complete ranking. 3.3 Typed Correspondence Decisions The typed decision operator realizes the direct correspondence se- mantics established in Section 2. Rather than requesting a serialized list of candidate scores, it presents each localized source–target pair as a bounded binary question whose positive outcome denotes that the two columns represent the same underlying attribute. Shared source state \( x_s \) Column:figo_grade Evidence:sampled values \(V(s)\) Candidate-specific question \( Q_s,t_1 \) Column:tumor_grade Evidence:sampled values \(V(t_1)\) Same underlying attribute? Bounded output \( d(s, t_1) ∈ [0, 1] \) Candidate-specific question \( Q_s,t_2 \) Column:figo_stage Evidence:sampled values \(V(t_2)\) Same underlying attribute? Bounded output \( d(s, t_2) ∈ [0, 1] \) reuse the same source state Figure 4: Typed correspondence decisions for two localized candidates. 3.3.1State–Question Construction. For a source column푠, JevNexus constructs a source state 푥 푠 = serialize 푁(푠),푉(푠) . For every푡 ∈ 퐶 푘 (푠), it constructs a candidate-specific typed ques- tion 푄 푠,푡 = question 푁(푡),푉(푡) that asks whether the candidate represents the same underlying attribute as the source. The positive criterion permits different surface forms when the real-world attribute is equivalent, whereas the negative criterion covers distinct attributes despite lexical or value-level similarity. Figure 4 instantiates this interface with the running example from Figure 1. The source profile is shared, but each candidate sup- plies a separate question and receives its own bounded match value. The state and questions contain only observable column evidence: they exclude retrieval scores and ranks, dataset identifiers, bench- mark metadata, and ground truth. Candidate identifiers serve only to associate returned values with questions and carry no semantic or ranking information. All questions for a source column can be submitted in one batched decision request while retaining candidate-specific con- texts. The source state is therefore encoded once, and each question Correspondences as Decisions: JevNexus for Decision-Centric Schema Matching exposes only the evidence of its own candidate rather than the com- plete candidate list. This organization preserves separately scored pairwise semantics without issuing one external request per pair. 3.3.2 Bounded Decision Inference. For every localized pair, the decision operator returns 푑(푠,푡)= 푃 휃 푦 푠,푡 = 1 | 푥 푠 ,푄 푠,푡 ∈ [0, 1]. The value is obtained directly from the typed binary decision in- terface rather than generated as a numeric string, candidate list, or explanation. Consequently, the matching pipeline does not need autoregressive answer decoding or score extraction from free-form text. Candidates are evaluated as separate questions and are not nor- malized with a softmax over퐶 푘 (푠). The values푑(푠,푡):푡 ∈ 퐶 푘 (푠) are therefore not constrained to sum to one: multiple candidates may receive strong evidence, and all candidates may receive weak evidence. The output contract requires exactly one finite value in[0,1] for every submitted candidate identifier. Missing identifiers, un- expected identifiers, duplicate fields, and out-of-range values are treated as invalid outputs rather than silently converted into corre- spondence scores. 3.4 Complementary Schema and Instance Evidence Direct semantic decisions do not exhaust the evidence available for schema matching. Column names can exhibit useful lexical regularities, and observed values can reveal correspondences that remain ambiguous from language-level semantics alone. JevNexus therefore computes a second scoreℎ(푠,푡)with a non-generative matching operator that combines schema- and instance-level sig- nals, following the classical principle of composing heterogeneous match functions [8, 10]. 3.4.1Schema-Level Evidence. For each source–target pair, the schema- level branch compares normalized column names with character- trigram and token-based similarities. The two measures are com- plementary: trigram overlap captures local spelling variation, while token similarity is more robust to separators, reordered fragments, and multiword names. Their maximum forms the name score ℎ name (푠,푡)= max ℎ tri (푠,푡),ℎ tok (푠,푡) . Because the evaluated inputs are flat relational tables, this branch uses column-level name evidence rather than introducing artificial hierarchy or neighborhood structure. 3.4.2Instance-Level Evidence. The instance-level branch operates on deterministically sampled table rows. For each table pair, it builds a shared TF–IDF corpus from the sampled values of all source and target columns and compares the value documents associated with each column pair. The resulting scoreℎ inst (푠,푡) ∈ [0,1]captures value-level overlap while down-weighting values that occur broadly across columns. If a table contains more rows than the configured sample budget, sampling uses a fixed seed so that the evidence is reproducible. 3.4.3 Evidence Aggregation and Selection. The enabled name and instance signals are combined by a fixed weighted average, e ℎ(푠,푡)= 푤 name ℎ name (푠,푡)+푤 inst ℎ inst (푠,푡) 푤 name +푤 inst . The canonical configuration assigns equal internal weight to the two branches. The operator computes e ℎ over the table pair and then applies bidi- rectional multiple selection. A pair is retained only if its score satis- fies the configured Top-푛, relative-to-best, and absolute-threshold conditions from both the source and target directions. The selected score defines ℎ(푠,푡)= e ℎ(푠,푡), whereas a localized pair not selected by the complementary op- erator receivesℎ(푠,푡)=0. Here, zero indicates that the selection procedure retained no positive complementary evidence for the pair. The complementary operator supplies scores only and never adds candidates to or removes candidates from퐶 푘 (푠). 3.5 Evidence Integration Decision evidence captures semantic correspondence, whereas the complementary score captures lexical and value regularities; neither signal is uniformly sufficient across heterogeneous schema pairs. JevNexus combines them into the integrated score 푓(푠,푡)= 훼 푑(푠,푡)+(1− 훼)ℎ(푠,푡),0≤ 훼 ≤ 1.(1) The coefficient훼controls the relative contribution of the two signals. Becauseℎ(푠,푡)is a matching score rather than a probability, the fused value푓(푠,푡)is used for ranking rather than interpreted as a match probability. Sorting the localized candidates by푓(푠,푡)produces the default fused order 휋 푓 푠 = sort 푡∈퐶 푘 (푠) 푓(푠,푡). This order is returned directly unless the routing criterion in Sec- tion 3.6 invokes listwise refinement. 3.6 Evidence-Aware Selective Refinement Pairwise and complementary evidence resolve most localized rank- ings, but a subset retains both conflicting signals and closely bal- anced leading candidates. JevNexus reserves cross-candidate rea- soning for this residual ambiguity instead of applying a listwise operator to every source column. 3.6.1 Routing Criterion. Let 푡 푑 (푠)= argmax 푡∈퐶 푘 (푠) 푑(푠,푡)and 푡 ℎ (푠)= argmax 푡∈퐶 푘 (푠) ℎ(푠,푡) denote the candidates preferred by the two evidence sources. All argmax and sorting operations use retrieval order as a deterministic tie breaker. If ℎ(푠,푡)= 0 for every localized candidate, the comple- mentary operator has expressed no positive preference and푡 ℎ (푠)is treated as unavailable. Evidence disagreement is defined as 퐷(푠)= 1 [ 푡 ℎ (푠) is available ∧ 푡 푑 (푠)≠ 푡 ℎ (푠) ] . Disagreement alone does not trigger refinement because fusion may still establish a decisive order. Runze Li, Hanchen Wang, Ying Zhang, and Wenjie Zhang Let푓 (1) (푠) ≥ 푓 (2) (푠)be the two largest fused scores and define their margin Δ 푓 (푠)= 푓 (1) (푠)− 푓 (2) (푠). A small margin indicates that the leading fused alternatives remain difficult to distinguish. The routing decision is therefore 푔(푠)= 1 퐷(푠)= 1 ∧Δ 푓 (푠)< 휏 ,(2) where휏is a fixed uncertainty threshold. For an empty or singleton candidate set,Δ 푓 (푠) is undefined and 푔(푠) is set to zero. 3.6.2Local Listwise Refinement. When푔(푠)=1, JevNexus selects the푚 푠 highest-scoring fused candidates, 퐴 푚 푠 (푠)= TopM 푡∈퐶 푘 (푠) 푓(푠,푡), 푚 푠 = min 푚,|퐶 푘 (푠)| . The local refiner receives the source profile as its query and the pro- files of퐴 푚 푠 (푠)as candidate documents. Each candidate document also includes its decision, complementary, and fused evidence val- ues so that the refiner can compare the alternatives in the context of the signals that produced the default order. The refiner returns a permutation휌 푠 of퐴 푚 푠 (푠)rather than a replacement score scale. To preserve the scale and distribution of the integrated evidence, let 푢 1 ≥ 푢 2 ≥·≥ 푢 푚 푠 be the fused score slots occupied by the selected candidates. If휌 푠 places candidate푡at refined position푟, the final localized score of 푡 is set to푢 푟 . Candidates outside 퐴 푚 푠 (푠) retain their fused scores. When푔(푠)=0, the fused order bypasses the refiner unchanged. When the gate opens but the returned permutation agrees with the fused order, the ranking also remains unchanged. Thus, the refinement stage can modify only the local head of a routed source and cannot change candidate membership or the overall fused score distribution. 3.7 Complete Ranking After optional refinement, the candidates in퐶 푘 (푠)are sorted by their resulting scores to form the localized head. Target columns outside 퐶 푘 (푠)are appended in their retained target-schema order, yielding the complete ranking휋 푠 over푇. This construction avoids inventing decision or complementary scores for pairs that the corresponding operators did not evaluate. Score ties in the localized head are resolved by retrieval order, while the untouched tail preserves target-schema order. An empty candidate set therefore returns the untouched target order, and a singleton set bypasses listwise refinement by construction. A ground-truth target outside퐶 푘 (푠)remains in the tail and cannot be recovered by downstream decision, integration, or refinement. 3.8 End-to-End Procedure Given푆,푇, the candidate bound푘, fusion coefficient훼, refinement width푚, and margin threshold휏, JevNexus executes the following procedure for every source column 푠 ∈ 푆 : (1)Localize at most푘candidates퐶 푘 (푠)and retain the remain- ing targets as an ordered tail. (2) Submit one typed correspondence question for every푡 ∈ 퐶 푘 (푠) and collect the decision evidence 푑(푠,푡). (3)Look up the complementary schema/instance evidence ℎ(푠,푡)and compute the integrated score푓(푠,푡)using Equa- tion 1. (4)Compute the evidence-disagreement indicator and fused Top-2 margin, and evaluate the routing rule in Equation 2. (5)If the gate opens, obtain a local permutation of the fused Top-푚and reassign their fused score slots; otherwise, retain the fused order. (6) Sort the localized candidates, append the untouched tail, and return the complete ranking 휋 푠 . The complementary score matrix is computed once per table pair, whereas localization, decision questions, routing, and optional refinement are organized per source column. All neural and non- neural operators remain fixed during online matching, and the procedure performs no online parameter updates. 3.8.1Running Example. Consider again the schematicfigo_grade example from Figure 1. Localization first placestumor_gradeand figo_stagein퐶 푘 (푠)together with any other retrieved candidates. The typed decision operator asks a separate correspondence ques- tion for each candidate, as shown in Figure 4. In this schematic example, suppose that attribute semantics yield 푑(푠, tumor_grade)> 푑(푠, figo_stage), while the shared tokenfigomakes the complementary evidence prefer the opposite candidate, ℎ(푠, tumor_grade)< ℎ(푠, figo_stage). After integration, suppose that the two candidates occupy the lead- ing fused positions and satisfyΔ 푓 (푠)< 휏. Because their evidence- specific Top-1 candidates disagree,퐷(푠)=1 and the conjunction in Equation 2 opens the refinement gate. In the illustrated path, the local refiner compares only the fused Top-푚alternatives and places tumor_gradeahead offigo_stageby permuting their fused score slots. The remaining localized candidates keep their fused scores, and all targets outside퐶 푘 (푠)are appended in their retained target- schema order. If either evidence agreement or a decisive fused margin had been present instead, the same example would have bypassed refinement and returned the fused order directly. 3.9 Inference Cost Localization bounds expensive semantic judgment to at most푘can- didates per source rather than applying it to all|푇|target columns. Across a schema pair, the number of typed questions is at most ∑︁ 푠∈푆 |퐶 푘 (푠)| ≤ |푆|푘. Because the questions for a source are batched into one decision request, the normal request count is proportional to|푆|rather than to|푆|푘. The decision interface returns bounded values directly and does not incur autoregressive generation of candidate identifiers, JSON syntax, or numeric score strings. The complementary operator computes its table-pair score ma- trix once and exposes the selected entries needed for localized fusion. Its scores are obtained without generative inference, and lookup during fusion is constant time per localized pair. Correspondences as Decisions: JevNexus for Decision-Centric Schema Matching Table 1: Dataset statistics. DatasetCases Source cols. Source width Target width GT pairs GT/case ChEMBL1803,06012–2312–232,05211.4 GDC1056916–17973625925.9 Magellan7414–94–9415.9 OpenData1806,87626–5126–514,57225.4 TPC-DI1802,91611–2212–221,98011.0 WikiData46013–2013–204010.0 Total56113,522–8,94415.9 Listwise refinement adds exactly ∑︁ 푠∈푆 푔(푠) model requests, each over at most푚candidates. An unconditional design would issue one such request for every eligible source col- umn, whereas JevNexus pays this additional cost only when evi- dence disagreement coincides with a small fused margin. The canon- ical values of푘,훼,푚, and휏, together with the concrete operator implementations, are reported in Section 4.1. 4 EXPERIMENTAL EVALUATION We evaluate JevNexus along four dimensions: end-to-end effec- tiveness, inference efficiency, component contributions, and the behavior of selective refinement. We additionally study hyperpa- rameter sensitivity and decompose residual errors into candidate- localization and downstream-ranking failures. 4.1 Experimental Setup Datasets. We use GDC-SM [17] and five datasets distributed with Valentine [13]. Table 1 summarizes the benchmark. It contains 561 cases, 13,522 source columns, and 8,944 ground-truth pairs. GDC is particularly challenging because each case maps into a common target schema with 736 columns. Compared methods. We compare against COMA [8], COMA++ [10], a distribution-based matcher, Similarity Flooding [19], ISResMat [9], Unicorn [34], and Magneto [17]. We run the official Magneto im- plementation and instantiate its replaceable LLM reranker with Qwen3.5-9B [24]. Magneto and JevNexus receive the same local- ized candidate sets and use the same candidate limit, isolating the quality and cost of downstream matching from candidate recall. The ablations retain only typed decision evidence, add complementary schema/instance evidence, apply listwise refinement uncondition- ally, or use the complete selective pipeline. Metrics and protocol. MRR is our primary ranking metric, Hits@1 measures whether a valid target is ranked first, and Recall@GT follows the Valentine and Magneto protocol: all pairs in a case are pooled by score, the top|퐺|are retained, and recall is computed against the ground truth. Recall@GT therefore also depends on score comparability across source columns. We report two aggregate views. Case-macro averages all cases equally; therefore ChEMBL, OpenData, and TPC-DI, with 180 cases each, receive more weight. Dataset-macro first averages cases within each dataset and then gives each dataset equal weight. JevNexus and all ablations complete all 561 cases. Magneto times out on one 0.000.250.500.751.00 Dataset-macro MRR COMA COMA++ Distribution SF ISResMat Unicorn Magneto JevNexus 0.250.500.751.00 Dataset-macro Hits@1 Figure 5: Overall matching effectiveness. OpenData case, and Distribution does not complete the ten GDC cases. Main effectiveness results cover all 561 expected cases and assign zero effectiveness to each execution failure; completed-case counts are reported explicitly. Configuration. The candidate limit is푘=20. The complemen- tary matcher uses푛=20, a relative-to-best tolerance of 0.15, an absolute threshold of 0, equal schema and instance weights, and an instance sample size of 500. Decision inference uses Open-Jev- 9B with maximum length 4096 and batch size 20 [2]. Fusion uses 훼=0.400, and refinement uses푚=3,휏=0.020, and Jina-Reranker- v3.5 [21]. The same fusion and routing configuration is used for every dataset without per-dataset fitting. Magneto uses Qwen3.5- 9B with temperature zero, at most 1,024 generated tokens, and disabled thinking. We retain its original benchmark serialization de- faults but use the zero-shot retriever rather than the GDC-specific fine-tuned variant. Neural methods run on the same RTX PRO 6000 host; timing excludes model startup and includes representation, retrieval, matching or reranking, and final ranking. All system-level latency summaries include the elapsed time of a failed run. All met- rics are computed from saved rankings with the same evaluator. Exact model revisions, prompts, software versions, and commands are provided in the artifact. Code and experimental artifacts are available at https://github.com/RazeenLI/JevNexus. 4.2 End-to-End Effectiveness Figure 5 summarizes performance across all six benchmark fami- lies. JevNexus obtains the highest dataset-macro MRR and Hits@1, followed by Magneto, while Magneto obtains the highest Re- call@GT. The remaining methods trail both on aggregate ranking quality, although their relative positions vary by dataset. Table 2 provides the per-dataset results; each dataset block re- ports MRR, Hits@1, and Recall@GT, with the best values in bold. JevNexus achieves the highest MRR and Hits@1 on GDC, ties for the best ranking results on Magellan and WikiData, and is close to the leading method on OpenData and TPC-DI. ChEMBL is the only family on which its MRR trails Magneto by more than two points, although it remains stronger than every other baseline. No method dominates every setting: Unicorn leads MRR and Hits@1 on OpenData, while Magneto leads on ChEMBL and narrowly on TPC-DI. Recall@GT gives a different view, with Magneto strongest on four datasets and JevNexus strongest on WikiData. Comparison with generative reranking. Because Magneto is the strongest generative method and the closest overall competitor, we Runze Li, Hanchen Wang, Ying Zhang, and Wenjie Zhang Table 2: Per-dataset matching effectiveness. ChEMBLGDCMagellanOpenDataTPC-DIWikiData MethodMRRH@1R@GTMRRH@1R@GTMRRH@1R@GTMRRH@1R@GTMRRH@1R@GTMRRH@1R@GT COMA0.7900.7150.5990.5210.4810.314 1.000 1.000 1.0000.8140.7810.6410.9030.8810.8760.7840.7480.665 COMA++0.8960.8470.7470.5610.5070.366 1.000 1.000 1.0000.9110.8910.5140.9720.9630.870 0.946 0.9460.873 Distribution0.5490.4380.4510.0000.0000.0000.6860.5610.5610.5630.4820.2820.7900.7220.5840.8080.7730.644 Similarity Flooding0.7930.6590.4470.5440.4670.280 1.000 1.000 1.0000.6500.5540.4200.8770.8250.6650.7660.7100.646 ISResMat0.8380.7820.7330.4480.3790.258 1.000 1.000 1.0000.6860.5870.5140.8680.7940.767 0.946 0.9460.873 Unicorn0.8140.7360.6840.5780.4940.3190.9630.9440.920 0.971 0.955 0.7930.8720.8170.6510.9000.8770.817 Magneto0.965 0.944 0.8830.7380.657 0.461 1.000 1.000 1.0000.9320.9120.753 0.983 0.968 0.9450.9400.9330.890 JevNexus0.9070.8780.783 0.794 0.7250.456 1.000 1.000 1.0000.9520.9390.6480.9800.9640.865 0.946 0.946 0.902 10 0 10 1 10 2 Mean latency (s, log) 0.7 0.8 0.9 Dataset-macro MRR 7.75 × Effectiveness vs. runtime 10 −1 10 0 10 1 10 2 Latency (s, log) 0.06--0.16 0.15--1.11 1.13--5.36 0.48--4.36 15.2--33.6 0.42--3.00 96.6--284 14.7--30.0 Median--P95 latency COMA COMA++ Distribution SF ISResMat Unicorn Magneto JevNexus Figure 6: Matching effectiveness and inference efficiency. additionally bootstrap paired case differences within each dataset before averaging the six dataset means. The analysis includes all 561 cases and assigns zero effectiveness to the timed-out Magneto case. The MRR difference (JevNexus−Magneto) is 0.004 with a 95% confidence interval of[−0.005,0.012]; the Hits@1 difference is 0.006 with interval[−0.007,0.020]. Both intervals contain zero. The Recall@GT difference is−0.046 with interval[−0.061,−0.032] (푝<0.001). Thus, their dataset-macro MRR and Hits@1 are sta- tistically comparable, whereas Magneto retains higher pooled correspondence recall. 4.3 Efficiency The left panel of Figure 6 places JevNexus in the effectiveness– runtime landscape of methods evaluated across all six benchmark families, while the right panel compares median and tail latency for all methods. Traditional matchers are faster but substantially less accurate, while JevNexus combines the highest dataset-macro MRR with much lower latency than Magneto and ISResMat. Includ- ing the timed-out run, JevNexus requires 15.929 seconds per case compared with 123.452 seconds for Magneto, a 7.750× speedup. Median and 95th-percentile latency are 14.694 and 30.029 seconds for JevNexus, versus 96.639 and 284.221 seconds for Magneto. The decision pipeline processes more input context but emits no autoregressive score tokens, whereas the completed Magneto runs emit 5.31 million recorded output tokens. The bounded output interface removes autoregressive score generation, and the latency advantage persists across both typical and tail cases. 4.4 Component Ablation Table 3 uses the same three metrics as the main table. Adding com- plementary schema/instance evidence to typed decisions produces 0255075100 Queries (%) Overall ChEMBL GDC Magellan OpenData TPC-DI WikiData Complementary evidence 02040 Changed queries (%) Refinement outcomes ImprovedUnchangedHarmedRoutedBypassed replay Figure 7: Effects of evidence integration and selective refine- ment. the largest consistent gain. Unconditional refinement is harmful on ChEMBL, TPC-DI, and WikiData. Selective refinement avoids this instability and produces the strongest aggregate variant, although its improvement over the unrefined fusion is deliberately modest. Case-macro MRR for removing both components, removing com- plementary evidence, removing selective refinement, always refin- ing, and JevNexus is 0.900, 0.918, 0.942, 0.928, and 0.944, respec- tively. Corresponding Hits@1 is 0.856, 0.888, 0.921, 0.896, and 0.925; Recall@GT is 0.668, 0.674, 0.763, 0.744, and 0.764. Selective refine- ment is therefore a targeted repair mechanism, not the primary source of accuracy. 4.5 Selective Routing Table 4 reports the online gate rather than an offline replay. It opens for 766 of 13,522 source columns (5.665%), avoiding 94.335% of the listwise calls made by unconditional refinement. OpenData and GDC contain the highest proportions of conflicted, low-margin decisions; the gate never opens on Magellan or WikiData. Figure 7 tests whether the two mechanisms change the right queries. Complementary evidence improves the reciprocal rank of 10.262% of ground-truth-bearing queries and harms 0.995%. Among the 204 routed source columns with at least one ground-truth cor- respondence, refinement improves the ranking for 30.882% and degrades it for 9.804%. Conversely, replaying unconditional refine- ment on bypassed queries harms 4.257% but improves only 2.175%. The gate therefore concentrates refinement where it is more likely to help while protecting comparatively stable rankings. The avoided calls are not the primary source of the speedup over Magneto; the overall comparison contrasts typed decision Correspondences as Decisions: JevNexus for Decision-Centric Schema Matching Table 3: Component ablation by dataset. ChEMBLGDCMagellanOpenDataTPC-DIWikiData VariantMRRH@1R@GTMRRH@1R@GTMRRH@1R@GTMRRH@1R@GTMRRH@1R@GTMRRH@1R@GT w/o Both0.8300.7660.6160.7390.6420.3760.9440.8950.8550.9180.8820.6200.9610.9310.7740.8680.8170.773 w/o Complement0.8690.8320.6380.797 0.7510.4000.9900.9800.9110.9480.9350.6260.9410.9030.7620.8890.8580.773 w/o Refinement0.9040.8730.7830.7810.7030.456 1.000 1.000 1.0000.9500.935 0.6470.9800.963 0.865 0.946 0.946 0.902 Always Refine0.9010.8650.762 0.7980.7300.454 1.000 1.000 1.0000.9520.9390.6450.9350.8890.8290.9240.902 0.902 JevNexus0.907 0.878 0.7830.7940.725 0.456 1.000 1.000 1.000 0.952 0.939 0.648 0.980 0.964 0.865 0.946 0.946 0.902 Table 4: Observed online routing by dataset. DatasetSources Disagree Low margin Calls Avoided Activation (%) ChEMBL3,0607117171112,9493.627 GDC569179189385316.678 Magellan41400410.000 OpenData6,8761,5961,5155216,3557.577 TPC-DI2,916338786962,8203.292 WikiData60970600.000 Overall13,5222,8373,21476612,7565.665 inference with autoregressive generation. Within JevNexus, un- conditional refinement takes 16.242 seconds per case and selective refinement takes 15.929 seconds. The gate’s main empirical role is therefore ranking control, with a smaller secondary runtime benefit. 4.6 Sensitivity Analysis Figure 8 reports one-factor-at-a-time diagnostic sweeps on the fixed 29-case development subset. Overall denotes the dataset-macro average; the sweeps characterize, rather than select, the reported setting. Fusion and gate. Overall Recall@GT rises from 0.749 at훼=0.000 to 0.808 at훼=0.400, then declines. The per-dataset curves show that no single weight dominates: GDC peaks at 0.400, whereas OpenData peaks at 0.600. Increasing휏from 0.020 to 0.100 raises activation from 5.443% to 13.165%, but does not improve Overall Recall@GT (0.808 versus 0.807). Refinement width. For푚=2,3,5, Overall Hits@1 is 0.915, 0.914, and 0.911. Only GDC changes visibly (0.776, 0.766, and 0.749); the other datasets are unchanged. Thus the canonical훼=0.400,휏= 0.020, and푚=3 is a fixed low-routing operating point, not a sharply tuned optimum. 4.7 Error Analysis Figure 9 separates candidate coverage from downstream ordering. Candidate coverage rises from 74.832% at푘=1 to 98.032% at푘=20. At the reported operating point, 92.039% of ground-truth pairs are ranked first, 3.321% reach the final Top-3 but not rank one, 2.661% enter the candidate Top-20 but remain below the final Top-3, and 1.957% fall outside the candidate set. The remaining 0.022% are the two WikiData pairs whose target column is absent from the raw target table. GDC is the most difficult benchmark for nearly all evaluated methods because each source table is matched against a common 736-column target schema. Nevertheless, JevNexus achieves the highest GDC MRR and Hits@1 among the compared methods and nearly matches the best Recall@GT. We therefore examine the er- rors that remain at this comparatively strong operating point. On GDC, 46.718% of its ground-truth pairs are ranked first, 24.324% are in the final Top-3 but below rank one, 12.355% are in the candi- date Top-20 but below the final Top-3, and 16.602% never enter the candidate set. Its 736-column target schema exposes both a local- ization ceiling and a local-ordering problem. By contrast, Magellan is saturated, and TPC-DI has no Top-20 localization misses. The decomposition does not by itself explain the Recall@GT difference. Strong MRR and Hits@1 can coexist with weaker Re- call@GT because the latter pools scores across source columns. The observed pattern is consistent with a cross-query score-comparability issue, but establishing calibration as its cause requires a dedicated calibration experiment; we therefore leave that claim to further analysis rather than infer it from ranking metrics alone. 5 RELATED WORK Traditional and Composite Schema Matching. Classical schema matching derives correspondences from names, constraints, schema structure, and data instances. Rahm and Bernstein organize this design space along schema- versus instance-level, element- versus structure-level, and linguistic versus constraint-based evidence [25]. Cupid combines linguistic and structural similarity [18], COMA provides an extensible architecture for composing multiple match- ers [8], and Similarity Flooding propagates evidence through graph structure [19]. COMA++ and later evaluation in Valentine further demonstrate the value of combining heterogeneous signals and the sensitivity of classical matchers to data and workload charac- teristics [10,13]. JevNexus retains this central insight—matching evidence is complementary—but combines it with neural seman- tic judgments and separates evidence fusion from optional cross- candidate reasoning. Learned Schema Matching. Learning-based methods replace or adjust hand-designed similarities using representations learned from data. EmbDI constructs local relational embeddings for schema matching and entity resolution [4], whereas ADnEV learns to ad- just and evaluate similarity matrices produced by existing matchers across domains [31]. More recent systems use pretrained language models: LSM combines a fine-tuned encoder with active learning for data-free matching against large target schemas [38]; Unicorn trains a unified model across multiple data-matching tasks [34]; and ISRes- Mat adapts a pretrained encoder in situ using schema and instance evidence [9]. These systems principally improve representations, training, or transfer. JevNexus instead studies the inference archi- tecture after candidate localization: how semantic correspondence Runze Li, Hanchen Wang, Ying Zhang, and Wenjie Zhang 0.000.250.500.751.00 Fusion weight α 0.4 0.6 0.8 1.0 Recall@GT Fusion weight .005.01.02.05.10 Gate threshold τ 0.6 0.8 1.0 Recall@GT Gate quality .005.01.02.05.10 Gate threshold τ 0 5 10 15 Activation (%) Gate activation 235 Refinement width m 0.8 0.9 1.0 Hits@1 Refinement width OverallChEMBLGDCMagellanOpenDataTPC-DIWikiData Figure 8: Hyperparameter sensitivity on the development subset. 151020 Candidate limit k Overall ChEMBL GDC Magellan OpenData TPC-DI WikiData 75909498 819899100 36657483 100100100100 70849097 8499100100 92959595 Candidate coverage (%) 0255075100 GT pairs (%) Stage-level errors Top-1Top-3In Top-20OutsideInvalid Figure 9: Candidate coverage and stage-wise error decompo- sition. should be represented, combined with complementary evidence, and selectively compared. LLM-Based Schema Matching. Foundation models have been prompted for a range of data cleaning and integration tasks, estab- lishing their potential for semantic matching without task-specific architectures [5,20,29,30,37]. For schema matching, Parciak et al. study how prompt scope and available context affect generated correspondences, decisiveness, and complementarity with lexical similarity [23]. Magneto introduces the closest retrieve–rerank ar- chitecture: an efficient semantic model localizes candidate columns before a generative LLM assigns scores and reranks the localized set [17]. JevNexus adopts candidate localization as a computational abstraction, but changes the main inference interface from gener- ated scores to bounded pairwise decisions, explicitly fuses com- plementary evidence, and reserves joint listwise comparison for ambiguous candidate sets. Decision-Oriented and Selective Reranking. Related matching sys- tems commonly cast correspondence as a bounded classification problem; for example, Ditto formulates entity matching as sequence- pair classification with pretrained language models [16], while Unicorn applies a shared matching interface across heterogeneous data-matching tasks [34]. Ranking research further distinguishes in- dependent scoring from joint listwise comparison [3,35]. RankT5 di- rectly produces ranking scores under pairwise or listwise losses [39], whereas RankGPT and ListT5 use language models to order can- didates jointly [33,36]. FIRST reduces the autoregressive cost of listwise LLM reranking by deriving an ordering from the logits of a single generated identifier [26]. More broadly, RouteLLM learns to invoke a stronger model selectively to balance quality and infer- ence cost [6,12,22,28]. JevNexus brings these concerns together for schema matching: it uses direct bounded decisions on the main path and an evidence-based, training-free gate to invoke local listwise refinement only when candidate-level comparison is warranted. 6 DISCUSSION AND LIMITATIONS Effectiveness–efficiency trade-off. Decision-centric matching re- tains competitive ranking effectiveness while substantially reducing the cost of semantic inference. Under a shared candidate generator, JevNexus is statistically comparable to the strongest tested gener- ative method in dataset-macro MRR and Hits@1, while reducing mean online runtime by 7.750×. The ablations indicate that com- plementary schema and instance evidence contributes the main consistent gain, confirming that semantic representations do not subsume all useful matching signals. Listwise reasoning has a nar- rower role: applying it uniformly can degrade the ranking, whereas selective refinement is useful when evidence conflicts and the lead- ing candidates remain close. The resulting separation makes pair- wise decisions the efficient default and reserves joint comparison for unresolved cross-candidate ambiguities. Localization and score comparability. The bounded candidate set remains a hard limit on downstream matching. If a ground-truth target is absent from퐶 푘 (푠), neither pairwise decisions nor listwise refinement can recover it. Simply increasing푘may improve cov- erage, but also increases inference cost and exposes the matcher to more distractors. An important direction is therefore adaptive localization that expands the candidate set only when the available evidence indicates that the initial set is insufficient. A separate lim- itation concerns the use of decision and fused scores across source queries. The strong query-level ranking results show that these scores are effective for ordering candidates for the same source, but their weaker pooled correspondence recall suggests limited comparability across source queries, motivating direct evaluation and calibration of cross-query scores. Transferability of fusion and routing. The current system uses fixed values for the fusion coefficient, refinement width, and routing margin. Sensitivity analysis identifies a stable, low-routing operat- ing point, but also shows that no single setting dominates across all benchmark families. This variation is expected because datasets differ in the informativeness of column names and values, the den- sity of plausible candidates, and the width of their target schemas. The shared configuration provides a stable operating point across Correspondences as Decisions: JevNexus for Decision-Centric Schema Matching datasets, while learning evidence weights, adapting candidate and refinement widths per query, and routing from the full pattern of evidence remain promising directions. Evaluation and system scope. The evaluation spans six hetero- geneous benchmark families, but it does not cover every schema- matching regime. In particular, several families contain relatively few cases, and the experiments use one decision backbone and one listwise refinement model. Broader evaluation across unseen domains, model families, and substantially larger target schemas is therefore needed to establish how far the observed trade-off trans- fers. Absolute latency depends on hardware, serving infrastructure, and batching; our measurements therefore characterize relative efficiency under a shared experimental environment. 7 CONCLUSION We presented JevNexus, a decision-centric approach to schema matching that combines typed pairwise judgments with comple- mentary schema and instance evidence, and invokes listwise rea- soning only for uncertain cases. This design preserves localized correspondence evidence while avoiding the cost and occasional degradation associated with unconditional generative reranking. Across six benchmark families, JevNexus achieves MRR and Hits@1 statistically comparable to those of Magneto while reduc- ing mean online runtime by 7.750×, although Magneto attains higher Recall@GT. The experiments further show that evidence fusion provides a consistent benefit and that selective refinement is most useful when the evidence sources disagree and the resulting ranking remains uncertain. Retrieval coverage remains the principal limitation: refinement cannot recover a correct correspondence ab- sent from the candidate set. Future work will therefore investigate adaptive candidate expansion and calibration across source queries, together with broader evaluation across domains and execution environments. REFERENCES [1] Michael J. Cafarella, Alon Halevy, Daisy Zhe Wang, Eugene Wu, and Yang Zhang. 2008. WebTables: Exploring the Power of Tables on the Web. Proceedings of the VLDB Endowment 1, 1 (2008), 538–549. https://doi.org/10.14778/1453856.1453916 [2]Zefan Cai. 2026. Open-Jev-9B. Hugging Face model repository.https: //huggingface.co/ZefanCai/Open-Jev-9B Accessed 2026-09-26. [3]Zhe Cao, Tao Qin, Tie-Yan Liu, Ming-Feng Tsai, and Hang Li. 2007. Learning to Rank: From Pairwise Approach to Listwise Approach. In Proceedings of the 24th International Conference on Machine Learning (ICML). 129–136.https: //doi.org/10.1145/1273496.1273513 [4]Riccardo Cappuzzo, Paolo Papotti, and Saravanan Thirumuruganathan. 2020. Creating Embeddings of Heterogeneous Relational Datasets for Data Integration Tasks. In Proceedings of the 2020 ACM SIGMOD International Conference on Management of Data. 1335–1349. https://doi.org/10.1145/3318464.3389742 [5]Houming Chen, Zhe Zhang, and H. V. Jagadish. 2026. ConStruM: A Structure- Guided LLM Framework for Context-Aware Schema Matching. arXiv preprint arXiv:2601.20482 (2026). https://doi.org/10.48550/arXiv.2601.20482 [6]Jasper Dekoninck, Maximilian Baader, and Martin Vechev. 2025. A Unified Approach to Routing and Cascading for LLMs. In Proceedings of the 42nd Interna- tional Conference on Machine Learning (ICML) (Proceedings of Machine Learning Research), Vol. 267. 12987–13010. [7] Yuhao Deng, Chengliang Chai, Lei Cao, Qin Yuan, Siyuan Chen, Yanrui Yu, Zhaoze Sun, Junyi Wang, Jiajun Li, Ziqi Cao, Kaisen Jin, Chi Zhang, Yuqing Jiang, Yuanfang Zhang, Yuping Wang, Ye Yuan, Guoren Wang, and Nan Tang. 2024. LakeBench: A Benchmark for Discovering Joinable and Unionable Tables in Data Lakes. Proceedings of the VLDB Endowment 17, 8 (2024), 1925–1938. https://doi.org/10.14778/3659437.3659448 [8]Hong-Hai Do and Erhard Rahm. 2002. COMA – A System for Flexible Combina- tion of Schema Matching Approaches. In Proceedings of the 28th International Con- ference on Very Large Databases (VLDB). 610–621. https://doi.org/10.1016/B978- 155860869-6/50060-3 [9]Xingyu Du, Gongsheng Yuan, Sai Wu, Gang Chen, and Peng Lu. 2024. In Situ Neu- ral Relational Schema Matcher. In Proceedings of the 40th IEEE International Con- ference on Data Engineering (ICDE). 138–150. https://doi.org/10.1109/ICDE60146. 2024.00018 [10]Daniel Engmann and Sabine Massmann. 2007. Instance Matching with COMA++. In BTW Workshops, Vol. 7. 28–37. [11]Raul Castro Fernandez, Ziawasch Abedjan, Famien Koko, Gina Yuan, Samuel Madden, and Michael Stonebraker. 2018. Aurum: A Data Discovery System. In Proceedings of the 34th IEEE International Conference on Data Engineering (ICDE). 1001–1012. https://doi.org/10.1109/ICDE.2018.00094 [12]Neha Gupta, Harikrishna Narasimhan, Wittawat Jitkrittum, Ankit Singh Rawat, Aditya Krishna Menon, and Sanjiv Kumar. 2024. Language Model Cascades: Token-Level Uncertainty and Beyond. In International Conference on Learning Representations (ICLR). [13]Christos Koutras, George Siachamis, Andra Ionescu, Kyriakos Psarakis, Jerry Brons, Marios Fragkoulis, Christoph Lofi, Angela Bonifati, and Asterios Katsifodi- mos. 2021. Valentine: Evaluating Matching Techniques for Dataset Discovery. In Proceedings of the 37th IEEE International Conference on Data Engineering (ICDE). 468–479. [14]Runze Li, Hanchen Wang, Wenjie Zhang, Binghao Li, Yu Zhang, Xuemin Lin, and Ying Zhang. 2025. FADTI: Fourier and Attention Driven Diffusion for Multivariate Time Series Imputation. arXiv preprint arXiv:2512.15116 (2025). https://doi.org/10.48550/arXiv.2512.15116 [15]Yize Li, Yongchao Dou, Felipe da Veiga Leprevost, Yifat Geffen, Anna P. Cali- nawan, François Aguet, et al.2023. Proteogenomic Data and Resources for Pan-Cancer Analysis. Cancer Cell 41, 8 (2023), 1397–1406. https://doi.org/10. 1016/j.ccell.2023.06.009 [16]Yuliang Li, Jinfeng Li, Yoshihiko Suhara, AnHai Doan, and Wang-Chiew Tan. 2020. Deep Entity Matching with Pre-Trained Language Models. Proceedings of the VLDB Endowment 14, 1 (2020), 50–60. https://doi.org/10.14778/3421424.3421431 [17]Yurong Liu, Eduardo H. M. Pena, Aécio S. R. Santos, Eden Wu, and Juliana Freire. 2025. Magneto: Combining Small and Large Language Models for Schema Matching. Proceedings of the VLDB Endowment 18, 8 (2025), 2681–2694. https: //doi.org/10.14778/3742728.3742757 [18]Jayant Madhavan, Philip A. Bernstein, and Erhard Rahm. 2001. Generic Schema Matching with Cupid. In Proceedings of the 27th International Conference on Very Large Data Bases (VLDB). 49–58. [19] Sergey Melnik, Hector Garcia-Molina, and Erhard Rahm. 2002. Similarity Flood- ing: A Versatile Graph Matching Algorithm and Its Application to Schema Match- ing. In Proceedings of the 18th IEEE International Conference on Data Engineering (ICDE). 117–128. https://doi.org/10.1109/ICDE.2002.994702 [20]Avanika Narayan, Ines Chami, Laurel Orr, and Christopher Ré. 2022. Can Foun- dation Models Wrangle Your Data? Proceedings of the VLDB Endowment 16, 4 (2022), 738–746. https://doi.org/10.14778/3574245.3574258 [21] Christina Nasika, Feng Wang, Antonis Minas Krasakis, and Han Xiao. 2026. jina-reranker-v3.5: An Efficient Listwise Reranker with Hybrid Attention and Self-Distillation. arXiv preprint arXiv:2607.18152 (2026). [22] Isaac Ong, Amjad Almahairi, Vincent Wu, Wei-Lin Chiang, Tianhao Wu, Joseph E. Gonzalez, M. Waleed Kadous, and Ion Stoica. 2025. RouteLLM: Learning to Route LLMs with Preference Data. In Proceedings of the 13th International Conference on Learning Representations (ICLR). [23]Marcel Parciak, Brecht Vandevoort, Frank Neven, Liesbet M. Peeters, and Stijn Vansummeren. 2024. Schema Matching with Large Language Models: An Experi- mental Study. In Proceedings of the VLDB 2024 Workshop on Tabular Data Analysis (TaDA). https://w.vldb.org/workshops/2024/proceedings/TaDA/TaDA.8.pdf [24]Qwen Team. 2026. Qwen3.5-9B. Hugging Face model repository.https:// huggingface.co/Qwen/Qwen3.5-9B Accessed 2026-09-26. [25]Erhard Rahm and Philip A. Bernstein. 2001. A Survey of Approaches to Automatic Schema Matching. The VLDB Journal 10, 4 (2001), 334–350. https://doi.org/10. 1007/s007780100057 [26]Revanth Gangi Reddy, JaeHyeok Doo, Yifei Xu, Md Arafat Sultan, Deevya Swain, Avirup Sil, and Heng Ji. 2024. FIRST: Faster Improved Listwise Reranking with Single Token Decoding. In Proceedings of the 2024 Conference on Empirical Meth- ods in Natural Language Processing. 8642–8652. https://doi.org/10.18653/v1/2024. emnlp-main.491 [27]Nils Reimers and Iryna Gurevych. 2019. Sentence-BERT: Sentence Embeddings using Siamese BERT-Networks. Proceedings of EMNLP-IJCNLP (2019), 3982–3992. [28]Tal Schuster, Adam Fisch, Jai Prakash Gupta, Mostafa Dehghani, Dara Bahri, Vinh Q. Tran, Yi Tay, and Donald Metzler. 2022. Confident Adaptive Language Modeling. In Advances in Neural Information Processing Systems, Vol. 35. [29]Nabeel Seedat and Mihaela van der Schaar. 2024. Matchmaker: Self-Improving Large Language Model Programs for Schema Matching. In Generative AI and Health Workshop at NeurIPS. https://arxiv.org/abs/2410.24105 [30]Eitam Sheetrit, Menachem Brief, Moshik Mishaeli, and Oren Elisha. 2024. Re- Match: Retrieval Enhanced Schema Matching with LLMs. arXiv preprint arXiv:2403.01567 (2024). https://doi.org/10.48550/arXiv.2403.01567 Runze Li, Hanchen Wang, Ying Zhang, and Wenjie Zhang [31]Roee Shraga, Avigdor Gal, and Haggai Roitman. 2020. ADnEV: Cross-Domain Schema Matching using Deep Similarity Matrix Adjustment and Evaluation. Proceedings of the VLDB Endowment 13, 9 (2020), 1401–1415. https://doi.org/10. 14778/3397230.3397237 [32]Kaitao Song, Xu Tan, Tao Qin, Jianfeng Lu, and Tie-Yan Liu. 2020. MPNet: Masked and Permuted Pre-training for Language Understanding. In Advances in Neural Information Processing Systems, Vol. 33. 16857–16867. [33]Weiwei Sun, Lingyong Yan, Xinyu Ma, Shuaiqiang Wang, Pengjie Ren, Zhumin Chen, Dawei Yin, and Zhaochun Ren. 2023. Is ChatGPT Good at Search? Investi- gating Large Language Models as Re-Ranking Agents. In Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing. 14918–14937. https://doi.org/10.18653/v1/2023.emnlp-main.923 [34] Jianhong Tu, Ju Fan, Nan Tang, Peng Wang, Guoliang Li, Xiaoyong Du, Xiaofeng Jia, and Song Gao. 2023. Unicorn: A Unified Multi-tasking Model for Supporting Matching Tasks in Data Integration. Proceedings of the ACM on Management of Data 1, 1 (2023), 84:1–84:26. https://doi.org/10.1145/3588938 [35] Fen Xia, Tie-Yan Liu, Jue Wang, Wensheng Zhang, and Hang Li. 2008. Listwise Approach to Learning to Rank: Theory and Algorithm. In Proceedings of the 25th International Conference on Machine Learning (ICML). 1192–1199. https: //doi.org/10.1145/1390156.1390306 [36]Soyoung Yoon, Eunbi Choi, Jiyeon Kim, Hyeongu Yun, Yireun Kim, and Seung won Hwang. 2024. ListT5: Listwise Reranking with Fusion-in-Decoder Improves Zero-shot Retrieval. In Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics. 2287–2308. https://doi.org/10.18653/v1/2024.acl- long.125 [37]Haochen Zhang, Yuyang Dong, Chuan Xiao, and Masafumi Oyamada. 2024. Jelly- fish: Instruction-Tuning Local Large Language Models for Data Preprocessing. In Proceedings of the 2024 Conference on Empirical Methods in Natural Language Pro- cessing (EMNLP). 8754–8782. https://doi.org/10.18653/v1/2024.emnlp-main.497 [38]Yunjia Zhang, Avrilia Floratou, Joyce Cahoon, Subru Krishnan, Andreas C. Müller, Dalitso Banda, Fotis Psallidas, and Jignesh M. Patel. 2023. Schema Matching using Pre-Trained Language Models. In Proceedings of the 39th IEEE International Conference on Data Engineering (ICDE). 1558–1571. https://doi.org/10.1109/ ICDE55515.2023.00123 [39]Honglei Zhuang, Zhen Qin, Rolf Jagerman, Kai Hui, Ji Ma, Jing Lu, Jianmo Ni, Xuanhui Wang, and Michael Bendersky. 2023. RankT5: Fine-Tuning T5 for Text Ranking with Ranking Losses. In Proceedings of the 46th International ACM SIGIR Conference on Research and Development in Information Retrieval. 2308–2313. https://doi.org/10.1145/3539618.3592047