GenTUS: Table Union Search via Constrained Generation of Semantic Table Identifiers

arXiv:2607.00833 · cs.DB · Submitted 2026-07-01 · Read on arXiv

Listen

Radio episode about this paper

Transcript

Introduction to the show: ident: Computer Science. The day in computer science: software, systems and theory on arXiv.

Tom: We're looking at the paper: "GenTUS: Table Union Search via Constrained Generation of Semantic Table Identifiers".

Jane: Here's the gist of the paper. Modern data lakes contain heterogeneous tables whose task-relevant information is often scattered across different schemas, sources, and naming conventions <ref:2607.00833#pg2> Table union search (TUS) retrieves tables that can be reliably unioned with a query table, supporting data discovery, enrichment, and downstream analytics <ref:2607.00833#pg2>.

Tom: GenTUS proposes a generative retrieval framework that reformulates TUS as constrained generation over discrete semantic table identifiers <ref:2607.00833#pg2> The core innovation of the proposed framework GenTUS reformulates TUS as constrained generation over discrete semantic table identifiers <ref:2607.00833#pg2>.

Jane: Instead of searching and reranking an explicit candidate pool, GenTUS assigns candidate tables compact unionability-aware identifiers and trains a generator to produce the identifiers of unionable tables directly from the query <ref:2607.00833#pg2>.

Tom: At query time, constrained decoding ensures that generated identifiers correspond to valid data-lake tables and returns them as ranked retrieval results <ref:2607.00833#pg2>.

Paper summary: Tom: So, wrapping up the "GenTUS: Table Union Search via Constrained Generation of Semantic Table Identifiers" paper. The authors are proposing this generative approach to TUS as a way to improve retrieval quality while cutting down on offline build time and online retrieval time.

Jane: It really boils down to reformulating table union search as constrained generation over discrete semantic table identifiers, which is a significant shift from the traditional encode search refine pipeline.

Lu: What this implies for the world of data lakes is that we can move away from relying on massive vector indexes and explicit candidate retrieval during query time for tasks like data discovery and enrichment.

Meng: For me, it means potentially lower operational costs because of the storage savings and faster online search, although I’d still be watching how reliable the generation holds up under extreme distribution shift.

Tom: The paper shows that GenTUS achieves a rank of one point zero five on average against a strong baseline, which is better than two point five seven.

Jane: So if you listen to this episode, remember that instead of searching through millions of candidates and re-scoring them, the system generates the right table identifiers directly from your query.

Conclusion: Tom: So, GenTUS is taking table union search and turning it into constrained generation over discrete semantic table identifiers.

Jane: It’s about ditching that whole encode search refine pipeline we talked about before, right?

Lu: Exactly. They are replacing the explicit candidate retrieval with a generator that produces those identifiers directly from the query.

Meng: From my side, I’m looking at how much offline training this takes and if it actually scales without needing massive infrastructure just to build that generator space.

Lalam: Based on what I see in the training signal construction, they are using a two-part objective: one for making sure the identifiers are semantically meaningful and another one specifically for pulling together tables that can actually be unioned.

Tom: That pairwise unionability supervision is interesting because it directly addresses the core problem of finding compatible tables.

Jane: It means the system isn't just guessing; it’s learning a structure where "unionable" is baked into the identifier space itself before any query even comes in.

Lu: And they enforce validity at inference time using a prefix tree structure, which guarantees that what the generator outputs actually maps back to a real table in that lake.

Meng: That sounds like it could reduce the storage overhead for candidate retrieval because you aren't maintaining a massive index of millions of candidates anymore.

Tom: The numbers on SANTOS large are pretty telling, showing online time going from thousands of seconds down to about seventy seconds. That’s a big jump in efficiency for real-time discovery.

Jane: So, the implication here is that we can finally get meaningful table union search without paying for the huge candidate pool storage and slow reranking steps every single time we ask a question.

Lalam: I'm fairly sure of this because they show it maintaining retrieval quality even when the pool size goes from one thousand tables all the way up to the whole lake.

Tom: It seems like a viable alternative to the complex encode search refine methods we’ve been using for TUS.

Jane: But they aren't done yet; they mentioned future work involving extending this to joint unionable and joinable table discovery, which could open up even more ways to connect data sources.

Shulun Zhang, Linting Wang, Yuwei Xu, Yingli Zhou, Chenhao Ma

The Chinese University of Hong Kong, Shenzhen · Fudan University

cs.DB

Submitted: 2026-07-01

Updated: 2026-11-13

Code: https://github.com/alanzhang1001/GenTUS

License: http://creativecommons.org/licenses/by/4.0/

Importance score: 96/100

The gist: Modern data lakes contain heterogeneous tables whose task-relevant information is often scattered across different schemas, sources…

Key concepts

Table Union Search (TUS)
TUS is a process used in modern data lakes to find tables that can be reliably combined with a given query table. It treats unionability as a semantic relationship, meaning two tables might be compatible even if their schemas or column names differ, allowing for richer data discovery.
Semantic Table Identifier
Each candidate table is assigned a compact identifier composed of tokens. These identifiers are designed to be meaningful and structured so that the retrieval process can be modeled as generating a sequence of these identifiers based on the query input.
Constrained Generation
This is the core mechanism where a trained generator takes the query table as input and generates a sequence of semantic table identifiers. Constraints ensure that only valid, existing data lake tables are generated, guiding the model to produce relevant unionable candidates directly.
Unionability-aware Contrastive Objective
This training signal is used during offline processing to teach the system which tables can be unioned together. It pulls annotated unionable tables closer in a latent space, ensuring that the resulting semantic identifiers for compatible tables share a similar structure.

Terminology

Summary

Modern data lakes contain heterogeneous tables whose task-relevant information is often scattered across different schemas, sources, and naming conventions <ref:2607. p. 0

Where the paper says it

cn Chenhao Ma The Chinese University of Hong Kong, Shenzhen Shenzhen, China machenhao@cuhk. cn Abstract Modern data lakes contain heterogeneous tables whose task-relevant information is often scattered across different schemas, sources, and naming conventions.

00833#pg2> Table union search (TUS) retrieves tables that can be reliably unioned with a query table, supporting data discovery, enrichment, and downstream analytics <ref:2607. p. 0
Where the paper says it

cn Abstract Modern data lakes contain heterogeneous tables whose task-relevant information is often scattered across different schemas, sources, and naming conventions. Table union search (TUS) retrieves tables that can be reliably unioned with a query table, supporting data discovery, enrichment, and downstream analytics.

p. 0
Where the paper says it

cn Abstract Modern data lakes contain heterogeneous tables whose task-relevant information is often scattered across different schemas, sources, and naming conventions. Table union search (TUS) retrieves tables that can be reliably unioned with a query table, supporting data discovery, enrichment, and downstream analytics.

00833#pg2>. GenTUS proposes a generative retrieval framework that reformulates TUS as constrained generation over discrete semantic table identifiers <ref:2607. p. 0
Where the paper says it

Although learning-based TUS methods improve table- or column-level representations, they still follow an encode–search–refine pipeline: candidate retrieval is followed by query–candidate matching or reranking, making quality dependent on candidate-pool recall and incurring growing latency and storage costs as the data lake scales. We propose GenTUS, a generative retrieval framework that reformulates TUS as constrained generation over discrete semantic table identifiers.

00833#pg2>

The core innovation of the proposed framework

GenTUS reformulates TUS as constrained generation over discrete semantic table identifiers <ref:2607.00833#pg2>. Instead of searching and reranking an explicit candidate pool, GenTUS assigns candidate tables compact unionability-aware identifiers and trains a generator to produce the identifiers of unionable tables directly from the query <ref:2607. p. 0

Where the paper says it

We propose GenTUS, a generative retrieval framework that reformulates TUS as constrained generation over discrete semantic table identifiers. Instead of searching and reranking an explicit candidate pool, GenTUS assigns candidate tables compact unionabilityaware identifiers and trains a generator to produce the identifiers of unionable tables directly from the query.

p. 0
Where the paper says it

We propose GenTUS, a generative retrieval framework that reformulates TUS as constrained generation over discrete semantic table identifiers. Instead of searching and reranking an explicit candidate pool, GenTUS assigns candidate tables compact unionabilityaware identifiers and trains a generator to produce the identifiers of unionable tables directly from the query.

00833#pg2>. At query time, constrained decoding ensures that generated identifiers correspond to valid data-lake tables and returns them as ranked retrieval results <ref:2607. p. 0
Where the paper says it

Instead of searching and reranking an explicit candidate pool, GenTUS assigns candidate tables compact unionabilityaware identifiers and trains a generator to produce the identifiers of unionable tables directly from the query. At query time, constrained decoding ensures that generated identifiers correspond to valid data-lake tables and returns them as ranked retrieval results.

p. 0
Where the paper says it

Instead of searching and reranking an explicit candidate pool, GenTUS assigns candidate tables compact unionabilityaware identifiers and trains a generator to produce the identifiers of unionable tables directly from the query. At query time, constrained decoding ensures that generated identifiers correspond to valid data-lake tables and returns them as ranked retrieval results.

00833#pg2>. This paradigm shift avoids explicit lake-wide ANN search and query–candidate reranking at inference time and eliminates the storage overhead of maintaining embeddings and retrieval indexes <ref:2607. p. 1
Where the paper says it

During retrieval, GenTUS generates target table identifiers directly via beam search over a prefix tree, eliminating the need for vector indexes, ANN search, or reranking models. This paradigm shift avoids explicit lake-wide ANN search and query– candidate reranking at inference time and eliminates the storage overhead of maintaining embeddings and retrieval indexes.

00833#pg2>.

How the training signal is built

The offline stage builds the generative candidate space used by GenTUS and trains the retrieval generator <ref:2607.00833#pg2>. This involves two conceptual steps:

  1. GenTUS constructs semantic identifiers for candidate tables, incorporating a unionability-aware contrastive objective to pull annotated unionable tables together in the latent space <ref:2607. p. 4

    Where the paper says it

    However, this quantization process by itself does not explicitly optimize identifiers for unionability. GenTUS therefore augments residual quantization with a unionability-aware contrastive objective, which pulls annotated unionable tables together in the quantizer latent space before discretization.

    00833#pg2>.

  2. GenTUS builds query-to-identifier training records and performs offline generative training, where the generator learns to output identifiers of annotated unionable candidates under incidental layout variations <ref:2607. p. 5

    Where the paper says it

    3 Training Instance Construction After semantic identifiers are assigned, GenTUS converts annotated query–candidate unionability pairs into identifier-level supervision for the retrieval generator. This subsection constructs positive generation records and augments them with structure-preserving query views, so the generator learns to output identifiers of annotated unionable candidates under incidental layout variations.

    00833#pg2>.

Where the gains come from

The effectiveness of GenTUS comes from its two core designs:

  1. Unionability-aware semantic identifier construction incorporates pairwise unionability supervision, so that unionable tables receive generation-friendly identifiers with shared coarse structure <ref:2607. p. 1

    Where the paper says it

    • We propose GenTUS, the first generative retrieval framework for TUS, reformulating table union search as constrained generation over discrete semantic table identifiers. • We design unionability-aware identifier construction that incorporates pairwise unionability supervision, so that unionable tables receive generation-friendly identifiers with shared coarse structure.

    00833#pg2>.

  2. Prefix-tree-based constrained decoding guarantees valid identifier generation and avoids explicit lake-wide candidate search during online retrieval <ref:2607. p. 1

    Where the paper says it

    • We design unionability-aware identifier construction that incorporates pairwise unionability supervision, so that unionable tables receive generation-friendly identifiers with shared coarse structure. • We propose prefix-tree-based constrained decoding that guarantees valid identifier generation and avoids explicit lake-wide candidate search during online retrieval.

    00833#pg2>.

Online retrieval mechanism

The online stage uses the generator and identifier structures produced offline to retrieve unionable tables through generation <ref:2607.00833#pg2>. Given a query table, GenTUS serializes it as the generator input, performs prefix-constrained decoding over valid semantic identifiers, and maps the completed identifiers back to candidate tables <ref:2607. p. 6

Where the paper says it

6 Online Retrieval The online stage uses the generator and identifier structures produced offline to retrieve unionable tables through generation. Given a query table, GenTUS serializes it as the generator input, performs prefix-constrained decoding over valid semantic identifiers, and maps the completed identifiers back to candidate tables (Section 6.

00833#pg2>. Validity is enforced by the prefix tree TrieD, which restricts the next token to prefixes that can still lead to an existing table identifier <ref:2607. p. 3
Where the paper says it

The decoder generates semantic identifiers token by token. At each decoding step, the prefix tree TrieD restricts the next token to prefixes that can still lead to an existing table identifier in SD.

00833#pg2>. The order of Rq is inherited from the constrained beam search, so GenTUS uses it directly as the retrieval ranking without post-hoc query–candidate alignment scoring over an ANN-retrieved candidate pool <ref:2607. p. 7
Where the paper says it

The final top-𝑘 result is 𝜋𝑘 (𝑞, D) = FirstK𝑘 (R𝑞). The order of R𝑞 is inherited from the constrained beam search, so GenTUS uses it directly as the retrieval ranking without post-hoc query–candidate alignment scoring over an ANN-retrieved candidate pool.

00833#pg2>.

System efficiency and scalability

GenTUS is consistently the most efficient method across offline build time, online retrieval time, and retrieval-artifact storage <ref:2607.00833#pg2>. For example, on SANTOS large it takes 72. p. 9

Where the paper says it

TUS Small TUS Large SANTOS Small SANTOS Large Wiki Union LakeBench 1K LakeBench 30K 1 10 100 1k 10k 100k Storage (MB, log scale) GenTUS TACTUS Starmie LIFTus Sherlock SATO Figure 8: Retrieval-artifact storage (log scale, MB): semantic identifiers and prefix tree for GenTUS, and embeddings, indexes or other retrieval artifacts for baselines. TUS_small it takes 71 seconds versus 463 for the next fastest baseline (6.

6 seconds versus 4434 for TACTUS (61×) in online time <ref:2607. p. 9
Where the paper says it

On SANTOS_large it takes 72. 6 seconds versus 4434 for TACTUS (61×), and on LakeBench_1k it takes 1, 849 seconds versus 66, 692 seconds.

00833#pg2>. Furthermore, GenTUS requires the least retrieval-artifact storage on every dataset <ref:2607. p. 9
Where the paper says it

These results show that the advantage of GenTUS holds across the full range of 𝐾 rather than at any single operating point. 3 Efficiency and Storage We report three system costs: offline build time (Figure 6), online end-to-end time (Figure 7), and retrieval-artifact storage (Figure 8).

00833#pg2>. The framework also supports low-cost table insertion by only requiring a local update to the SID tree without rebuilding any index <ref:2607. p. 1
Where the paper says it

To address this, we organize all valid identifiers into a prefix tree and impose prefix constraints during decoding, ensuring that every generated identifier maps to an existing table in the data lake. Online retrieval then reduces to constrained beam search over the SID tree, with the decoding budget governed by beam width and identifier depth rather than explicit search over the data lake; inserting a new table requires only a local update to the SID tree without rebuilding any index.

00833#pg2>.

Robustness and ablation analysis

Experiments show that GenTUS maintains retrieval quality as the candidate pool grows, remaining stable across pool sizes from 1,000 tables to the full lake <ref:2607.00833#pg2>. Ablation studies confirm that semantic IDs should be both semantically meaningful and uniquely addressable, while the unionability-aware objective provides additional gains <ref:2607.00833#pg2>. Prefix-constrained decoding guarantees valid table identifiers and improves retrieval quality when those additional valid candidates include relevant tables <ref:2607. p. 1

Where the paper says it

• We design unionability-aware identifier construction that incorporates pairwise unionability supervision, so that unionable tables receive generation-friendly identifiers with shared coarse structure. • We propose prefix-tree-based constrained decoding that guarantees valid identifier generation and avoids explicit lake-wide candidate search during online retrieval.

00833#pg2>. The results show that generative retrieval is a viable and practical alternative to encode–search–refine pipelines for TUS <ref:2607. p. 0
Where the paper says it

Although learning-based TUS methods improve table- or column-level representations, they still follow an encode–search–refine pipeline: candidate retrieval is followed by query–candidate matching or reranking, making quality dependent on candidate-pool recall and incurring growing latency and storage costs as the data lake scales. We propose GenTUS, a generative retrieval framework that reformulates TUS as constrained generation over discrete semantic table identifiers.

00833#pg2>.

Conclusion

GenTUS, a generative framework that reformulates table union search as constrained generation over discrete semantic table identifiers, improves retrieval quality while reducing offline build time, online retrieval time, and index storage <ref:2607.00833#pg2>. It also remains effective as the candidate pool grows and supports low-cost table insertion <ref:2607. p. 7

Where the paper says it

The resulting final identifier is inserted into the valid identifier set SD, the identifier mapping 𝜙, and the prefix tree TrieD, while all existing identifiers remain unchanged. If the latent dimension is 𝑑𝑧 , the update cost for identifier assignment and trie insertion is𝑂(𝐿𝐾𝑑𝑧 + | sid(𝑇 ′ )|), excluding the cost of table encoding.

00833#pg2>. Future work includes extending the formulation to joint unionable-table and joinable-table discovery, refreshing the identifier space and the generator incrementally under substantial distribution shift, and scaling the approach to even larger data lakes <ref:2607.00833#pg2>.

--- Page 1 ---

Generative Retrieval for Table Union Search Shulun Zhang Linting Wang Yuwei Xu Yingli Zhou Chenhao Ma Abstract Modern data lakes contain heterogeneous tables whose task-relevant information is often scattered across different schemas, sources, and naming conventions <ref:2607.00833#pg2> Table union search (TUS) retrieves tables that can be reliably unioned with a query table, supporting data discovery, enrichment, and downstream analytics <ref:2607.00833#pg2>. Although learning-based TUS methods improve table- or column-level representations, they still follow an encode–search–refine pipeline: candidate retrieval is followed by query–candidate matching or reranking, making quality dependent on candidate-pool recall and incurring growing latency and storage costs as the data lake scales <ref:2607. p. 0

Where the paper says it

Table union search (TUS) retrieves tables that can be reliably unioned with a query table, supporting data discovery, enrichment, and downstream analytics. Although learning-based TUS methods improve table- or column-level representations, they still follow an encode–search–refine pipeline: candidate retrieval is followed by query–candidate matching or reranking, making quality dependent on candidate-pool recall and incurring growing latency and storage costs as the data lake scales.

00833#pg2>. We propose GenTUS, a generative retrieval framework that reformulates TUS as constrained generation over discrete semantic table identifiers <ref:2607. p. 0
Where the paper says it

Although learning-based TUS methods improve table- or column-level representations, they still follow an encode–search–refine pipeline: candidate retrieval is followed by query–candidate matching or reranking, making quality dependent on candidate-pool recall and incurring growing latency and storage costs as the data lake scales. We propose GenTUS, a generative retrieval framework that reformulates TUS as constrained generation over discrete semantic table identifiers.

00833#pg2>. Instead of searching and reranking an explicit candidate pool, GenTUS assigns candidate tables compact unionability-aware identifiers and trains a generator to produce the identifiers of unionable tables directly from the query <ref:2607.00833#pg2>. At query time, constrained decoding ensures that generated identifiers correspond to valid data-lake tables and returns them as ranked retrieval results <ref:2607.00833#pg2>. Experiments on seven public TUS benchmarks show that GenTUS achieves the best overall retrieval quality, with an average rank of 1. p. 0
Where the paper says it

At query time, constrained decoding ensures that generated identifiers correspond to valid data-lake tables and returns them as ranked retrieval results. Experiments on seven public TUS benchmarks show that GenTUS achieves the best overall retrieval quality, with an average rank of 1.

05 compared to 2. p. 0
Where the paper says it

Experiments on seven public TUS benchmarks show that GenTUS achieves the best overall retrieval quality, with an average rank of 1. 05 compared to 2.

57 for the strongest baseline, while substantially reducing online latency, retrieval-artifact storage, and incremental update cost <ref:2607. p. 0
Where the paper says it

05 compared to 2. 57 for the strongest baseline, while substantially reducing online latency, retrieval-artifact storage, and incremental update cost.

00833#pg2>.

--- Page 2 ---

Prior TUS methods encode-search-refine 1 Query Table Encode Query Embeddings (table/column)ANN index Candidate Retrieval C candidates C >> k! Large pool Heavy storage Candidate Refinement query q T1 q + T1 s1 T2 q + T2 s2 TN q + TN sN Sort candidates T17 0. p. 1

Where the paper says it

Large pool Heavy storage 4. Candidate Refinement query q T1 q + T1 s1 T2 q + T2 s2 T3 q + T3 s3 TN q + TN sN .

92 T03 0.89 T41 0.86 Tk 0.75! C candidate scores / high latency GenTUS Generate Semantic IDs Encoder-Decoder No ANN index, No per-candidate scoring map IDs to tables +18% ↑ 12. p. 1
Where the paper says it

Top-k Tables GenTUS 2. Generate Semantic IDs Encoder-Decoder No ANN index, No per-candidate scoring map IDs to tables +18% ↑ 12.

6× faster ↓ 13× smaller ↓

--- Page 3 ---

2 Preliminaries 2.1 Table Union Search Let D = T1,. p. 2

Where the paper says it

2 Preliminaries 2. 1 Table Union Search Let D = {𝑇1, .

..,TN denote the candidate data lake, where each table Ti consists of a set of columns Ci and a collection of tuples defined over Ci Given a query table q, table union search aims to retrieve a ranked list of candidate tables from D that are unionable with q Following prior work [17, 25, 33, 38], we treat unionability as a semantic relation between tables rather than a strict schemamatching problem: two tables may be unionable even when their column names, schemas, or metadata differ, as long as their columns can be aligned under compatible semantics <ref:2607. p. 2
Where the paper says it

Given a query table 𝑞, table union search aims to retrieve a ranked list of candidate tables from D that are unionable with 𝑞. Following prior work [17, 25, 33, 38], we treat unionability as a semantic relation between tables rather than a strict schemamatching problem: two tables may be unionable even when their column names, schemas, or metadata differ, as long as their columns can be aligned under compatible semantics.

00833#pg2>. We write μ(q,T) for a unionability score between q and a candidate T, with larger values indicating stronger unionability <ref:2607. p. 2
Where the paper says it

Following prior work [17, 25, 33, 38], we treat unionability as a semantic relation between tables rather than a strict schemamatching problem: two tables may be unionable even when their column names, schemas, or metadata differ, as long as their columns can be aligned under compatible semantics. We write 𝜇(𝑞,𝑇 ) for a unionability score between 𝑞 and a candidate 𝑇 , with larger values indicating stronger unionability.

00833#pg2>. Definition 2. p. 2
Where the paper says it

We write 𝜇(𝑞,𝑇 ) for a unionability score between 𝑞 and a candidate 𝑇 , with larger values indicating stronger unionability. Definition 2.

1 (Top-k Table Union Search). p. 2
Where the paper says it

Definition 2. 1 (Top-𝑘 Table Union Search).

Following the standard TUS formulation [25, 33], given a data lake D, a query table q, and a positive integer k, top-k table union search returns a ranked list of k distinct tables πk (q, D) = (T(1),. p. 2
Where the paper says it

1 (Top-𝑘 Table Union Search). Following the standard TUS formulation [25, 33], given a data lake D, a query table 𝑞, and a positive integer 𝑘, top-𝑘 table union search returns a ranked list of 𝑘 distinct tables 𝜋𝑘 (𝑞, D) = (𝑇(1) , .

..,T(k)) ordered by a unionability score μ: μ(q,T(j)) ≥ μ(q,T(j+1)) for j = 1,. p. 2
Where the paper says it

Following the standard TUS formulation [25, 33], given a data lake D, a query table 𝑞, and a positive integer 𝑘, top-𝑘 table union search returns a ranked list of 𝑘 distinct tables 𝜋𝑘 (𝑞, D) = (𝑇(1) , . ,𝑇(𝑘)) ordered by a unionability score 𝜇: 𝜇(𝑞,𝑇(𝑗)) ≥ 𝜇(𝑞,𝑇(𝑗+1)) for 𝑗 = 1, .

.., k − 1 <ref:2607. p. 0
Where the paper says it

Unlike keywordbased table search or join-oriented discovery, unionability cannot be determined from keyword overlap or shared keys alone, because unionable tables may use different column names, have missing metadata, and share only sparse value overlap. arXiv:2607.

00833#pg2>.

--- Page 4 ---

2.2 Generative Retrieval In this paper we study a generative formulation of table union search. p. 2

Where the paper says it

We partition the annotated pairs G ⊆ Q × D into a training split Gtrain ⊆ Qtrain × D and a test split Gtest ⊆ Qtest × D, where Qtrain ∩ Qtest = ∅; all annotation-dependent offline training signals use only Gtrain. 2 Generative Retrieval In this paper we study a generative formulation of table union search.

Each candidate table Ti ∈ D is assigned a discrete semantic identifier si = (si,1,. p. 2
Where the paper says it

2 Generative Retrieval In this paper we study a generative formulation of table union search. Each candidate table 𝑇𝑖 ∈ D is assigned a discrete semantic identifier 𝑠𝑖 = (𝑠𝑖,1, .

.., si,li), with tokens drawn from a finite code vocabulary In our implementation, si consists of an L-token residual-quantized base code, optionally followed by a suffix token for collision resolution Let SD = T i Ti ∈ D be the set of valid identifiers under the current candidate data lake, and Φ: SD → D the bijective mapping from identifiers back to tables <ref:2607. p. 2
Where the paper says it

In our implementation, 𝑠𝑖 consists of an 𝐿-token residual-quantized base code, optionally followed by a suffix token for collision resolution. Let SD = {𝑠𝑖 | 𝑇𝑖 ∈ D} be the set of valid identifiers under the current candidate data lake, and 𝜙 : SD → D the bijective mapping from identifiers back to tables.

00833#pg2>.

--- Page 5 ---

5.1 Table Representation Following the metadata-agnostic table-centric design of TACTUS [38], GenTUS first derives contextualized column representations from serialized cell values and then attention-pools them into a table embedding <ref:2607. p. 4

Where the paper says it

4 train a generator that maps query tables to identifiers of unionable candidates. 1 Table Representation Following the metadata-agnostic table-centric design of TACTUS [38], GenTUS first derives contextualized column representations from serialized cell values and then attention-pools them into a table embedding.

00833#pg2>. The table-level representation is then projected and normalized as the table embedding <ref:2607. p. 4
Where the paper says it

A global query q𝜂 attention-pools the column representations into a table-level representation: h tab 𝑇 = MultiHeadAttn𝜂 (q𝜂, H col 𝑇 , H col 𝑇 ). The table-level representation is then projected and normalized as the table embedding: e𝑇 = 𝑔𝜂 (h tab 𝑇 ) ∥𝑔𝜂 (h tab 𝑇 ) ∥2 .

00833#pg2>. This continuous table embedding eT is then quantized into a discrete semantic identifier in the next step <ref:2607. p. 4
Where the paper says it

The table-level representation is then projected and normalized as the table embedding: e𝑇 = 𝑔𝜂 (h tab 𝑇 ) ∥𝑔𝜂 (h tab 𝑇 ) ∥2 . This continuous table embedding e𝑇 is then quantized into a discrete semantic identifier in the next step.

00833#pg2>.

--- Page 6 ---

5.2 Semantic Identifier Construction Design goals GenTUS represents each candidate table with a semantic identifier so that table retrieval can be formulated as sequence generation over a finite identifier space <ref:2607. p. 4

Where the paper says it

2 Semantic Identifier Construction Design goals. GenTUS represents each candidate table with a semantic identifier so that table retrieval can be formulated as sequence generation over a finite identifier space.

p. 4
Where the paper says it

2 Semantic Identifier Construction Design goals. GenTUS represents each candidate table with a semantic identifier so that table retrieval can be formulated as sequence generation over a finite identifier space.

00833#pg2>. To this end, GenTUS constructs identifiers with residual quantization <ref:2607. p. 4
Where the paper says it

A useful identifier should satisfy two requirements: it should be compact enough to be generated reliably by an autoregressive decoder, and it should preserve the semantic structure needed for TUS so that unionable tables are placed close together in the identifier space. To this end, GenTUS constructs identifiers with residual quantization.

p. 4
Where the paper says it

A useful identifier should satisfy two requirements: it should be compact enough to be generated reliably by an autoregressive decoder, and it should preserve the semantic structure needed for TUS so that unionable tables are placed close together in the identifier space. To this end, GenTUS constructs identifiers with residual quantization.

00833#pg2>. However, GenTUS augments residual quantization with a unionability-aware contrastive objective, which pulls annotated unionable tables together in the quantizer latent space before discretization <ref:2607. p. 4
Where the paper says it

However, this quantization process by itself does not explicitly optimize identifiers for unionability. GenTUS therefore augments residual quantization with a unionability-aware contrastive objective, which pulls annotated unionable tables together in the quantizer latent space before discretization.

p. 4
Where the paper says it

However, this quantization process by itself does not explicitly optimize identifiers for unionability. GenTUS therefore augments residual quantization with a unionability-aware contrastive objective, which pulls annotated unionable tables together in the quantizer latent space before discretization.

00833#pg2>.

--- Page 7 ---

5.4 Generator Training Given the identifier-level records constructed in Section 5. p. 5

Where the paper says it

The resulting triples form Rpairwise. 4 Generator Training Given the identifier-level records constructed in Section 5.

p. 5
Where the paper says it

The resulting triples form Rpairwise. 4 Generator Training Given the identifier-level records constructed in Section 5.

p. 5
Where the paper says it

The resulting triples form Rpairwise. 4 Generator Training Given the identifier-level records constructed in Section 5.

p. 5
Where the paper says it

The resulting triples form Rpairwise. 4 Generator Training Given the identifier-level records constructed in Section 5.

p. 5
Where the paper says it

The resulting triples form Rpairwise. 4 Generator Training Given the identifier-level records constructed in Section 5.

p. 5
Where the paper says it

The resulting triples form Rpairwise. 4 Generator Training Given the identifier-level records constructed in Section 5.

3, GenTUS trains a retrieval generator to map serialized query tables to semantic identifiers of unionable candidate tables <ref:2607. p. 5
Where the paper says it

4 Generator Training Given the identifier-level records constructed in Section 5. 3, GenTUS trains a retrieval generator to map serialized query tables to semantic identifiers of unionable candidate tables.

p. 5
Where the paper says it

4 Generator Training Given the identifier-level records constructed in Section 5. 3, GenTUS trains a retrieval generator to map serialized query tables to semantic identifiers of unionable candidate tables.

p. 5
Where the paper says it

4 Generator Training Given the identifier-level records constructed in Section 5. 3, GenTUS trains a retrieval generator to map serialized query tables to semantic identifiers of unionable candidate tables.

p. 5
Where the paper says it

4 Generator Training Given the identifier-level records constructed in Section 5. 3, GenTUS trains a retrieval generator to map serialized query tables to semantic identifiers of unionable candidate tables.

p. 5
Where the paper says it

4 Generator Training Given the identifier-level records constructed in Section 5. 3, GenTUS trains a retrieval generator to map serialized query tables to semantic identifiers of unionable candidate tables.

p. 5
Where the paper says it

4 Generator Training Given the identifier-level records constructed in Section 5. 3, GenTUS trains a retrieval generator to map serialized query tables to semantic identifiers of unionable candidate tables.

00833#pg2>. The full generator objective is Lgen = Lce + λr Lpairwise <ref:2607. p. 6
Where the paper says it

This loss is applied to hard-negative ranking records in Rpairwise and calibrates the generator so that positive identifiers receive higher average sequence log-likelihood than hard-negative identifiers. The full generator objective is Lgen = Lce + 𝜆𝑟 Lpairwise.

00833#pg2>.

--- Page 8 ---

5.5 Training Instance Construction GenTUS constructs positive generation records and augments them with structure-preserving query views, so the generator learns to output identifiers of annotated unionable candidates under incidental layout variations <ref:2607. p. 5

Where the paper says it

3 Training Instance Construction After semantic identifiers are assigned, GenTUS converts annotated query–candidate unionability pairs into identifier-level supervision for the retrieval generator. This subsection constructs positive generation records and augments them with structure-preserving query views, so the generator learns to output identifiers of annotated unionable candidates under incidental layout variations.

p. 5
Where the paper says it

3 Training Instance Construction After semantic identifiers are assigned, GenTUS converts annotated query–candidate unionability pairs into identifier-level supervision for the retrieval generator. This subsection constructs positive generation records and augments them with structure-preserving query views, so the generator learns to output identifiers of annotated unionable candidates under incidental layout variations.

00833#pg2>. It also builds indexing records that ground table contents to identifiers and hard-negative ranking records that provide pairwise ranking supervision <ref:2607. p. 5
Where the paper says it

This subsection constructs positive generation records and augments them with structure-preserving query views, so the generator learns to output identifiers of annotated unionable candidates under incidental layout variations. It also builds indexing records that ground table contents to identifiers and hard-negative ranking records that provide pairwise ranking supervision.

p. 5
Where the paper says it

This subsection constructs positive generation records and augments them with structure-preserving query views, so the generator learns to output identifiers of annotated unionable candidates under incidental layout variations. It also builds indexing records that ground table contents to identifiers and hard-negative ranking records that provide pairwise ranking supervision.

00833#pg2>.

--- Page 9 ---

5.4 Generator Training Given the identifier-level records constructed in Section 5.3, GenTUS trains a retrieval generator to map serialized query tables to semantic identifiers of unionable candidate tables <ref:2607.00833#pg2>. The training records provide two complementary supervision signals <ref:2607. p. 5

Where the paper says it

3, GenTUS trains a retrieval generator to map serialized query tables to semantic identifiers of unionable candidate tables. The training records provide two complementary supervision signals.

p. 5
Where the paper says it

3, GenTUS trains a retrieval generator to map serialized query tables to semantic identifiers of unionable candidate tables. The training records provide two complementary supervision signals.

p. 5
Where the paper says it

3, GenTUS trains a retrieval generator to map serialized query tables to semantic identifiers of unionable candidate tables. The training records provide two complementary supervision signals.

p. 5
Where the paper says it

3, GenTUS trains a retrieval generator to map serialized query tables to semantic identifiers of unionable candidate tables. The training records provide two complementary supervision signals.

p. 5
Where the paper says it

3, GenTUS trains a retrieval generator to map serialized query tables to semantic identifiers of unionable candidate tables. The training records provide two complementary supervision signals.

00833#pg2>.

--- Page 10 ---

5.4 Generator Training Given the identifier-level records constructed in Section 5.3, GenTUS trains a retrieval generator to map serialized query tables to semantic identifiers of unionable candidate tables <ref:2607.00833#pg2>. The training records provide two complementary supervision signals <ref:2607.00833#pg2>.

--- Page 11 ---

5.4 Generator Training Given the identifier-level records constructed in Section 5.3, GenTUS trains a retrieval generator to map serialized query tables to semantic identifiers of unionable candidate tables <ref:2607.00833#pg2>. The training records provide two complementary supervision signals <ref:2607.00833#pg2>.

--- Page 12 ---

5.4 Generator Training Given the identifier-level records constructed in Section 5.3, GenTUS trains a retrieval generator to map serialized query tables to semantic identifiers of unionable candidate tables <ref:2607.00833#pg2>. The training records provide two complementary supervision signals <ref:2607.00833#pg2>.

--- Page 14 ---

Table Preprocessing Candidate Tables from Data Lake Query Table Table-to-Sentence Serialization Query Table Sentence Candidate Table Sentences Sampled Values 18 9364 7263 9971… Amy Sam Tom Luka … … Serialized Sentence [COL][VAL] 2014-05-01 … [COL][VAL] 18 9364 7263 … [COL][VAL] Amy Sam Tom … tbl air schedule → [COL] [VAL] 2014-05-01 … [COL][VAL] 18 … tbl passenger counts → [COL] [VAL] 52 … [COL][VAL] Ship … query table → [COL][VAL] Boston Chicago New York …

--- Page 13 ---

2 Offline Processing 2A Semantic ID Construction candidate Residual Quantization

--- Page N ---

5.2 Semantic Identifier Construction Design goals GenTUS represents each candidate table with a semantic identifier so that table retrieval can be formulated as sequence generation over a finite identifier space <ref:2607.00833#pg2>. To this end, GenTUS constructs identifiers with residual quantization <ref:2607.00833#pg2>. However, GenTUS augments residual quantization with a unionability-aware contrastive objective, which pulls annotated unionable tables together in the quantizer latent space before discretization <ref:2607.00833#pg2>.

--- Page N ---

5.4 Generator Training Given the identifier-level records constructed in Section 5.3, GenTUS trains a retrieval generator to map serialized query tables to semantic identifiers of unionable candidate tables <ref:2607.00833#pg2>. The training records provide two complementary supervision signals <ref:2607.00833#pg2>.

--- Page N ---

5.5 Training Instance Construction GenTUS constructs positive generation records and augments them with structure-preserving query views, so the generator learns to output identifiers of annotated unionable candidates under incidental layout variations <ref:2607.00833#pg2>. It also builds indexing records that ground table contents to identifiers and hard-negative ranking records that provide pairwise ranking supervision <ref:2607.00833#pg2>.

Green — the paper says it, and the page is printed. Red — no passage carries it, so treat that sentence as unverified. The same check runs on your own text: check a summary against its source, $5.

Improvements for AI systems

  1. Unified Search Framework

The improved system can reformulate table union search as constrained generation over discrete semantic table identifiers, eliminating the need for vector indexes, ANN search, or reranking models. This allows the system to retrieve tables directly from a generator conditioned on the query.

  1. Unionability-Aware Identifier Construction

The system incorporates a unionability-aware contrastive objective during training that pulls annotated unionable tables together in the quantizer latent space before discretization, ensuring that unionable tables receive identifiers with shared coarse codes.

  1. Constrained Decoding for Validity Guarantee

The system uses a prefix tree to enforce prefix constraints during decoding, which prevents invalid identifiers from consuming beam-search capacity and guarantees that every generated identifier corresponds to an existing table in the data lake.

Abstract

Modern data lakes contain heterogeneous tables, and the information needed for a task is often scattered across many tables with different schemas, sources, and naming conventions. Table union search (TUS) retrieves tables whose rows are semantically compatible with a query table for union, supporting data discovery, enrichment, and downstream analytics. Existing TUS methods follow an encode-search-refine pipeline: they embed tables or columns, retrieve a candidate pool from a vector index, and refine it by query-candidate matching or reranking. As a result, a unionable table missed by candidate retrieval cannot be recovered later, and searching and refining a pool for every query adds latency and storage costs that grow with the data lake. We propose GenTUS, a generative retrieval framework that reformulates TUS as constrained generation over discrete semantic table identifiers. GenTUS assigns each data-lake table a compact identifier learned under pairwise unionability supervision, which encourages unionable tables to share identifier prefixes, and trains a generator to produce the identifiers of unionable tables directly from the query table. At query time, prefix-constrained beam search guarantees that every generated identifier maps to an existing table, and its output order serves directly as the ranking, with no candidate pool to search or rerank. Experiments on seven public TUS benchmarks show that GenTUS achieves the best overall retrieval quality, with the largest gains on the larger and more diverse benchmarks, while requiring substantially lower online retrieval time and retrieval-artifact storage than prior TUS methods.

Sources

Related papers