Table-as-Search (TaS): Structured Table Retrieval

• Table-as-Search is a search paradigm that treats tables as primary objects, preserving rows, columns, joins, and other structural signals to enhance semantic relevance.
• Research in TaS advances from probabilistic field decomposition to transformer-based methods, integrating dense semantic vectors with sparse lexical cues to boost retrieval metrics.
• The paradigm underpins applications in dataset search, table union, and long-horizon table completion, addressing challenges like metadata sparsity and token-budget constraints.

Searching arXiv for papers on “Table-as-Search” and closely related table retrieval formulations. Table-as-Search (TaS) denotes a family of formulations in which tables, or table-derived structures, are treated as the primary search objects rather than as incidental context for document retrieval or SQL execution. In current usage, the term covers at least two closely related lines of work: retrieval over corpora or data lakes where the output is one or more relevant tables, and a structured planning framework that reformulates long-horizon information seeking as table completion (Boutaleb et al., 17 Nov 2025, Terrenzi et al., 18 May 2026, Lan et al., 6 Feb 2026). Across both lines, the central design move is to preserve structure—fields, rows, columns, entities, joins, alignments, or pending cells—so that search can optimize semantic relevance together with relational organization.

1. Conceptual scope and research lineage

Early table search work treated a table as a structured object with heterogeneous evidence sources rather than as a flat text fragment. In scientific table search, MaitreD decomposed each table into document-level fields, table-level fields, and cell-level fields, then ranked by a probabilistic framework of the form $p(t \mid q) \propto p(t)\,p(q \mid t)$, where the prior $p(t)$ favored numeric-content-rich tables (Gao et al., 2017). The same paper also introduced TableArXiv, built from the arXiv January 2014 snapshot in Physics and related categories, with 341,573 extracted tables and 105 expert-created topics (Gao et al., 2017).

Neural retrieval work shifted the emphasis from explicit field weighting to learned representation. BERT-based ad hoc table retrieval adapted structured tables to a token-limited encoder by selecting rows, columns, or cells and concatenating them with table context fields; on WikiTables, “Hybrid-BERT-Row-Max” reached MAP $=0.6311$, MRR $=0.6673$, and NDCG@10 $=0.6519$, exceeding STR and simpler BERT baselines (Chen et al., 2020). StruBERT then fused textual fields and tabular structure by combining row-based and column-based linearizations with horizontal and vertical self-attention, and improved content-based retrieval on Query-by-Example from TaBERT’s NDCG@5 $=0.5877$ to $0.6345$ (Trabelsi et al., 2022).

More recent work has made the field heterogeneity of tables explicit again, but within neural and hybrid frameworks. THYME argued that titles benefit from coarse-grained semantic matching, whereas headers and cells often require fine-grained lexical matching, and implemented this with a dense semantic score plus a field-aware sparse lexical score controlled by a Mixture-of-Field-Experts gate (Li et al., 4 Mar 2025). PIPER, EasyTUS, TACTUS, GenTUS, and related systems extended TaS from web-style table retrieval to large, metadata-poor data lakes in which the dominant tasks are dataset search, union search, join-aware composition, and multi-table retrieval (Terrenzi et al., 18 May 2026, Otto, 4 Nov 2025, Sun et al., 18 Mar 2026, Zhang et al., 1 Jul 2026, Boutaleb et al., 17 Nov 2025).

Taken together, these works indicate that TaS is not a single retrieval model but a design paradigm whose defining property is that table structure remains operational throughout ranking, candidate generation, or planning.

2. Query modalities and formal problem statements

The TaS literature spans several distinct query modalities and output spaces.

Task family	Query object	Representative formulations
Keyword or natural-language table retrieval	Text query	Scientific table search (Gao et al., 2017), PIPER dataset search (Terrenzi et al., 18 May 2026)
Related-table or query-by-example retrieval	Table query	OGDC relatedness tasks (Glass et al., 2023), StruBERT content-based retrieval (Trabelsi et al., 2022)
Table union search	Query table	EasyTUS (Otto, 4 Nov 2025), TACTUS (Sun et al., 18 Mar 2026), GenTUS (Zhang et al., 1 Jul 2026)
Multi-table retrieval	Natural-language query	Iterative join-aware retrieval for NL2SQL (Boutaleb et al., 17 Nov 2025)
Novel table search	Query table plus unionable candidates	ANTs and NTS variants (Kassaie et al., 7 Mar 2026)
Long-horizon information seeking as table completion	Natural-language query	TaS planner over external table state (Lan et al., 6 Feb 2026)

In keyword-style settings, the query is free text and the output is a ranked list of tables whose textual, structural, or content-derived features are relevant. MaitreD modeled this as structured probabilistic retrieval over eight fields, while PIPER defined the TaS task as ranking tables whose content and structure make them relevant to a free-form natural-language query for downstream analysis, integration, or exploration (Gao et al., 2017, Terrenzi et al., 18 May 2026).

In table-to-table settings, the query itself is a table. OGDC formalized three notions of relatedness—same dataset, same organization, and weighted tag overlap—and provided both related-table search and keyword-search benchmarks over roughly $195{,}448$ distinct tables collected from seven CKAN portals (Glass et al., 2023). In table union search, the objective is to find top-$k$ tables whose union with a query table produces a semantically consistent extension of rows; this is the task addressed by EasyTUS, TACTUS, and GenTUS (Otto, 4 Nov 2025, Sun et al., 18 Mar 2026, Zhang et al., 1 Jul 2026).

Composition-oriented TaS generalizes beyond returning a single table. In “Exploring Multi-Table Retrieval Through Iterative Search,” the input is a natural-language query $Q$ and a datalake of tables $p(t)$0, and the output is a small set $p(t)$1, $p(t)$2, whose joint content can answer $p(t)$3. The paper defines a joint score that balances semantic relevance, coverage of query sub-concepts, and structural coherence via joinability (Boutaleb et al., 17 Nov 2025). Novel Table Search introduces a second-stage ranking problem over unionable tables, seeking tables that contribute new information relative to the query table rather than near-duplicates of it (Kassaie et al., 7 Mar 2026).

A separate but increasingly cited use of the term redefines search itself as a table-completion process. In that framework, a query is mapped to a schema $p(t)$4, rows correspond to candidate entities, columns correspond to constraints or information to be collected, filled cells record evidence, and empty cells encode an explicit search plan (Lan et al., 6 Feb 2026). This suggests that “Table-as-Search” now names both a retrieval paradigm and a state-management paradigm.

3. Representation and ranking architectures

A recurrent issue in TaS is how much of a table to encode and at what granularity. The earliest probabilistic systems handled this by field decomposition and smoothing. MaitreD assigned different emphasis to article title, abstract, caption, referring sentences, footnotes, column headers, row headers, and cell values, and additionally used entity extraction and quantity-type expansion via QUDT units to bridge query–table mismatch (Gao et al., 2017).

Transformer-based retrievers initially confronted the token-budget problem directly. “Table Search Using a Deep Contextualized LLM” proposed a content selector that ranks rows, columns, or cells by salience relative to the query using fastText-based similarity, after which only the top-ranked items are serialized into BERT together with caption, section title, page title, and header row (Chen et al., 2020). The reported ablations found row-based selection superior to column-based or cell-based alternatives, and max-salience strongest for rows (Chen et al., 2020).

Structure-aware encoders then made row and column organization part of the representation itself. StruBERT linearized each table into column-based sequences and row-based sequences, pooled BERT token states into cell embeddings, applied vertical and horizontal self-attention, and compared tables through both coarse-grained $p(t)$5-style signals and fine-grained row/column interactions computed with a lightweight miniBERT layer (Trabelsi et al., 2022). THYME followed a different route: a shared Transformer produces both a dense semantic vector and sparse lexical logits, title pooling uses max aggregation, header and cell pooling are field-specific, and a gating network dynamically weights title, header, and cell evidence before the final hybrid score $p(t)$6 at inference (Li et al., 4 Mar 2025).

Data-lake systems increasingly avoid publisher metadata altogether. PIPER profiles each table statistically, generating per-column features such as inferred datatype, number of distinct values, fraction of missing entries, coverage span, and type-specific statistics; an LLM then turns the serialized profile into multiple pseudoqueries, which are embedded and indexed for dense retrieval (Terrenzi et al., 18 May 2026). TUNES takes an entity-oriented route: it identifies primary keys and attributes, instantiates a Neo4j property graph, combines BM25 over entity documents with semantic search using bge-m3 embeddings, and augments the retrieved set with entities returned by LLM-generated Cypher queries (Nguyen et al., 23 Aug 2025).

A plausible implication is that TaS representation learning has diverged into three stable families: field-aware text ranking, holistic table embedding, and structure-preserving entity or graph indexing. The literature does not treat these families as mutually exclusive; several systems combine them.

4. Composition-oriented search in data lakes

Data-lake TaS places particular emphasis on compositionality: whether retrieved tables can be unioned, joined, or otherwise combined into a coherent answer.

In Table Union Search, EasyTUS implements a three-step pipeline of table serialization, table representation, and vector search. It uses a zero-shot, value-only serialization strategy with random row sampling, mean-pooled LLM embeddings, and HNSW-based approximate nearest-neighbor retrieval by cosine similarity (Otto, 4 Nov 2025). TACTUS argues that prior TUS methods are overly column-centric and instead learns one embedding per table to capture holistic unionability, using positive table pair construction, latent-positive exclusion, hard-negative mining, attentive table encoding, adaptive candidate retrieval, and dual-evidence reranking that combines table-level and column-level signals (Sun et al., 18 Mar 2026).

GenTUS departs from the standard encode-search-rerank architecture altogether. It assigns each table a unionability-aware semantic identifier by residual quantization of learned table embeddings, trains an encoder–decoder model to generate the identifiers of unionable tables directly from a serialized query table, and enforces validity at inference through constrained decoding over a prefix trie of known identifiers (Zhang et al., 1 Jul 2026). Novel Table Search adds yet another objective: among already unionable tables, prefer those that are syntactically novel. ANTs approximates novelty through attribute-level scores that combine low syntactic similarity with semantic similarity, and ranks candidate tables by $p(t)$7 (Kassaie et al., 7 Mar 2026).

Join-oriented composition introduces a different constraint. The iterative multi-table retrieval framework for open-domain question answering over data lakes defines a joint objective over relevance, coverage, and joinability and explicitly treats retrieval as an iterative search process. Its concrete instantiation, Greedy Join-Aware Retrieval, uses pre-computed coarse relevance $p(t)$8, fine-grained relevance matrix $p(t)$9, and join compatibility $=0.6311$0, with a utility that sums coarse, coverage, and join gains using weights $=0.6311$1, $=0.6311$2, and $=0.6311$3 (Boutaleb et al., 17 Nov 2025).

These systems share a common departure from classical document retrieval: admissible results are not merely relevant in isolation. They must also support a downstream table operation—union, join, augmentation, or novelty-aware diversification.

5. Benchmarks and empirical findings

Benchmark design has been central to TaS because evaluation depends strongly on the task definition. TableArXiv targeted scientific keyword search with graded relevance labels over 105 topics (Gao et al., 2017). OGDC scaled evaluation to approximately $=0.6311$4 distinct open-government tables and defined automatic labels for related-table search and keyword search via metadata such as dataset, organization, and tags (Glass et al., 2023). Query-by-Example, WikiTables, PMC, NQ-TABLES, OTT-QA, FetaQA, NTCIR-15, SPIDER, BIRD, FIBEN, BEAVER-DW, BEAVER-NW, TUSBench, WikiUnion, LakeBench, GAIA, BrowseComp-ZH, and WideSearch subsequently extended TaS evaluation across web tables, data lakes, and long-horizon search settings (Trabelsi et al., 2022, Li et al., 4 Mar 2025, Terrenzi et al., 18 May 2026, Boutaleb et al., 17 Nov 2025, Otto, 4 Nov 2025, Zhang et al., 1 Jul 2026, Lan et al., 6 Feb 2026).

Reported gains are substantial but task-specific. In scientific table search, full MaitreD improved over BM25F from NDCG@20 $=0.6311$5 to $=0.6311$6, ERR@20 $=0.6311$7 to $=0.6311$8, and MAP@100 $=0.6311$9 to $=0.6673$0 (Gao et al., 2017). On OGDC related-table search over the Canada site, Siamese reached NDCG@20 $=0.6673$1 for same-dataset relatedness, $=0.6673$2 for same-organization relatedness, and $=0.6673$3 for tag-overlap relatedness, all above BM25 (Glass et al., 2023). THYME achieved NQ-TABLES NDCG@5 $=0.6673$4, NDCG@10 $=0.6673$5, and R@1 $=0.6673$6, and OTT-QA NDCG@5 $=0.6673$7, NDCG@10 $=0.6673$8, and R@1 $=0.6673$9, significantly outperforming BIBERT+SPLADE (Li et al., 4 Mar 2025).

For metadata-poor dataset search, PIPER obtained FetaQA R@10 $=0.6519$0 versus $=0.6519$1 for pT+QGpT, while on OTT-QA the metadata-strong setting favored baselines, with PIPER at $=0.6519$2 versus $=0.6519$3 (Terrenzi et al., 18 May 2026). On NTCIR-15 complex natural-language queries, PIPER reported MAP $=0.6519$4, P@10 $=0.6519$5, R@10 $=0.6519$6, and nDCG@10 $=0.6519$7, all above Dense-BGE (Terrenzi et al., 18 May 2026).

In union search, EasyTUS reported average MAP improvements of up to $=0.6519$8, data-preparation speedups up to $=0.6519$9, and query-processing speedups up to $=0.5877$0, while retaining $=0.5877$1 of its accuracy in metadata-absent variants (Otto, 4 Nov 2025). TACTUS achieved MAP@k $=0.5877$2 on SANTOS Small, $=0.5877$3 on TUS Small, $=0.5877$4 on TUS Large, and $=0.5877$5 on WikiUnion, while reducing candidate set size and improving purity (Sun et al., 18 Mar 2026). GenTUS reported the best overall retrieval quality across seven public TUS benchmarks, with average rank $=0.5877$6 compared to $=0.5877$7 for TACTUS, as well as $=0.5877$8 faster online retrieval and $=0.5877$9 smaller storage (Zhang et al., 1 Jul 2026).

In multi-table retrieval for NL2SQL, the iterative join-aware method was competitive with a mixed-integer-programming baseline while being much faster: on SPIDER at $0.6345$0, greedy obtained $0.6345$1 Recall and $0.6345$2 Complete Recall in about $0.6345$3 s versus $0.6345$4 s; on BIRD at $0.6345$5, $0.6345$6 versus $0.6345$7 in $0.6345$8 s versus $0.6345$9 s; and on BEAVER-NW at $195{,}448$0, $195{,}448$1 versus $195{,}448$2 in $195{,}448$3 s versus $195{,}448$4 s (Boutaleb et al., 17 Nov 2025).

The main empirical lesson is not that one architecture dominates universally, but that gains are largest when the retrieval objective aligns tightly with the downstream table operation and with the actual evidence available in the corpus.

6. Table completion as long-horizon search state

A distinct research line uses the phrase “Table-as-Search” for long-horizon agentic information seeking rather than corpus retrieval. In that formulation, a deterministic schema-construction function maps a query $195{,}448$5 to a schema $195{,}448$6, where $195{,}448$7 is a set of candidate entities, $195{,}448$8 is a set of constraint columns, and $195{,}448$9 is a set of information columns; at time $k$0, the system maintains a table $k$1 whose empty cells denote pending search steps (Lan et al., 6 Feb 2026).

The framework unifies three paradigms by schema choice: Deep Search when $k$2, Wide Search when $k$3, and DeepWide Search when both are nonempty (Lan et al., 6 Feb 2026). The planner alternates between expanding rows and populating empty cells, while the external table persists search state that would otherwise be compressed into a fragile plain-text context window (Lan et al., 6 Feb 2026).

The reported results indicate that this state externalization matters. On BrowseComp-ZH, TaS-MA with Gemini-2.5-Flash improved from $k$4 to $k$5, and with GPT-5 Medium-Think from $k$6 to $k$7, relative to ReAct-MA (Lan et al., 6 Feb 2026). On DeepWide Search over 20 e-commerce BD queries, TaS-MA with Claude-S4 reached Col-F1 $k$8 and Item-P $k$9, above Gemini DeepResearch and ReAct-MA (Lan et al., 6 Feb 2026). The authors also reported that the performance gap versus ReAct-MA widens as query complexity increases, and that Planner quality is the primary bottleneck (Lan et al., 6 Feb 2026).

This usage broadens TaS from “search over tables” to “search through a table-shaped state space.” It remains structurally consistent with retrieval-oriented TaS because the table is again the mechanism that organizes evidence and constrains admissible next actions.

7. Limitations, ambiguities, and open problems

Several limitations recur across the literature. Metadata sparsity remains a major problem for data-lake search, motivating systems such as PIPER and EasyTUS that either replace metadata with content-derived profiles or ignore metadata entirely (Terrenzi et al., 18 May 2026, Otto, 4 Nov 2025). Token-budget constraints remain a structural issue for Transformer encoders, and early BERT-based work still relied on fixed salience heuristics that cannot dynamically choose the optimal granularity of rows, columns, or cells per query (Chen et al., 2020). In TUNES, graph-search errors accounted for approximately $Q$0 of search errors, while entity identification failed on highly irregular table layouts in $Q$1 of analyzed cases (Nguyen et al., 23 Aug 2025).

LLM-dependent TaS systems introduce additional concerns: pseudoquery generation and query optimization incur quality and cost dependencies, and synthetic query generation may hallucinate (Terrenzi et al., 18 May 2026). EasyTUS reported that on the purely numeric-heavy ecb_union dataset, D3L slightly outperformed its LLM-based approach, which the paper identifies as a known limitation of pure-LLM methods (Otto, 4 Nov 2025). GenTUS reduces index size and latency, but its own discussion notes that a generative retriever depends on a well-structured identifier space and that beam search may struggle if the codebook does not align with semantics (Zhang et al., 1 Jul 2026).

Complexity-theoretic limits also appear repeatedly. The one-shot mixed-integer formulation for multi-table retrieval has $Q$2 worst-case complexity, motivating greedy iterative approximations (Boutaleb et al., 17 Nov 2025). Novel Table Search similarly proves the associated optimization problem NP-hard and therefore turns to penalization-based approximation through ANTs (Kassaie et al., 7 Mar 2026). In the long-horizon planning setting, planner degradation rather than sub-agent degradation produced the largest drops, indicating that search control itself can dominate tool quality (Lan et al., 6 Feb 2026).

A persistent ambiguity is terminological. “Table-as-Search” now refers both to retrieval of tables as indexed objects and to the reformulation of search as table completion. This suggests a broader unifying interpretation: TaS is any search system in which table structure is not merely input formatting but the principal computational substrate for ranking, composition, or control.