acceptodds
Under review as a conference paper at ICLR 2027

HCVR: Hypothesize, Compose, and Verify for Precise Multi-Table Retrieval

Abstract

Text-to-SQL is increasingly deployed over a *pool* of tables, a data-lake-style repository that consolidates many source systems without database boundaries or declared foreign keys. Answering a query then starts with retrieval, which must find both the tables that hold the requested information and the tables needed to join them. Pooled-table retrievers to date rank candidates by how well each table matches the query, and a match alone cannot tell whether a table adds information or a connection that the tables already selected do not. We instead ask what role each table plays in answering the query: a table is a *provider* if it supplies one information requirement, a *connector* if it joins providers that would otherwise remain unjoined, and is excluded if it serves neither role. Under this criterion, competing providers of the same requirement and competing connectors of the same providers are resolved rather than all retained. We propose HCVR (**H**ypothesize, **C**ompose, and **V**erify for multi-table **R**etrieval), a training-free framework that treats every candidate as a hypothesis about its role. HCVR parses the query into typed requirements and retrieves provider candidates for each; it composes providers by bounded beam search and adds the connectors between them along minimum-hop paths in an inferred join graph; and it declares a role absent only after cross-verifying the composed set against an independent LLM selection. The whole procedure takes two to three LLM calls per query. On pooled BIRD, Spider, and MMQA with two retrieval LLMs, HCVR exceeds the best baseline in every setting by 10.7 to 31.2 precision points and 8.0 to 25.5 F1 points, while the fraction of queries whose gold tables are all retrieved stays within 3.9 points of the best baseline. With the retrieved tables frozen, the execution accuracy of two downstream SQL-generation LLMs improves on every corpus, by up to 11.8 points.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Multi-Table Retrieval, Text-to-SQL, Information Retrieval, Large Language Models

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_RfgjOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

HCVR reframes multi-table retrieval from similarity scoring to role assignment (provider/connector/none) through a three-stage pipeline (Hypothesize → Compose → Verify) that requires no training. It achieves large gains in Precision and F1 over all baselines on pooled BIRD/Spider/MMQA and leads downstream execution accuracy.

Critical strengths

S1. The role-based paradigm is a clean and principled departure from existing scoring methods. Instead of asking “how similar is this table to the query?”, HCVR explicitly asks “what job does this table do?” — a distinction that directly addresses the core failure mode of prior work: retaining many relevant-looking but ultimately useless tables.

Critical weaknesses

W1. Perfect Recall (PR) is not improved by HCVR and in fact lags behind CORE-T in several settings​ (e.g., 84.6 vs 87.0 on BIRD/Qwen; 93.0 vs 94.9 on Spider/Qwen, Table 1). For NL2SQL, missing even one gold table makes correct SQL execution impossible, so a method that does not enhance PR limits its downstream benefit. This undermines the paper's central claim that "precise retrieval helps SQL generation without sacrificing coverage" — since coverage (recall) is precisely what is sacrificed in the settings where it matters most.

W2. Methodological novelty is limited. Every component (LLM decomposition, dense retrieval, beam search, LLM verification) is off-the-shelf. The contribution is essentially a workflow design rather than a new algorithm. The paper would benefit from introducing a novel technique at one of the stages.

Questions for authors

Q1. Would HCVR still improve over end-to-end NL2SQL systems with built-in schema linking (e.g., CHESS)?

Q2. The method introduces several hyperparameters (beam width K=10, weight α=0.5, compatibility threshold τ=0.3, max hop h=5). Are the results sensitive to these choices? Was any systematic sensitivity analysis performed?

Q3. Has the paper considered latency or other computational costs? HCVR requires 2.5–2.8 LLM calls per query. How does this overhead affect practical deployability, especially under latency-sensitive scenarios?

Reviewer_d9GAOpenReview ↗
Rating
3
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper addresses multi-table retrieval for text-to-SQL over a pooled corpus of tables in which source-database boundaries and declared foreign keys are unavailable. The authors propose a role-based criterion, under which a table is kept only if it acts as a provider that supplies an information requirement or as a connector that joins providers that would otherwise remain unjoined. Building on this criterion, they introduce HCVR, a training-free framework with three stages. Hypothesize parses the query into typed entity–predicate–attribute requirements and retrieves provider candidates for each. Compose selects providers via bounded beam search and adds connectors along minimum-hop paths in an inferred join graph. Verify reconciles the top composition with an independent LLM selection, invoking a conservative deletion review when needed. HCVR achieves the highest precision and F1 in every setting.

Critical strengths
  • The paper is clearly written and easy to follow.
  • Framing pooled-table retrieval in terms of the roles that tables play in answering a query (provider, connector, or neither) is an interesting and intuitive perspective.
  • The experiments are extensive and well organized.
Critical weaknesses

W1. The contribution of connector inference is not isolated. The paper presents C2 (connector entanglement) as one of its three core challenges and claims that the connector inference in Compose addresses it. Indeed, removing connector inference lowers the final PR by 6.9–10.5 points (Table 11). However, it remains unclear whether this drop actually occurs on queries that require a connector. By the paper's own definition, a connector is a table that joins providers that would otherwise remain unjoined. Connector-only tables can therefore exist only in queries with three or more gold tables, which account for 14.3% of of all queries pooled across the three corpora (6.4% on Spider; Table 9). If the authors' interpretation holds, the PR drop from removing connector inference should be concentrated in queries with ≥3 gold tables and should be negligible for single-table queries. On Spider, however, even if HCVR lost every query with ≥3 gold tables that it currently retrieves completely, PR would drop by at most about 5.6 points. The observed drop is 8.9 points, while PR_U before verification remains unchanged. This suggests that part of the drop may stem not from the connector role itself but from other pathways, such as changes in P₁ or in the candidate union U and the resulting changes in Verify's branching. Table 9 only compares HCVR with CORE-T by the number of gold tables, and Table 11 reports the ablation over all queries, so neither allows the contribution of connector inference to be assessed. I would ask the authors to report the composition ablations, including w/o connector inference, broken down by the number of gold tables (1 / 2 / ≥3).

W2. The independence of Verify is weaker than claimed. In the Introduction, when motivating C3, the paper argues that rechecking a selection with the evidence that produced it cannot expose that evidence's own errors, and thus a second judgement must rest on evidence the first did not use. In practice, however, Verify functions less as an independent check on the errors of the first judgement and more as a re-selection within the candidate space that the first judgement has already defined. The candidate union U from which SELECT chooses is itself the output of Hypothesize and Compose, and ADD likewise adds tables only from within U.

W3. The evaluation setting does not clearly reflect the motivating scenario. The paper is motivated by a data-lake setting in which tables from multiple source systems are pooled without database boundaries or foreign keys. However, the evaluation pools are constructed by merging the databases of BIRD (11 DBs, 75 tables) and Spider (20 DBs, 81 tables), and MMQA (710 tables) is itself built on the Spider databases. In this setup, it is unclear how many duplicate or near-duplicate tables providing the same information actually exist, and the paper does not report how often competing providers occur.

Questions for authors
  • Could the authors report the composition ablations in Tables 3 and 11, particularly w/o connector inference, broken down by the number of gold tables (1 / 2 / ≥3) with Llama-3.1-8B-Instruct? In particular, how large is the PR drop on single-table queries, where a connector-only table cannot exist by definition?

  • How many of HCVR’s final perfect-recall failures are due to gold tables missing from U, and how many are due to Verify’s decisions within U?

Reviewer_hjBNOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes a training-free retriever for text-to-SQL over pooled tables without database boundaries or declared foreign keys. HCVR extracts typed information requirements, retrieves candidate providers for each requirement, composes alternative table sets with beam search and inferred connector paths, and reconciles the top composition with an LLM selection through conditional deletion review. Experiments on pooled BIRD, Spider, and MMQA with two retrieval LLMs show higher precision and F1 than the reported baselines, while downstream evaluation with two SQL generators demonstrates consistent execution-accuracy improvements.

Critical strengths

Substantial and consistent improvements on the evaluated retrieval task.

Critical weaknesses
  1. The novelty is limited. Prior work already models complementary information and bridging tables.

  2. The semantic selector does not see the composition ranking, but both views are restricted to the same retained union U and use the same compatibility evidence. Consequently, their agreement cannot detect a necessary table excluded from U.

Questions for authors
  1. How does HCVR compare with other methods (such as the greedy join-aware method)? This would help establish what explicit role assignments contribute beyond coverage-aware search.

  2. How often do P1 and R share an incorrect omission caused by candidate construction or join inference? Please distinguish gold tables absent from U from erroneous deletions within U, and report whether the mechanism handles the key-domain-collision example above.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.