acceptodds
Under review as a conference paper at ICLR 2027

SQuAR: Synthetic Queries as Join-Aware Retrieval Anchors for Text-to-SQL

Abstract

Schema retrieval is critical in text-to-SQL systems over large database collections, where a system must identify both the target database and the tables needed to answer a natural-language question. Existing approaches either rely on costly query-time LLM reasoning or struggle to preserve multi-table structural relationships as the schema pool scales. To address this limitation, we present SQuAR, an open-pool schema retrieval approach that uses synthetic questions as join-aware retrieval anchors. Offline, SQuAR samples multi-table sub-schemas along foreign-key paths and indexes schema-grounded synthetic questions as retrieval anchors. At query time, it aggregates database-level and table-level evidence from retrieved anchors without query-time LLM calls. We further show that table-level recall alone is insufficient, as missing even one required table or selecting an incorrect database makes the retrieved schema insufficient to construct the target SQL query. Thus, we introduce Execution-Ready Coverage (ERC), which requires the correct database and all required tables, and Execution-Ready Jaccard (ERJ), which additionally penalizes over-retrieval. We evaluate SQuAR on four benchmarks and their unified pool of 37 databases and 619 tables. Across the four benchmarks, SQuAR improves Table Recall and ERC by up to 17.5 and 26.1 percentage points over the strongest evaluated baselines, while also maintaining a median retrieval latency of approximately 0.1 seconds. In the unified open pool, SQuAR improves DB R@1, Table Recall, and ERC by 15.5, 8.7, and 20.1 percentage points, respectively.

Primary area:
generative models
Keywords:
Text-to-SQL, Schema Retrieval, Large Language Models, Synthetic Data, Retrieval-Augmented Generation

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_aB58OpenReview ↗
Rating
2
Confidence
4
Informativeness
1
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes SQuAR, a schema retrieval method that identifies the target database and the tables required to answer a question in an open pool of multiple databases. In the offline stage, SQuAR samples multi-table sub-schemas along foreign-key paths, generates synthetic questions for each sub-schema, and indexes them as retrieval anchors. At query time, it retrieves anchors similar to the user question and aggregates their associated databases and tables through voting, without any LLM calls. The authors also argue that table recall alone cannot tell whether the retrieved schema is sufficient to construct the target SQL, and they propose two metrics, ERC and ERJ. Across four benchmarks, SQuAR improves Table Recall and ERC over the baselines.

Critical strengths
  • The paper is clearly written and easy to follow.
  • The motivation is convincing: missing even a single required table makes the correct SQL impossible to generate.
  • The experiments are well organized around explicit research questions.
Critical weaknesses
  • The core idea of SQuAR is to generate synthetic questions that cover diverse table combinations and to use them as retrieval anchors so that all required tables are recovered. However, the offline pipeline closely follows DBCopilot [1]. DBCopilot already (i) constructs a schema graph in which a virtual root node connects to database nodes, which in turn connect to their tables linked by key relationships, (ii) samples sub-schemas via random walks starting from the root node, and (iii) generates a pseudo-question for each sampled sub-schema in a schema-to-question manner. SQuAR reproduces these three steps. The main methodological difference lies in how the generated questions are used: DBCopilot trains a generative router on them, whereas SQuAR indexes them for kNN retrieval followed by frequency voting. Moreover, the base models differ substantially. DBCopilot uses T5-base for both its schema-to-question generator and its router, whereas SQuAR uses Qwen2.5-14B-Instruct for question generation and Qwen3-Embedding-4B for retrieval. It is unclear whether the gains over DBCopilot stem from the proposed anchor-based retrieval or simply from stronger base models.

  • Using generated questions to bridge the gap between user queries and retrieval targets is well established in information retrieval. doc2query [2] expands documents with predicted queries, and PAQ [3] goes further: its retriever, RePAQ, directly retrieves the most similar generated question and returns the attached answer. SQuAR’s online stage closely mirrors this paradigm, with table sets in place of answers. In text-to-SQL, the Contextual RAG component of CSR-RAG [4] likewise retrieves the top-k most similar questions from a collection of question–SQL pairs and returns the union of their associated table sets. None of these works are cited or discussed. Furthermore, frequency-based voting over the retrieved anchors is hard to regard as a separate technical contribution.

  • The paper states that an execution-ready schema should contain all required tables while excluding unnecessary ones (Challenge 2). However, SQuAR returns the top-K tables by vote count for a fixed, pre-set K. Apart from removing tables outside the predicted database, it has no query-adaptive mechanism for deciding how many tables to return. Indeed, SQuAR’s ERJ is lower than that of CORE-T [5] on both Spider and BIRD at every K.


[1] Wang et al., DBCopilot: Natural Language Querying over Massive Databases via Schema Routing, EDBT, 2025.

[2] Nogueira et al., Document Expansion by Query Prediction, arXiv, 2019.

[3] Lewis et al., PAQ: 65 Million Probably-Asked Questions and What You Can Do With Them, TACL, 2021.

[4] Singh et al., CSR-RAG: An Efficient Retrieval System for Text-to-SQL on the Enterprise Scale, arXiv, 2026.

[5] Soliman et al., CORE-T: COherent REtrieval of Tables for Text-to-SQL, arXiv, 2026.

Questions for authors
  1. Have you compared SQuAR with DBCopilot when both use the same synthetic questions? For example, DBCopilot's router could be trained on the 10K questions generated by Qwen2.5-14B-Instruct, or SQuAR's index could be built from questions produced by DBCopilot's question generator. Such a comparison would clarify whether the gains come from the anchor-based retrieval or from the stronger base models.

  2. How does SQuAR compare with a simple baseline that uses the same embedder (Qwen3-Embedding-4B)? One such baseline would densely retrieve tables using their descriptions and columns, and optionally expand the result to neighboring tables via foreign keys. What is the net gain attributable to the anchor-based approach itself?

  3. Are there cases where ERC leads to a substantially different conclusion from the existing complete Recall metric? If not, what justifies proposing ERC as a new metric?

6 comments
AuthorsOpenReview ↗

Response to Critical Weakness W1

Response to Critical Weakness W1

We acknowledge that DBCopilot introduced schema-graph construction, random-walk sampling, and synthetic question generation. Building on these components, SQuAR introduces a different retrieval formulation for open-pool schema retrieval: rather than training a generative router, it directly indexes synthetic question–schema pairs and aggregates retrieved evidence to identify the correct database and recover complete sets of related tables, without autoregressive decoding.

This distinction offers four practical advantages:

(1) Schema extensibility: DBCopilot learns schema routing through a trained generative model, whereas SQuAR can incorporate new databases and tables by generating and indexing additional anchors, without retraining a schema-specific router.

(2) Online efficiency: SQuAR replaces autoregressive schema decoding with dense retrieval and evidence aggregation, achieving 3.5–11.2× faster median retrieval than DBCopilot across our four benchmarks (Table 7).

(3) Synthetic-example efficiency: DBCopilot originally uses 100K synthetic training examples per collection, whereas SQuAR uses only 10K anchors per benchmark. With 10× fewer synthetic examples, SQuAR-UAE achieves 19.34 and 21.12 percentage points higher ERC@5 on Spider and BIRD, respectively, compared with the original DBCopilot.

(4) Order-independent evidence aggregation: Rather than generating schema elements as an autoregressive sequence, SQuAR directly aggregates database and table evidence from multiple retrieved question–schema pairs, without relying on an ordering of schema elements during decoding.

To address backbone fairness, we additionally evaluated SQuAR using UAE-Large-V1 (335M) instead of Qwen3-Embedding-4B. Since DBCopilot employs T5-base (220M) as a generative router rather than a dedicated embedding model, UAE-Large-V1 provides a comparison at a similar parameter scale, although the architectures differ. SQuAR-UAE achieves 95.55%/89.50% ERC@5 on Spider/BIRD, compared with 76.21%/68.38% for DBCopilot. These results show that SQuAR's gains do not depend solely on its original 4B-parameter embedding backbone.

To control for differences in synthetic question generation, we additionally trained DBCopilot's T5-base router using the exact same 10K LLM-generated question–schema pairs used to construct SQuAR's retrieval index. Under this matched-data setting, DBCopilot achieves 50.29%/35.20% ERC@5 on Spider/BIRD, compared with 95.55%/89.50% for SQuAR-UAE. Although the retrieval architectures remain different, these results provide complementary evidence that SQuAR's improvements are not solely attributable to differences in synthetic question generation or its original larger embedding backbone.

MethodModelSpider ERC@5BIRD ERC@5BEAVER-DW ERC@5BEAVER-NW ERC@5
DBCopilot (original)T5-base (220M)76.21%68.38%8.26%12.5%
DBCopilot (matched SQuAR 10K data)T5-base (220M)50.29%35.20%0.83%0.00%
SQuAR-UAEUAE-Large-V1 (335M)95.55%89.50%21.49%20.45%
AuthorsOpenReview ↗

Response to Critical Weakness W3 and Question Q1

Response to Critical Weakness W3

We acknowledge that SQuAR uses a fixed retrieval budget KK rather than query-adaptive selection. Our primary objective is complete schema recovery, since missing a required table or selecting the wrong database prevents the construction of the gold SQL.

This trade-off highlights the importance of both ERC and ERJ. ERC measures complete, correctly routed retrieval, while ERJ penalizes unnecessary tables. CORE-T achieves higher ERJ on Spider and BIRD, reflecting better selectivity, whereas SQuAR achieves higher ERC, reflecting better completeness.

Importantly, CORE-T's higher selectivity did not translate into higher downstream execution accuracy (EA) with our fixed GPT-5.5 generator. SQuAR achieves the highest EA among evaluated retrieval methods (Figure 4). However, ERJ may be more valuable for smaller-context models or schemas with extensive per-table metadata, where over-retrieval is more costly. Query-adaptive selection for SQuAR is a promising future extension to improve its selectivity while preserving completeness.

Response to Question Q1

We thank the reviewer for this suggestion. To address this concern, we trained DBCopilot's T5-base router (220M parameters) using the exact same 10K Qwen-generated question–schema pairs used to construct SQuAR's retrieval index. We additionally evaluated SQuAR using UAE-Large-V1 (335M parameters) instead of Qwen3-Embedding-4B (4B) to reduce the disparity in backbone size.

Comparison with DBCopilot using matched synthetic data (ERC@5, %).

DatasetDBCopilot (Original)DBCopilot (Matched 10K)SQuAR-UAE
Spider76.2150.2995.55
BIRD68.3835.2089.50
BEAVER-DW8.260.8321.49
BEAVER-NW12.500.0020.45

As shown in the table, SQuAR-UAE achieves higher ERC@5 than both the original DBCopilot and its matched-data variant across all four benchmarks. On Spider/BIRD, SQuAR-UAE achieves 95.55%/89.50% ERC@5, compared with 50.29%/35.20% for DBCopilot trained on the identical synthetic dataset.

By using the exact same synthetic questions for both methods and a substantially smaller embedding backbone for SQuAR, these results demonstrate that SQuAR's improvements cannot be attributed solely to the stronger Qwen2.5-14B-Instruct model used for offline question generation or the original Qwen3-Embedding-4B retrieval backbone.

AuthorsOpenReview ↗

Response to Question Q2 and Question Q3

Response to Question Q2

We thank the reviewer for this suggestion. To directly address this question, we implemented a direct table retrieval baseline using the same Qwen3-Embedding-4B backbone as SQuAR. Tables are represented by their database and table names, columns, key relationships, and purpose descriptions. The baseline performs dense retrieval with one-hop foreign-key expansion.

SQuAR versus Simple Baseline using Qwen3-Embedding-4B (ERC, %).

DatasetKDirect-Table + FKSQuAR
Spider589.0796.13
1089.4696.81
1589.4696.81
BIRD583.7791.40
1092.3797.46
1594.5297.52
BEAVER-DW514.0522.31
1028.9349.59
1543.8066.12
BEAVER-NW57.9517.05
1014.7728.41
1520.4543.18

SQuAR achieves higher ERC in all 12 comparisons, with improvements ranging from 3.00 to 22.73 percentage points. Notably, on BEAVER-DW and BEAVER-NW, SQuAR improves ERC@15 from 43.80%/20.45% to 66.12%/43.18%, respectively. Our existing QGpT and CORE-T comparisons provide complementary evidence. QGpT retrieves tables using enriched representations containing schema metadata, sample rows, and synthetic questions, while CORE-T combines table-level retrieval with LLM-based refinement and joinability information. Using their respective embedding backbones, BGE-M3 (QGpT) and UAE-Large-V1 (CORE-T), SQuAR-BGE achieves 94.78% ERC@5 versus 85.01% for QGpT, while SQuAR-UAE achieves 95.55% versus 87.52% for CORE-T on Spider.

SQuAR versus competing methods using matched embedding backbones (ERC, %).*

DatasetKSQuAR-BGEQGpTSQuAR-UAECORE-T
Spider594.7885.0195.5587.52
Spider1094.7875.7396.1388.10
Spider1594.9775.6395.8487.72
BIRD588.5982.0189.5079.34
BIRD1096.2886.9697.1384.8
BIRD1596.3586.9097.2085.66
BEAVER-DW516.5315.7021.4910.74
BEAVER-DW1039.6727.2742.1516.53
BEAVER-DW1555.3734.7160.3316.53
BEAVER-NW518.182.2720.459.09
BEAVER-NW1029.555.6832.9520.45
BEAVER-NW1553.417.9548.8617.0

SQuAR-BGE outperforms QGpT in all 12 settings, while SQuAR-UAE outperforms CORE-T in all 12 settings. Overall, SQuAR achieves higher ERC in all 24 matched-backbone comparisons, providing further evidence that its gains cannot be attributed solely to its stronger embedding backbone.

Response to Question Q3

We thank the reviewer for this question. ERC can lead to different conclusions about retrieval effectiveness that better align with downstream SQL generation. Unlike Complete Recall (CR), ERC requires both correct database identification and retrieval of all required tables, two essential requirements for generating the target SQL.

Consider IterJAR and CORE-T at K = 15:

Complete Recall (CR), ERC, and downstream execution accuracy (EA) at K = 15 (%).

DatasetMethodCRERCEA
SpiderIterJAR97.4960.4450.0
CORE-T95.7087.7271.7
BIRDIterJAR96.5470.9335.5
CORE-T87.1685.6642.0

Despite achieving higher CR on both datasets, IterJAR yields lower downstream execution accuracy. On Spider, CORE-T achieves 21.7 percentage points higher EA, despite its lower CR, alongside a 27.3-point improvement in ERC. The same pattern holds on BIRD, where CORE-T achieves 6.5 points higher EA and 14.7 points higher ERC.

These examples illustrate that ERC measures whether retrieval produces an end-to-end-ready schema for downstream SQL generation. As demonstrated in Section 5.5 (Figure 4), higher ERC is generally associated with higher execution accuracy (EA) using the same GPT-5.5 generator. Thus, ERC provides a more task-relevant measure of open-pool schema retrieval by assessing whether the retrieved schema contains the correct database and all necessary tables to support SQL generation.

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

This paper addresses the problem of open-pool schema retrieval, where a method must identify the target database and retrieve all tables required to answer the question over a collection of databases. It proposes a method, SQuAR that generates synthetic questions offline from foreign-key connected table sub-schemas and indexes them as dense retrieval anchor. At inference time, it uses nearest-neighbor retrieval with frequency-based voting to rank databases and tables. This reduces LLM calls at inference time and shifts computation offline. It also introduce ERC and ERJ, two retrieval metrics that measure complete, correctly routed retrieval and account for over-retrieval. The method achieves high ERC and table recall and low latency.

Critical strengths
  • Accurately identified the gap in measuring the retrieval of sufficient set of tables that supports query execution. The proposed metrics are useful.
  • The proposed method is simple and effective. Also low-latency and LLM-related costs are valid requirements in many real-world applications.
  • Empirical results sufficiently capture the main claims made in the paper.
Critical weaknesses
  • There are some missing related works and baselines. Including these will help distinguish the novelty and establish the effectiveness of the proposed approach. Some of these include REaR (ACL 2026), Fine-Grained Table Retrieval for Open-Domain Tabular Question Answering (AKBC 2026).
  • The unified-pool experiment is useful but it only focuses on 37 databases and 619 tables. This might be too small to fully establish scalability to the much larger schema collections. Given their LLM-free and open-pool retrieval setting, it would be interesting to see how the performance generalizes to more realistic data-lake settings.
Questions for authors

N/A

1 comment
AuthorsOpenReview ↗

Response to Reviewer SRtJ

Response to Critical Weakness W1

We thank the reviewer for highlighting these relevant works. We agree that both REaR (ACL 2026) and DCTR (AKBC 2026) are relevant approaches to multi-table retrieval.

REaR focuses on expanding and refining initially retrieved tables using structural joinability, whereas SQuAR incorporates foreign-key relationships during offline synthetic-question generation and retrieves database and table evidence through precomputed anchors.

To further establish SQuAR's effectiveness, we additionally evaluated REaR as an experimental baseline across all four benchmarks at K ∈ {5, 10, 15}. The following table compares SQuAR with REaR's final refinement stage.

SQuAR versus REaR across retrieval budgets (ERC, %).

DatasetKREaRSQuAR
Spider587.2396.13
1073.9896.81
1561.9096.81
BIRD589.1891.40
1079.0197.46
1568.9097.52
BEAVER-DW512.4022.31
1032.2349.59
1548.7666.12
BEAVER-NW512.5017.05
1019.3228.41
1521.5943.18

SQuAR achieves higher ERC than REaR in all 12 comparisons across the four benchmarks and three retrieval budgets, providing further evidence of the effectiveness of its anchor-based retrieval formulation for complete, correctly routed schema recovery.

DCTR employs LLM-based query decomposition and component-level retrieval, whereas SQuAR directly retrieves database and table evidence using a single query embedding over precomputed question–schema anchors.

We will incorporate the REaR experimental results and discuss both REaR and DCTR in the revised manuscript, clarifying their methodological differences from SQuAR.

Response to Critical Weakness W2

We thank the reviewer for this valuable suggestion. Our unified-pool experiment evaluates retrieval over a heterogeneous collection combining four benchmarks, comprising 37 databases, 619 tables, 5,477 columns, and 2,777 annotated questions. In terms of table count, this unified pool is 7.6×, 8.3×, 6.4×, and 1.7× the size of the individual Spider, BIRD, BEAVER-DW, and BEAVER-NW evaluation pools, respectively.

Importantly, SQuAR maintains strong retrieval effectiveness in this expanded setting. At K = 5, SQuAR achieves 85.31% ERC, compared with 65.21% for QGpT and 54.41% for DBCopilot. Moreover, SQuAR's anchor-based indexing can incorporate additional databases and tables without requiring corpus-wide pairwise compatibility computations. In comparison, extending CORE-T and IterJAR to the unified pool is more computationally demanding due to their corpus-wide compatibility preprocessing, which involves approximately 191K table pairs and 15M column pairs.

These results provide encouraging evidence of SQuAR's effectiveness and practical scalability advantages in heterogeneous schema retrieval. We agree that evaluating substantially larger, real-world data lakes would provide further insight. Such an evaluation would require representative queries with annotated target databases and complete required-table sets, potentially requiring new benchmarks and substantial annotation effort. We appreciate this suggestion and identify larger-scale evaluation as an important direction for future work.

Reviewer_cLm1OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
1
Read review
Paper summary

This paper presents SQuAR, a schema retrieval method that generates synthetic questions over connected table subsets offline, and collects database/table information from similar synthetic questions at inference time.

Critical strengths

I like the idea of generating synthetic questions over related tables, as it brings table relationships into retrieval without needing LLM calls at query time.

Critical weaknesses
  1. The paper mentions that the numbers of tables and columns for sub-schema sampling come from empirical distributions of each dataset, but does not mention which split. And from the description it seems to be from the dev set. If so, this is a form of test-set leakage, even if it is only size statistics, which biases the anchors towards the dev/test set.
  2. SQuAR used Qwen3-Embedding-4B, while other baselines use much smaller retrievers. It is unclear to me how much of the gain comes from the proposed design vs the stronger backbone.
  3. Also, it is unclear to me why there’s a discrepancy in the numbers reported in Table 2 and Figure 11's normal method. Table 2 reports ERC@5 as 0.96/0.91, while Figure 11’s normal method numbers are around 0.55-0.60 (Spider/BIRD). Even though the two runs use different splits, the paper does not explain why the results differ this much.
Questions for authors

Check weaknesses

3 comments
AuthorsOpenReview ↗

Response to Reviewer cLm1

Response to Critical Weakness W1

We thank the reviewer for highlighting this concern. Although we did not use any evaluation question text, SQL queries, or exact gold table sets to generate the synthetic anchors, we conducted an additional experiment to assess the robustness of our method independently of the sub-schema sampling procedure:

  • Spider and BIRD: Both table and column count distributions are now estimated exclusively from the training splits.
  • BEAVER-DW and BEAVER-NW: Since these datasets do not provide training splits, we instead employ fixed, evaluation-independent sampling priors, using a decay-based distribution.

We regenerated the 10,000 synthetic anchors per dataset under the revised protocol and repeated the evaluation. The new results are as follows:

Table 1: Original versus Alternative SQuAR retrieval performance (ERC, %).

DatasetKOriginalAlternative
Spider596.1395.94
Spider1096.8196.32
Spider1596.8196.23
BIRD591.4091.26
BIRD1097.4697.46
BIRD1597.5297.52
BEAVER-DW522.3121.49
BEAVER-DW1049.5942.98
BEAVER-DW1566.1263.64
BEAVER-NW517.0515.90
BEAVER-NW1028.4123.86
BEAVER-NW1543.1835.22

The alternative sampling protocol produces nearly unchanged ERC on Spider and BIRD across retrieval budgets. On BEAVER, the decrease is more pronounced, potentially because the fixed decay-based prior favors smaller sub-schemas and provides less coverage of larger table combinations required by some queries. Unlike Spider and BIRD, BEAVER does not provide training data from which representative query-complexity distributions can be estimated.

Nevertheless, SQuAR outperforms the strongest existing baseline in ERC across all four benchmarks and retrieval budgets under the alternative protocol. This demonstrates that SQuAR maintains strong complete-schema retrieval effectiveness without relying on evaluation-derived sampling statistics

Response to Critical Weakness W2

We thank the reviewer for raising this concern. To investigate whether SQuAR's improvements stem primarily from its stronger embedding backbone, we additionally evaluated SQuAR using BGE-M3 (used by QGpT) and UAE-Large-V1 (used by CORE-T), alongside Qwen3-Embedding-4B.

All SQuAR variants use identical sets of 10k synthetic anchors per dataset, retrieval settings, and voting methods, with only the embedding backbone changed. For this experiment, we use the alternative sampling protocol described in W1.

Table 2: ERC (%) comparison across embedding backbones and competing methods.

DatasetKSQuAR-Qwen (Alt.)SQuAR-BGE (Alt.)QGpTSQuAR-UAE (Alt.)CORE-T
Spider595.9495.0785.0195.6587.52
Spider1096.3295.0775.7395.8488.10
Spider1596.2394.5875.6396.1387.72
BIRD591.2687.9482.0189.9079.34
BIRD1097.4696.2886.9696.8784.8
BIRD1597.5296.4886.9097.2085.7
BEAVER-DW521.4911.5715.7016.5310.74
BEAVER-DW1042.9834.7127.2738.0216.53
BEAVER-DW1563.6451.2434.7161.9816.53
BEAVER-NW515.9017.042.2720.459.09
BEAVER-NW1023.8626.135.6827.2720.45
BEAVER-NW1535.2231.817.9543.1817.05

On Spider and BIRD, SQuAR maintains comparable performance with substantially smaller embedding models. BEAVER-DW is more sensitive to backbone choice, particularly with BGE-M3, whereas UAE-Large-V1 outperforms Qwen3-Embedding-4B on BEAVER-NW at all three retrieval budgets.

Importantly, SQuAR-BGE outperforms QGpT in 11 of 12 settings, while SQuAR-UAE outperforms CORE-T in all 12 settings. Overall, SQuAR achieves higher ERC in 23 of 24 matched-backbone comparisons across the four datasets and three retrieval budgets.

Response to Critical Weakness W3

We thank the reviewer for this observation. The main difference is the substantially larger retrieval space in Figure 11. Table 2 in the paper evaluates the original Spider/BIRD development databases (20/11 databases), whereas Figure 11 uses query-level resplits covering 166/80 databases to enable training and evaluation database overlap for workload-aware sampling.

Both experiments retain the same 10K-anchor budget, reducing the average anchors per database from 500 to 60 on Spider and 909 to 125 on BIRD. This expanded retrieval space is associated with substantial decreases in DB Recall@1 (96.32% -> 68.9% on Spider; 98.89% -> 76.81% on BIRD) and ERC@5 (96.13% -> 55.08% on Spider; 91.40% -> 59.33% on BIRD).

Importantly, Figure 11 compares Normal and Workload under identical evaluation conditions, ensuring a fair comparison of workload-aware sampling.

Thanks for providing results from the additional experiments, especially the backbone comparison, and clarification on the Table2 vs Fig11 difference. However, I would like to maintain my score.

Thank you for acknowledging the effort we have put into addressing all the concerns you raised. However, we would appreciate knowing whether there are any remaining technical reasons that influenced your decision to maintain your score, so that we can provide further clarification with evidence. Our ultimate goal is to refine and strengthen our technical contribution and convince you to raise your score.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Summary for the Area Chair: Rebuttal Findings and Manuscript Revisions

Dear AC,

We thank the reviewers for their constructive feedback and for recognizing the motivation, practical relevance, effectiveness, and clarity of our work. During the discussion period, we conducted several additional experiments and incorporated the new findings into a revised manuscript. We summarize the main findings below:

  • Robustness and evaluation independence. We regenerated the synthetic anchors using training-only statistics for Spider and BIRD and evaluation-independent sampling distributions for BEAVER. SQuAR preserves its performance advantage across all four benchmarks, reinforcing the robustness of our original findings (cLm1-W1).

  • Effectiveness of anchor-based retrieval. We conducted three complementary experiments: (i) matched embedding backbones, outperforming competing methods in 23 out of 24 settings under the evaluation-independent sampling protocol; (ii) training DBCopilot using identical synthetic question–schema pairs; and (iii) comparing against direct table retrieval with foreign-key expansion using the same backbone, outperforming it in all 12 settings. These results demonstrate that SQuAR’s gains are not merely attributable to stronger underlying models (cLm1-W2, aB58-W1, aB58-Q1, aB58-Q2).

  • Additional baselines and methodological contributions. We evaluated REaR (ACL 2026), outperforming it in all 12 settings, and clarified SQuAR’s methodological distinctions from related work. These distinctions translate into practical advantages, including 3.5–11.2× faster median retrieval than DBCopilot and schema extensibility without router retraining (SRtJ-W1, aB58-W1, aB58-W2).

  • Evaluation and experimental clarifications. We empirically demonstrated how ERC better reflects downstream SQL execution accuracy than Complete Recall, explained the workload-aware performance discrepancy, and discussed scalability and selectivity trade-offs (aB58-Q3, cLm1-W3, SRtJ-W2, aB58-W3).

Taken together, SQuAR’s contribution extends beyond the application of established retrieval techniques: it introduces an effective retrieval formulation for complete multi-table schema recovery, workload-adaptive anchor generation, and an evaluation framework aligned with downstream SQL execution. The consistent empirical gains and practical advantages further substantiate the significance of these contributions.

We would greatly appreciate your consideration of the revised manuscript and the additional experimental evidence.

Thank you for your time and consideration.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.