CORE-T: COherent REtrieval of Tables for Text-to-SQL

Soliman, Hassan; Gupta, Vivek; Roth, Dan; Gurevych, Iryna

EMNLP 2026 Main

CORE-T
COherent REtrieval of Tables
for Text-to-SQL

Retrieving compact, relevant, and join-coherent table sets from pooled open-book collections.

Hassan Soliman¹, Vivek Gupta², Dan Roth^3,4, Iryna Gurevych¹

¹ UKP Lab, TU Darmstadt and ATHENE ² Arizona State University ³ University of Pennsylvania ⁴ Oracle AI

Paper ArXiv Code Data

Abstract

TL;DR

Dense retrieval finds relevant tables but also many distractors. CORE-T enriches tables offline, selects a coherent subset with one LLM call, and restores strongly compatible tables. It improves table-selection F1 by up to 22.7 points and multi-table execution accuracy by up to 24.4 points over dense retrieval, while outperforming join-aware and agentic baselines in nearly every available comparison.

Realistic text-to-SQL workflows often require joining multiple tables. As a result, accurately retrieving the relevant set of tables becomes a key bottleneck for end-to-end performance. We study an open-book setting where queries must be answered over large, heterogeneous table collections pooled from many sources, without clean scoping signals such as database identifiers. Here, dense retrieval achieves high recall but returns many distractors, while join-aware alternatives often rely on extra assumptions and/or incur high inference overhead. We propose CORE-T, a scalable, training-free framework that enriches tables with LLM-generated purpose metadata and pre-computes a lightweight table-compatibility cache. At inference time, dense retrieval returns top-K candidates; a single LLM call selects a coherent, joinable subset, and a two-step additive adjustment stage restores strongly compatible tables. Across BIRD, SPIDER, MMQA, and BEAVER, CORE-T improves over dense retrieval by up to 22.7 points in table-selection F1 while returning up to 40% fewer tables, and by up to 24.4 points in multi-table execution accuracy, and uses 1.64-4.20 times fewer total selection tokens than LLM-intensive baselines.

Motivation

Why open-book retrieval is hard

Without a database identifier, retrieval must disambiguate similar tables and preserve the complete join path.

Closed-book setting

Database ID known

User querySearch starts with scope

One known schema Retrieval stays inside the correct database

Remove the scoping signal

Open-book setting

Database ID unavailable

Same queryMust discover its scope

Pooled collection Relevant Distractor Bridge Similar name Relevant Other source

Disambiguate similar tables

Preserve bridge tables

Remove distractors

Compact, relevant, join-coherent table set

Pipeline

How CORE-T works

Offline

Build semantic and relational evidence once, then reuse it for every question.

Semantic track

Table snapshotHeaders plus five sampled rows

Purpose enrichmentAn LLM describes the role of each table

Dense indexSnapshot and purpose are embedded together

Structural track

Column pairsCandidate links across table pairs

Compatibility scoreHeaders, values, uniqueness, and subset evidence

Compatibility cacheStrong table-to-table links stored for reuse

Reusable evidence Purpose-enriched index Compatibility cache

Online

Convert one user question into a compact, join-ready schema with a single selection call.

User queryThe natural-language information need

1. RetrieveFetch a high-recall top-K candidate pool

2. SelectOne LLM call chooses a relevant, coherent seed set

3. AdjustRecover strong neighbors, then add only positive-gain tables

Compact join-coherent table set for SQL generation

What we contribute

Retrieval designed for coherent sets

Training-free, join-coherent retrieval

CORE-T combines purpose-enriched relevance with cached table-to-table compatibility, without requiring database IDs, gold foreign keys, or new model training.

Compact online inference

One LLM selection call replaces iterative or multi-draft inference, while the adjustment stage provides targeted recall recovery.

Pooled multi-database evaluation

Experiments span BIRD, SPIDER, MMQA, and BEAVER in open-book settings with table retrieval and end-to-end SQL execution metrics.

Evaluation

Four open-book benchmarks

Tables are pooled across databases or schemas and database identifiers are removed.

BIRD75tables
1,534 queries

SPIDER81tables
1,034 queries

MMQA710tables
1,105 queries

BEAVER463tables
209 queries

MMQA uses a stratified one-third subset. BEAVER is already open-book; the other corpora are pooled to approximate integrated multi-source analytics.

Performance

Results across every baseline

Complete main-paper results for table selection, downstream SQL execution, and selection efficiency.

Table 2. Table selection

Average number of returned tables, precision, F1, and perfect recall. All methods use UAE-Large-V1 for dense retrieval.

Method	BIRD (n=1,534, gold avg.=1.95)				SPIDER (n=1,034, gold avg.=1.51)				MMQA (n=1,105, gold avg.=2.20)				BEAVER (n=209, gold avg.=4.44)
Method	Avg.	P	F1	PR	Avg.	P	F1	PR	Avg.	P	F1	PR	Avg.	P	F1	PR
Llama-3.1-8B-Instruct selector
DR@5	5.0	34.9	49.4	83.0	5.0	29.4	43.9	95.6	5.0	30.8	42.6	50.3	5.0	33.9	36.5	13.9
ReAct	4.8	43.1	55.6	86.1	4.5	39.7	52.2	92.8	6.1	29.6	40.4	53.1	8.4	9.3	12.2	8.1
JAR@5	5.0	35.5	50.1	86.0	5.0	29.6	44.2	96.7	-	-	-	-	-	-	-	-
REAR	5.0	36.1	51.0	88.3	5.0	28.2	42.1	90.8	5.0	28.4	39.2	47.8	5.0	33.6	35.5	8.1
ARM	5.3	40.4	53.5	90.9	-	-	-	-	-	-	-	-	-	-	-	-
CORE-T	4.1	50.0	62.3	90.0	4.2	40.2	53.8	96.6	5.1	38.4	49.2	61.3	5.1	39.3	39.5	15.3
Qwen-2.5-7B-Instruct selector
DR@5	5.0	35.2	49.8	84.2	5.0	29.6	44.1	96.4	5.0	31.1	43.0	51.0	5.0	34.8	37.3	14.8
ReAct	3.3	65.6	71.6	81.3	2.7	67.6	75.5	93.4	4.9	42.1	50.4	55.4	8.2	21.0	25.3	16.3
JAR@5	5.0	35.5	50.1	86.0	5.0	29.5	44.1	96.3	-	-	-	-	-	-	-	-
REAR	5.0	36.1	51.0	88.3	5.0	28.2	42.1	90.8	5.0	28.4	39.2	47.8	5.0	33.6	35.5	8.1
ARM	3.5	59.2	68.3	84.6	-	-	-	-	-	-	-	-	-	-	-	-
CORE-T	3.1	63.1	72.3	87.0	3.0	54.4	66.8	94.9	4.1	48.3	56.4	59.4	5.3	42.8	43.0	17.7

Best scoreSecond-best scoreAvg. average tables returned, lower is betterP precisionPR perfect recall

17 / 19query-level F1 comparisons

On query-level table-selection F1, with Qwen-2.5-7B-Instruct as selector, CORE-T is significantly better in 17 of 19 dataset-baseline comparisons, tied with ReAct on BIRD, and significantly lower only against ReAct on SPIDER.

Table 3. End-to-end query execution accuracy

Execution accuracy with Llama-3.1-8B-Instruct as table selector, reported for multi-table queries, all queries, and the perfect-recall subset.

Method	BIRD (n=1,534)			SPIDER (n=1,034)			MMQA (n=1,105)			BEAVER (n=209)
Method	EX_MT76.4%	EX_all100%	EX_PR90.0%	EX_MT44.4%	EX_all100%	EX_PR96.6%	EX_MT99.6%	EX_all100%	EX_PR61.3%	EX_MT98.6%	EX_all100%	EX_PR15.3%
Llama-3.2-3B SQL generator
DR@5	3.3	4.0	3.9	10.0	13.9	13.9	2.1	2.1	1.8	0.5	0.5	0.0
ReAct	9.9	10.8	10.4	12.4	20.0	20.0	2.5	2.5	2.2	0.0	0.0	0.0
JAR@5	7.8	8.3	8.1	10.9	14.4	14.4	-	-	-	-	-	-
REAR	9.6	10.0	9.8	26.4	29.6	29.3	8.4	8.3	6.9	1.0	1.0	0.0
ARM	4.9	5.1	5.0	-	-	-	-	-	-	-	-	-
CORE-T	15.7	16.6	16.2	34.4	34.5	34.1	18.9	18.8	15.8	0.5	0.5	0.5
Oracle	24.8	27.8	27.8	45.1	58.9	58.9	45.4	45.4	45.4	1.5	1.9	1.9
Gemma-3-4B SQL generator
DR@5	14.6	20.0	19.0	38.8	51.5	51.1	18.9	18.9	16.8	5.3	5.3	1.0
ReAct	15.3	21.2	20.5	40.1	51.8	51.5	18.0	17.9	15.2	3.4	3.3	0.5
JAR@5	18.3	23.0	22.2	41.4	51.4	51.3	-	-	-	-	-	-
REAR	15.9	20.1	19.6	35.3	45.1	44.7	16.1	16.1	12.9	4.9	4.8	0.5
ARM	16.6	22.3	21.7	-	-	-	-	-	-	-	-	-
CORE-T	16.6	21.9	21.1	43.8	54.2	53.8	22.8	22.8	20.8	3.9	3.8	1.0
Oracle	24.7	30.4	30.4	53.8	65.7	65.7	47.6	47.5	47.5	4.4	4.8	4.8
GPT-4o-mini SQL generator
DR@5	34.6	40.0	37.7	53.2	65.4	64.7	31.6	31.7	27.8	5.3	5.3	1.9
ReAct	38.1	42.5	40.9	50.5	63.7	63.4	33.2	33.1	28.7	3.4	3.3	0.5
JAR@5	36.8	41.6	40.4	55.3	66.0	65.7	-	-	-	-	-	-
REAR	38.4	42.3	40.6	52.3	62.3	61.7	28.9	29.0	25.2	4.4	4.3	0.5
ARM	37.9	42.3	41.5	-	-	-	-	-	-	-	-	-
CORE-T	38.6	43.0	41.6	56.9	66.7	66.2	35.8	35.8	33.3	5.3	5.7	1.9
Oracle	47.8	50.7	50.7	64.5	71.8	71.8	65.8	65.7	65.7	6.8	7.2	7.2

Best non-oracle scoreSecond-best non-oracle scoreEX_MT multi-table queriesEX_all all queriesEX_PR queries with perfect table recall

Table 4. Selection efficiency

Selection-step tokens in millions with UAE-Large-V1 embeddings and Llama-3.1-8B-Instruct as selector. Factors are relative to CORE-T.

Method	BIRD			SPIDER			MMQA			BEAVER
Method	Input M	Output M	Total M	Input M	Output M	Total M	Input M	Output M	Total M	Input M	Output M	Total M
ARM	51.74.79x	0.730.43x	52.44.20x	-	-	-	-	-	-	-	-	-
ReAct	43.54.03x	1.070.63x	44.53.57x	24.24.99x	0.740.65x	24.94.16x	26.64.65x	0.840.68x	27.53.95x	2.61.25x	1.215.05x	3.831.64x
CORE-T	10.81.0x	1.711.0x	12.51.0x	4.81.0x	1.141.0x	5.991.0x	5.71.0x	1.231.0x	6.961.0x	2.11.0x	0.241.0x	2.331.0x

Cell color relative cost vs. CORE-TAt or below CORE-TModerate overheadHigh overhead

Takeaway. CORE-T provides the strongest overall precision-recall and efficiency trade-off: it dominates the dense and join-aware baselines wherever they are available, usually outperforms the agentic alternative, and never produces a significant downstream execution loss.

Remaining challenge. BEAVER's anonymized schemas and many empty tables make it a demanding stress test. CORE-T leads in F1 and perfect recall, but absolute execution accuracy remains low even with oracle tables.

Reproducibility

Getting started

The repository exposes each offline and online stage as a standalone Python script.

01 Create the environment

# Python 3.12 with uv
uv python install 3.12
uv venv .venv
source .venv/bin/activate
uv pip install -r requirements.txt

02 Prepare tables and caches

python offline_preprocessing/table_preprocessor.py \
  --dataset bird \
  --llm-model "huggingface:Qwen/Qwen2.5-7B-Instruct" \
  --embedding-model "fireworks:WhereIsAI/UAE-Large-V1"

python offline_preprocessing/metadata_generator.py \
  --dataset bird \
  --llm-model "huggingface:Qwen/Qwen2.5-7B-Instruct" \
  --embedding-model "fireworks:WhereIsAI/UAE-Large-V1"

python offline_preprocessing/compatibility_calculator.py \
  --dataset bird \
  --llm-model "huggingface:Qwen/Qwen2.5-7B-Instruct" \
  --embedding-model "fireworks:WhereIsAI/UAE-Large-V1"

03 Retrieve, select, and adjust

python dense_retriever.py \
  --dataset bird --top-k 10 \
  --llm-model "huggingface:Qwen/Qwen2.5-7B-Instruct" \
  --embedding-model "fireworks:WhereIsAI/UAE-Large-V1"

python table_selector.py \
  --dataset bird --top-k 10 \
  --llm-model "huggingface:Qwen/Qwen2.5-7B-Instruct" \
  --embedding-model "fireworks:WhereIsAI/UAE-Large-V1"

python table_augmentor.py \
  --dataset bird --top-k 10 \
  --llm-model "huggingface:Qwen/Qwen2.5-7B-Instruct" \
  --embedding-model "fireworks:WhereIsAI/UAE-Large-V1"

Model provider IDs and API credentials are configured through command-line options and a local .env file. See the repository README for complete commands, downloads, SQL generation, and evaluation.

Reference

Citation

If you use CORE-T, please cite the current arXiv version.

@misc{soliman2026core-t,
    title         = {CORE-T: COherent REtrieval of Tables for Text-to-SQL},
    author        = {Hassan Soliman and Vivek Gupta and Dan Roth and Iryna Gurevych},
    year          = {2026},
    eprint        = {2601.13111},
    archivePrefix = {arXiv},
    primaryClass  = {cs.CL},
    url           = {https://arxiv.org/abs/2601.13111}
}

CORE-TCOherent REtrieval of Tablesfor Text-to-SQL