Retrieve Only Relevant Tables Whether Few or Many: Adaptive Table Retrieval Method

Taehee Kim; Seungbin Yang; Jihwan Kim; Jaegul Choo

cs.IRApr 12, 2026

Retrieve Only Relevant Tables Whether Few or Many: Adaptive Table Retrieval Method

Authors: Taehee Kim, Seungbin Yang, Jihwan Kim, Jaegul Choo

Organizations: KAIST AI · Letsur

Abstract

Retrieving relevant tables from extensive databases for a given natural language query is essential for accurately answering questions in tasks such as text-to-SQL. Existing table retrieval approaches select a pre-determined set of k tables with the highest similarity to the query. However, the number of required tables varies across queries and cannot be known in advance. Enforcing a fixed number of retrieved tables regardless of the query may either retrieve an undersized set, failing to obtain all necessary evidence, or retrieve an oversized pool, including irrelevant tables. To address this issue, we propose an adaptive table retrieval method that adjusts the number of tables retrieved according to the requirements of each query. Specifically, we utilize an adaptive thresholding mechanism to selectively retrieve tables and integrate a sliding-window reranking algorithm to efficiently process a large table corpus. Extensive experiments on Spider, BIRD, and Spider 2.0 demonstrate that our method effectively addresses the limitations of the top-k retrieval strategy, improving performance in retrieval and downstream tasks. Our code and data are available at https://github.com/sbY99/Adaptive-Table-Retrieval.

Figures & tables

Figure 1: Rather than rely on a rigid fixed $k$ retrieval strategy, ATR retrieves only relevant tables. Gray indicates tables required by the query but not retrieved, red denotes irrelevant tables, and blue highlights retrieved relevant tables.

Figure 2: Retrieving irrelevant tables introduces noise, degrading performance on the text-to-SQL task. Execution accuracy consistently decreases as more irrelevant tables are added to the ground-truth tables.

Figure 3: Overview of the ATR framework. (A) Inference: ATR takes a query and candidate tables as input to compute relevance logits. It dynamically selects the subset of tables (e.g., $T_{2},T_{3}$ ) whose logits exceed that of the adaptive threshold token ( $T_{th}$ ), thereby automatically determining the specific number of tables needed for the query. (B) Training: ATR is optimized using three complementary objectives: 1) Adaptive Thresholding ( $L_{AT}$ ) trains the threshold token to serve as a decision boundary, enforcing relevant tables to score higher ( $L_{1}$ ) and irrelevant ones lower ( $L_{2}$ ) than $T_{th}$ ; 2) Relevance Calibration ( $L_{RC}$ ) maximizes the logit margin between relevant and irrelevant tables; and 3) Semantic Grouping ( $L_{SG}$ ) adopts contrastive learning to pull embeddings of joinable tables that are sharing join keys closer together while pushing non-joinable ones apart.

Encoder-based
		Spider				BIRD				Spider 2.0
	k	P	R	CR	F1	P	R	CR	F1	P	R	CR	F1
Contriever	3	46.5	94.0	89.2	60.1	45.6	72.9	55.1	54.6	24.7	37.6	24.8	27.4
	5	29.3	97.4	95.5	43.7	31.2	82.1	68.6	44.2	19.1	45.1	32.6	24.4
	10	15.0	99.2	98.7	25.6	18.5	96.1	92.6	30.5	13.3	56.5	44.8	19.5
UAE	3	46.1	93.0	88.0	59.5	48.2	79.0	63.5	58.3	27.1	40.6	27.6	29.7

Table 1: Retrieval performance comparison with baseline methods, evaluated using Precision (P), Recall (R), Complete Recall (CR), and F1 scores (F1). We use SGPT-5.8B as the backbone model for Murre. ATR consistently outperforms all baselines across datasets. The best and second-best scores for each metric are highlighted in bold and underlined , respectively.

		Spider			BIRD			Spider 2.0
	k	Llama (8B/70B)	Qwen (7B/32B)	Gemma (4B/27B)	Llama (8B/70B)	Qwen (7B/32B)	Gemma (4B/27B)	Llama (8B/70B)	Qwen (7B/32B)	Gemma (4B/27B)
Encoder-based
Contriever	3	54.6 / 64.8	65.4 / 66.3	58.0 / 67.4	21.7 / 40.4	35.4 / 46.0	21.2 / 38.9	1.1 / 3.4	0.5 / 3.9	0.2 / 2.5
	5	53.3 / 65.0	65.7 / 68.6	58.3 / 67.7	22.4 / 44.5	37.0 / 49.8	24.1 / 39.7	1.1 / 3.2	1.4 / 3.7	0.7 / 3.2
	10	54.6 / 67.1	66.0 / 70.1	60.3 / 67.6	24.3 / 47.4	39.1 / 53.0	25.6 / 43.2	1.1 / 3.5	0.9 / 4.1	0.2 / 3.4
UAE	3	55.7 / 65.0	63.6 / 68.0	52.5 / 68.1	22.6 / 40.7	35.1 / 45.4	20.1 / 38.8	0.7 / 3.5	1.1 / 3.4	0.5 / 3.4

Table 2: Text-to-SQL execution accuracy comparison across different table retrieval methods. ATR consistently outperforms baseline retrievers on the Spider, BIRD, and Spider 2.0 datasets.

Figure 4: Comparison of execution accuracy and average token length for the text-to-SQL task across different retrieval methods. ATR achieves higher accuracy with fewer tokens compared to the best-performing top-

k

approach.

Figure 5: Execution accuracy on the Spider 2.0 dataset across varying numbers of required tables.

|\mathcal{C}_{q}|

denotes the number of tables required per query.

Figure 6: Analysis of training loss hyper-parameters. Irr. indicates the number of retrieved tables irrelevant to the query, and Irr.+Join. represents the retrieved tables that are irrelevant but can be joined with the ground-truth tables.

	Spider		BIRD		Spider 2.0
	R	CR	R	CR	R	CR
ATR	99.5	99.2	98.2	96.0	72.4	64.4
$-$ (1) BCE	99.0	98.7	97.5	95.8	68.2	60.8
$-$ (2) Contrastive	99.0	98.4	97.4	95.2	69.1	60.1
$-$ (1) & (2)	96.4	94.4	91.8	85.7	67.7	58.2

Table 5: Ablation study on table representation losses.

Appendix figures & tables11 assets

Supplementary material from the paper’s appendix.

Appendix

		Spider				BIRD				Spider 2.0
	k	P	R	CR	F1	P	R	CR	F1	P	R	CR	F1
RankGPT-4o-mini	3	48.2	96.8	93.5	62.1	54.5	88.0	76.2	65.6	33.2	48.5	33.8	35.2
RankGPT-4o-mini	5	29.9	98.9	97.9	44.5	35.0	92.7	84.6	49.7	24.6	56.1	42.1	30.5
RankGPT-4o-mini	10	15.1	99.3	99.0	25.5	18.7	97.5	94.6	30.9	15.3	64.0	52.2	21.9
RankGPT-5-mini	3	48.9	98.2	95.5	63.1	57.3	91.6	81.6	66.3	36.6	53.6	37.5	39.4
RankGPT-5-mini	5	30.0	99.5	98.9	44.8	36.4	95.7	90.0	51.5	25.8	59.3	44.1	32.4

Appendix

Table 8: Performance comparison with LLM-based retrievers and rerankers. Bold indicates the highest score, and italic indicates the second-highest score.

Encoder-based
		Spider				BIRD				Spider 2.0
	k	Qwen3	DeepSeek-V2	GPT-5	Gemini-2.5	Qwen3	DeepSeek-V2	GPT-5	Gemini-2.5	Qwen3	DeepSeek-V2	GPT-5	Gemini-2.5
Contriever	3	71.7	61.7	60.4	73.7	45.2	29.9	35.1	48.8	3.2	1.8	3.2	2.5
	5	74.2	61.9	63.4	75.0	50.3	32.3	37.5	54.0	3.2	1.6	5.1	3.9
	10	75.9	62.3	64.8	76.5	53.7	36.8	40.4	58.1	3.7	1.6	5.8	4.4
UAE	3	71.9	61.5	60.6	73.3	46.1	31.8	34.2	50.5	3.7	1.6	4.6	2.5

Appendix

Table 9: Text-to-SQL execution accuracy comparison for the various models, including open source and proprietary LLMs, across different table retrieval methods.

Dataset	Metric	Sliding Window (Ours)	No Sliding Window	Reduction
Spider	Avg Peak	30.62 MB	63.98 MB	52.1%
Spider	Max Peak	44.01 MB	80.67 MB	45.4%
BIRD	Avg Peak	51.76 MB	93.32 MB	44.5%
BIRD	Max Peak	94.31 MB	145.13 MB	35.0%
Spider 2.0	Avg Peak	66.52 MB	340.57 MB	80.5%
Spider 2.0	Max Peak	284.44 MB	1027.88 MB	72.3%

Appendix

Table 10: GPU memory consumption profiling on the test sets. Avg Peak refers to the average maximum memory allocated per sample, and Max Peak denotes the single highest memory spike observed during inference.

Case 1: Cross-Database Joins
Query	“Could you provide, for the United States, France, China, Italy, Spain, Germany, and Iran, the total number of confirmed COVID-19 cases as of April 20, 2020, along with the number of cases per 100,000 people based on their total 2020 populations calculated by summing all relevant population entries from the World Bank data”
Gold	covid19_jhu_csse.summary , world_bank_wdi.indicators_data
ATR	covid19_jhu_csse.summary , world_bank_wdi.indicators_data
Top-5	world_bank_health_population.health_nutrition_population , world_bank_health_population.health_nutrition_population , world_bank_health_population.health_nutrition_population , world_bank_global_population.population_by_country , world_bank_wdi.series_times
Case 2: Ambiguous Table References
Query	“Calculate the average sales per order for each customer within distinct RFM segments, considering only ‘delivered’ orders”

Appendix

Table 11: Edge-case retrieval results. Green denotes correctly retrieved gold tables; red denotes irrelevant or missed tables.

Question
Please list player names which are higher than 180.
Tables Retrieved by ATR	Tables Retrieved by Top- $k$
"Player"	"Player" "Match" "League" "Team" "superhero"
SQL Generated by ATR	SQL Generated by Top- $k$

Retrieved Table Schema

Appendix

Table 12: A case study illustrating a scenario where the retrieval size

k

exceeds the number of tables required to resolve the query. As the top-

k

approach retrieves unnecessary tables, it propagates noise, degrading downstream SQL generation. Irrelevant tables are in red, and unnecessary tables that do not contribute to answering the query are highlighted in yellow. Retrieved table schema displays schema information for all retrieved tables.

Question
Please calculate the monthly average levels of PM10, PM2.5 FRM, PM2.5 non-FRM, volatile organic emissions, SO2 (scaled by a factor of 10), and Lead (scaled by a factor of 100) air pollutants in California for the year 2020.
Tables Retrieved by ATR	Tables Retrieved by Top- $k$
"pm25_frm_daily_summary" "lead_daily_summary" "pm10_daily_summary" "pm25_nonfrm_daily_summary" "so2_daily_summary" "voc_daily_summary"	"pm25_frm_daily_summary" "lead_daily_summary" "pm10_daily_summary" "pm25_frm_hourly_summary" "pm10_hourly_summary" "pm25_nonfrm_daily_summary" "so2_daily_summary" "voc_daily_summary"
SQL Generated by ATR	SQL Generated by Top- $k$

Retrieved Table Schema

Appendix

Table 13: A case study illustrating a scenario where the retrieval size

k

is smaller than the number of tables required to resolve the query. As the top-

k

approach fails to retrieve all necessary tables, it produces inaccurate SQL, whereas ATR retrieves all essential tables, enabling correct SQL generation. Tables not retrieved but crucial for SQL generation are displayed in light gray.

Figure 7: An illustrative example of the sliding window reranking process in ATR with four input tables, window size $W=2$ , and retention size $R=1$ . Tables highlighted in red indicate those whose rankings are finalized but excluded from the final set, while tables in green represent the tables that are ultimately retrieved. The input tables are initially ordered by query-table similarity in descending order from left to right. ATR processes tables starting from the lowest-ranked ones (rightmost) within the window. In each step, ATR compares the logits of table tokens with the threshold token ( $T_{th}$ ) and retains the top- $R$ tables for the next iteration.

Figure 8: Prompt template for Spider and BIRD datasets.

Figure 9: Prompt template for Spider 2.0 (BigQuery dialect)

Figure 10: Prompt template for Spider 2.0 (Snowflake dialect)

Figure 11: Prompt template for Spider 2.0 (SQLite dialect)

Table Retrieval

Table retrieval finds relevant tables and schema elements for question answering, analytics, and text-to-SQL.

See all 19 papers

CORE-T: COherent REtrieval of Tables for Text-to-SQL
Jan 19, 2026Hassan Soliman, Vivek Gupta, Dan Roth +1Text-to-SQL Information Retrieval
Ubiquitous Knowledge Processing Lab (UKP Lab), Department of Computer Science TU Darmstadt and National Research Center for Applied Cybersecurity ATHENE, Germany · Arizona State University · University of Pennsylvania
JoinGR: Learning to Traverse Join Graphs for Table Retrieval
Oct 1, 2026Sandipan De, Abhijit Chakraborty, Sambaran Bandyopadhyay +1Graph-Based Retrieval Table Retrieval
Arizona State University · Adobe Research, India
FGTR: Fine-Grained Multi-Table Retrieval via Hierarchical LLM Reasoning
Mar 13, 2026Chaojie Sun, Bin Cao, Tiantian Li +3Table Retrieval LLM Reasoning
Zhejiang University of Technology Hangzhou, China · University of Aberdeen Aberdeen, UK

Jan 19, 2026J/K move · Enter open · S save

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

Hassan Soliman, Vivek Gupta, Dan Roth, Iryna Gurevych

Ubiquitous Knowledge Processing Lab (UKP Lab), Department of Computer Science TU Darmstadt and National Research Center for Applied Cybersecurity ATHENE, Germany · Arizona State University · University of Pennsylvania

Text-to-SQL Information Retrieval Table Retrieval

Open paper AskPDF arXiv

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 (DR) 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, DR 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 DR 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 1.64-4.20x fewer total selection tokens than LLM-intensive baselines.

Abstract

Table Retrieval

Explore similar work

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

JoinGR: Learning to Traverse Join Graphs for Table Retrieval

FGTR: Fine-Grained Multi-Table Retrieval via Hierarchical LLM Reasoning