TableSeek: Structure-Preserving Agentic Evidence Seeking over Heterogeneous Table Corpora

Jiaming Tian; Liyao Li; Wentao Ye; Haobo Wang; Lihua Yu; Zujie Ren; Gang Chen; Junbo Zhao

cs.AISep 28, 2026

TableSeek: Structure-Preserving Agentic Evidence Seeking over Heterogeneous Table Corpora

Authors: Jiaming Tian, Liyao Li, Wentao Ye, Haobo Wang, Lihua Yu, Zujie Ren, Gang Chen, Junbo Zhao

Organizations: Zhejiang University · Bank of Hangzhou Co., Ltd. · Zhejiang Lab

Abstract

Open-domain table retrieval seeks tables that contain sufficient evidence for answering a question or verifying a claim. Yet semantic relevance is often misleading: topically similar tables may lack the required facts, while answer-bearing evidence is often confined to a few cells whose meaning depends on surrounding schema and table context. Heterogeneous schemas, value formats, and serializations further weaken one-shot matching. We present TableSeek, a structure-preserving agentic search framework for heterogeneous table corpora. Instead of ranking tables once, an LLM agent iteratively follows sparse clues, inspects schema-preserving previews, identifies schema- and value-level mismatches, and refines its investigation. TableSeek uses cells and schemas as evidence anchors while retaining complete tables as evidence units, enabling fine-grained localization without losing the context required for interpretation and answerability checking. Without relying on retriever training or a precomputed semantic index, TableSeek produces transparent evidence-seeking trajectories and achieves competitive end-to-end performance against strong retrieval-and-reranking pipelines on heterogeneous table benchmarks. These results suggest that active, structure-preserving evidence seeking is a promising paradigm for open-domain table retrieval.

Figures & tables

Figure 1: Semantic relevance

\neq

evidence sufficiency. The left table mentions Australia and Capital Region but lacks the actual capital city. The right table contains the exact answer Canberra , but uses Seat of Government instead of Capital , resulting in low lexical overlap with the query.

Figure 2: Overview of TableSeek for question answering over heterogeneous table corpora. Stage 1 scores tables against generated search patterns to form a ranked candidate list. Stage 2 inspects table previews and iteratively refines the patterns until a candidate is selected. Stage 3 checks answerability on the selected table, falling back to the next candidate when evidence is insufficient. Here, $F_{i}$ denotes a table format, and $T_{i}$ denotes the $i$ -th ranked table.

Dataset	Split	#Queries	#Tables
TCR-Bench	–	209	637
TARGET-TabFact	valid	12,792	1,696
TARGET-TabFact	valid-1k	1,000	1,696
TARGET-TabFact	test	12,779	1,695
TARGET-TabFact	test-1k	1,000	1,695

Table 1: Statistics of benchmarks and table corpora used in our experiments.

Method	Config	Recall	F1
Dense	E4B	0.172	0.118
Rerank (top-3)	E4B+R4B	0.359	0.269
Rerank (top-5)	E4B+R4B	0.421	0.317
TableSeek (5-step)	Qwen3-30B	0.416	0.284
TableSeek (10-step)	Qwen3-30B	0.507	0.361

Table 2: Main results on TCR-Bench (Mixed format) with Qwen3-30B as the QA/backbone model. Recall is R@1 for Dense/Rerank and R@all for TableSeek. Full results with additional backbones are given in the appendix.

Method	Config	Recall	Acc	Acc-Sup
Dense	Stella	0.519	0.721	0.489
Rerank (top-3)	Stella+R4B	0.623	0.752	0.562
Rerank (top-5)	Stella+R4B	0.656	0.754	0.572
TableSeek (5-step)	Qwen3-30B	0.617	0.754	0.566
TableSeek (10-step)	Qwen3-30B	0.648	0.764	0.589

Table 3: Results on TARGET-TabFact-test-1k with Qwen3-30B as the QA/backbone model. Retrieval is R@1 for Dense/Rerank and R@all for TableSeek.

Metric	Dense	Rerank		TableSeek
Metric	Dense	top-3	top-5	5step	10step
R@1	0.172	0.359	0.421	–	–
DS@1	0.252	0.478	0.540	–	–
R@all	–	–	–	0.416	0.507
DS@all	–	–	–	0.725	0.803
R@final	–	–	–	0.354	0.431

Table 4: Recall and Discriminative Score on TCR-Bench. Single-pass baselines are evaluated at their single retrieval decision (R@1/DS@1); TableSeek is evaluated both across all QA calls (R@all/DS@all) and at the final call used for answer generation (R@final/DS@final).

Figure 3: F1 across table formats and robustness statistics on TCR-Bench, using Qwen3-30B.

Figure 4: Wall-clock time breakdown by stage (dense encoding, reranking, LLM/QA) across methods on TCR-Bench.

Method	Calls (4B)	Calls (30B)
Dense	1055	1055
Rerank (top-3)	1682	1682
Rerank (top-5)	2100	2100
TableSeek (5-step)	918	921
TableSeek (10-step)	1581	1305

Table 5: Total inference calls on TCR-Bench. Dense/Rerank are independent of the QA backbone.

Strategy	R@all	F1	Avg. Steps	Time (h)
5-step setting
5-step (Full)	0.416	0.284	4.407	2.335
w/o fallback	0.364	0.263	3.919	2.094
w/o refine	0.239	0.173	3.555	1.733
match-only	0.359	0.249	4.435	1.841
match-only QA	0.359	0.065	4.445	1.132

Table 6: Ablation study on refine, fallback, and Structure-Preserving mechanisms on TCR-Bench (Mixed format) using Qwen3-30B.

Figure 5: Top-1 hit rate as a function of refinement round. Iterative refinement continues to improve retrieval quality for the stronger Qwen3-30B model well beyond the first round, whereas the smaller Qwen3-4B model plateaus almost immediately, indicating that the marginal benefit of refinement scales with model capability.

Figure 6: Average action counts per query on TCR-Bench for the default TableSeek configuration, excluding the initial ‘ pattern-gen ‘ step (fixed to 1 per query).

Appendix figures & tables29 assets

Supplementary material from the paper’s appendix.

Appendix

Hyperparameter	Value
Max LLM steps per query ( max_steps )	5
Candidate tables shown per refine ( n_candidates )	5
Max tables selectable per refine ( max_select_tables )	1
Table head preview length ( preview_chars )	800
Max matched rows per candidate ( max_matched_rows )	5
Max matched rows per pattern ( max_rows_per_pattern )	1

Appendix

Table 7: Default TableSeek (5s1-small) configuration.

Configuration	n_candidates	max_select_tables
5s1	5	1
5s3	5	3
10s1	10	1
10s3	10	3
10s5	10	5
r3	3	1

Appendix

Table 8: Refine configurations explored in this work.

Parameter	Value
temperature	$0.0$
max_tokens	$8192$

Appendix

Table 9: Decoding parameters used across all experiments.

Hyperparameter	Value
Table source ( table_source )	gold
Tables used for answering ( top_k )	1

Appendix

Table 10: Direct QA hyperparameters.

Hyperparameter	Value
Retrieved tables ( top_k )	5
Similarity	Cosine
Embedding normalization	$\ell_{2}$

Appendix

Table 11: Dense retrieval hyperparameters.

Hyperparameter	Value
Dense candidates as input	top-3/top-5
Reranked tables ( top_k )	3/5

Appendix

Table 12: Reranking hyperparameters.

Model	TP	Attn. backend	Ctx. len.	Max pref.	Chunk. pref.	Mem. frac.	Other
Qwen3-30B-A3B-Thinking-2507	4	FlashInfer	65,536	65,536	4,096	0.8	-
Qwen3-4B-Instruct-2507	2	FlashInfer	40,960	40,960	4,096	0.85	-
GLM-Z1-32B-0414	4	FlashInfer	32,768	32,768	4,096	0.8	-
Llama-3.1-8B-Instruct	2	Default	40,960	40,960	4,096	0.8	-

Appendix

Table 13: SGLang serving configuration for different models.

Model	TP	Mem. util.	Max len.	Other
Qwen3.5-9B	4	0.9	40,960	Default
Qwen3.6-35B-A3B	4	0.85	32,768	Default
Qwen3.6-27B	4	0.8	40,960	Default
TableGPT-R1	4	0.85	40,960	Default

Appendix

Table 14: vLLM serving configuration for different models.

Method	Config	Format	Recall	F1 (Qwen3-30B)	F1 (Qwen3-4B)	F1 (LLaMA)
Dense	E4B	CSV	0.158	0.116	0.054	0.079
Dense	E4B	Mixed	0.172	0.118	0.064	0.088
Dense	E4B	HTML	0.134	0.122	0.048	0.056
Dense	E4B	Markdown	0.187	0.134	0.077	0.109
Rerank (top-3)	E4B+R4B	CSV	0.354	0.284	0.102	0.173
Rerank (top-3)	E4B+R4B	Mixed	0.359	0.269	0.127	0.172

Appendix

Table 24: Full backbone comparison on TCR-Bench with Qwen3-30B, Qwen3-4B, and LLaMA. Recall is R@1 for Dense/Rerank and R@all for TableSeek.

Method	Config	Recall	F1
Oracle	GLM-Z1	1.000	0.501
Dense	E4B+GLM-Z1	0.172	0.099
Rerank (top-3)	E4B+R4B+GLM-Z1	0.359	0.197
Rerank (top-5)	E4B+R4B+GLM-Z1	0.421	0.264
TableSeek (5-step-5s1-small)	GLM-Z1	0.258	0.170
TableSeek (10-step-5s1-small)	GLM-Z1	0.306	0.180

Appendix

Table 25: Supplementary backbone comparison on TCR-Bench with additional models beyond the three primary backbones. Recall is R@1 for Dense/Rerank and R@all for TableSeek.

Method	Split	Config	Retrieval	Acc	Acc-Sup	Acc-Ref
Dense	test	Qwen3-EB+Qwen3-4B	0.495	0.628	0.407	0.851
Dense	test-1k	Qwen3-EB+Qwen3-4B	0.498	0.610	0.405	0.823
Dense	test	Stella+Qwen3-4B	0.520	0.637	0.435	0.840
Dense	test-1k	Stella+Qwen3-4B	0.519	0.627	0.446	0.815
Dense	test	Qwen3-EB+LLaMA	0.495	0.611	0.308	0.918
Dense	test-1k	Qwen3-EB+LLaMA	0.498	0.598	0.297	0.910

Appendix

Table 26: Results on TARGET-TabFact-test and TARGET-TabFact-test-1k. Retrieval is R@1 for Dense/Rerank and R@all for TableSeek.

Method	Split	Config	Retrieval	Acc	Acc-Sup	Acc-Ref
Dense	valid	Qwen3-EB+Qwen3-4B	0.492	0.633	0.415	0.857
Dense	valid-1k	Qwen3-EB+Qwen3-4B	0.515	0.610	0.395	0.814
Dense	valid	Stella+Qwen3-4B	0.516	0.639	0.434	0.849
Dense	valid-1k	Stella+Qwen3-4B	0.523	0.629	0.418	0.830
Dense	valid	Qwen3-EB+LLaMA	0.492	0.610	0.315	0.914
Dense	valid-1k	Qwen3-EB+LLaMA	0.515	0.610	0.303	0.902

Appendix

Table 27: Results on TARGET-TabFact-valid and TARGET-TabFact-valid-1k. Retrieval is R@1 for Dense/Rerank and R@all for TableSeek.

Comparison	Backbone LLM	F1 A	F1 B	$\Delta$	95% CI	$p$
TableSeek (10-step) vs Rerank (top-5)	Qwen3.6-27B	0.592	0.359	+0.233	[0.167, 0.302]	$<0.001$
TableSeek (5-step) vs Rerank (top-3)	Qwen3.6-27B	0.545	0.323	+0.222	[0.152, 0.294]	$<0.001$
TableSeek (10-step) vs Rerank (top-5)	Qwen3-30B	0.361	0.317	+0.045	[-0.028, 0.118]	0.229
TableSeek (5-step) vs Rerank (top-3)	Qwen3-30B	0.284	0.259	+0.024	[-0.044, 0.095]	0.482
TableSeek (10-step) vs TableSeek (5-step)	Qwen3-30B	0.361	0.284	+0.077	[0.028, 0.128]	0.002

Appendix

Table 28: Paired significance analysis on TCR-Bench using per-instance QA F1 scores. Bootstrap uses 10,000 resamples.

Model	CSV	Markdown	HTML	Mixed
Qwen3-4B	0.239	0.301	0.266	0.270
LLaMA	0.491	0.485	0.417	0.465
Qwen3-30B	0.717	0.748	0.719	0.734

Appendix

Table 29: Oracle QA results (F1) on TCR-Bench under different table formats.

Model	Split	Acc	Acc-Sup	Acc-Ref
Qwen3-4B	valid	0.732	0.690	0.775
Qwen3-4B	valid-1k	0.723	0.683	0.764
Qwen3-4B	test	0.695	0.654	0.737
Qwen3-4B	test-1k	0.695	0.654	0.737
LLaMA	valid	0.699	0.528	0.875
LLaMA	valid-1k	0.684	0.488	0.871

Appendix

Table 30: Oracle QA results on TabFact.

Method	Model(s)	Format	R@1	R@2	R@3
Dense	Qwen3-Embedding-4B	CSV	0.158	0.311	0.464
Dense	Qwen3-Embedding-4B	Mixed	0.172	0.321	0.469
Dense	Qwen3-Embedding-4B	HTML	0.134	0.287	0.474
Dense	Qwen3-Embedding-4B	Markdown	0.187	0.340	0.512
Dense	jina-embeddings-v4	CSV	0.129	0.282	0.368
Dense	jina-embeddings-v4	Mixed	0.163	0.306	0.426

Appendix

Table 31: Retrieval-only results on TCR-Bench.

Method	Model(s)	Split	R@1	R@2	R@3
Dense	Qwen3-Embedding-4B	test	0.495	0.596	0.649
Dense	Qwen3-Embedding-4B	test-1k	0.498	0.589	0.646
Dense	jina-embeddings-v4	test	0.439	0.528	0.579
Dense	jina-embeddings-v4	test-1k	0.450	0.542	0.593
Dense	stella_en_1.5B_v5	test	0.520	0.615	0.664
Dense	stella_en_1.5B_v5	test-1k	0.519	0.614	0.668

Appendix

Table 32: Retrieval-only results on TARGET-TabFact-test and TARGET-TabFact-test-1k.

Method	Model(s)	Split	R@1	R@2	R@3
Dense	Qwen3-Embedding-4B	valid	0.492	0.592	0.647
Dense	Qwen3-Embedding-4B	valid-1k	0.515	0.606	0.641
Dense	jina-embeddings-v4	valid	0.428	0.519	0.570
Dense	jina-embeddings-v4	valid-1k	0.432	0.522	0.579
Dense	stella_en_1.5B_v5	valid	0.516	0.618	0.668
Dense	stella_en_1.5B_v5	valid-1k	0.523	0.611	0.661

Appendix

Table 33: Retrieval-only results on TARGET-TabFact-valid and TARGET-TabFact-valid-1k.

Method	Dense Calls	Rerank Calls	LLM Calls	Total Calls	Dense Time	Rerank Time	LLM Time	Total Time
Dense (Qwen3-4B)	846	0	209	1055	0.940	0	0.125	1.065
Rerank-3 (Qwen3-4B)	846	627	209	1682	0.940	0.964	0.119	2.023
Rerank-5 (Qwen3-4B)	846	1045	209	2100	0.940	1.597	0.122	2.659
Dense (Qwen3-30B)	846	0	209	1055	0.940	0	0.765	1.705
Rerank-3 (Qwen3-30B)	846	627	209	1682	0.940	0.964	0.711	2.615
Rerank-5 (Qwen3-30B)	846	1045	209	2100	0.940	1.597	0.704	3.241

Appendix

Table 34: Number of inference calls and wall-clock runtime (hours) on TCR-Bench.

Strategy	R@all	F1	Avg. Steps	Time (h)
5-step setting
Full	0.416	0.284	4.407	2.335
w/o refine	0.239	0.173	3.555	1.733
w/o regenerate	0.254	0.190	4.148	1.921
w/o select	0.402	0.305	4.909	3.031
10-step setting

Appendix

Table 35: Fine-grained ablation of the refine mechanism, decomposed into regenerate and select, on TCR-Bench (Mixed format) using Qwen3-30B.

Round	Qwen3-30B		Qwen3-4B		Qwen3-30B (TabFact)		GLM-Z1	Qwen3.5-9B
Round	5-step	10-step	5-step	10-step	5-step	10-step	10-step	10-step
0	0.206	0.201	0.153	0.153	0.441	0.444	0.062	0.306
1	0.301	0.301	0.158	0.158	0.488	0.486	0.206	0.349
2	0.311	0.330	0.158	0.158	0.498	0.493	0.206	0.359
3	0.330	0.340	0.163	0.163	0.505	0.493	0.210	0.359
4	–	0.364	–	0.158	–	0.506	0.220	0.364

Appendix

Table 36: Top-1 hit rate by refinement round, for two independent runs (5-step and 10-step budgets) of each model configuration.

Strategy	R@all	F1	Eff	Avg. Steps	Time (h)
5step-5s1	0.416	0.284	0.683	4.407	2.335
5step-5s3	0.445	0.307	0.690	4.407	2.609
5step-10s1	0.397	0.294	0.741	4.383	2.807
5step-10s3	0.431	0.297	0.689	4.407	3.129
5step-10s5	0.469	0.316	0.674	4.354	3.259
10step-5s1	0.507	0.361	0.712	6.244	3.466

Appendix

Table 37: Effect of refine strategy on TCR-Bench (Mixed format), using Qwen3-30B. Eff

=

F1/R@all.

Strategy	R@all	F1	Eff	Avg. Steps	Time (h)
5step-r3	0.378	0.269	0.712	4.431	2.172
5step-r5	0.416	0.291	0.700	4.459	2.580
5step-r10	0.416	0.307	0.738	4.349	2.850
10step-r3	0.498	0.375	0.753	6.196	3.174
10step-r5	0.550	0.384	0.698	6.067	3.535
10step-r10	0.517	0.374	0.723	6.038	4.069

Appendix

Table 38: Effect of the rerank-based refine strategy (r

n

) on TCR-Bench (Mixed format), using Qwen3-30B. Eff

=

F1/R@all.

Format	10step-5s1	10step-5s3	10step-r5
CSV	0.346	0.367	0.362
Mixed	0.361	0.352	0.384
HTML	0.364	0.356	0.352
Markdown	0.364	0.389	0.346

Appendix

Table 39: F1 of selected 10-step refine strategies across table formats on TCR-Bench, using Qwen3-30B.

Method	Dense	Rerank (top-3)	Rerank (top-5)	TableSeek (5step)	TableSeek (10step)
CSV	0.116	0.284	0.356	0.283	0.346
Mixed	0.118	0.269	0.317	0.284	0.361
HTML	0.122	0.279	0.353	0.292	0.364
Markdown	0.134	0.314	0.362	0.318	0.364
Mean	0.123	0.287	0.347	0.294	0.359
Range	0.018	0.045	0.045	0.035	0.018

Appendix

Table 40: F1 across table formats and robustness statistics on TCR-Bench, using Qwen3-30B.

	TableSeek (10-step)				Rerank (top-5)
Format	Run1	Run2	Run3	Avg	Run1	Run2	Run3	Avg
CSV	0.346	0.340	0.340	0.342	0.356	0.351	0.352	0.353
Mixed	0.361	0.360	0.363	0.362	0.317	0.312	0.317	0.315
HTML	0.364	0.364	0.364	0.364	0.353	0.353	0.353	0.353
Markdown	0.364	0.358	0.375	0.366	0.362	0.362	0.362	0.362
Mean	–	–	–	0.358	–	–	–	0.346

Appendix

Table 41: Multi-run robustness analysis for Rerank (top-5) and TableSeek (10-step). Robustness statistics are computed from the format-wise average F1 scores across three independent runs.

Method	10-step-5s3	10-step-r5
CSV	0.367	0.362
Mixed	0.352	0.384
HTML	0.356	0.352
Markdown	0.389	0.346
Mean	0.366	0.361
Range	0.037	0.038

Appendix

Table 42: Robustness analysis of alternative refine strategies.

Configuration	R@All	F1	Avg. Steps	Time (h)
10-step setting
full	0.512	0.371	6.153	10.079
big	0.502	0.370	6.139	3.498
medium	0.493	0.372	6.340	3.606
small	0.507	0.361	6.244	3.466
5-step setting

Appendix

Table 43: Effect of retrieved evidence size on TCR-Bench (Mixed format) using Qwen3-30B. Larger evidence windows generally improve retrieval performance but incur substantially higher runtime costs.

Method	Config	Recall T	F1 T
Oracle	Qwen3-30B	1.000	0.269
Dense	E4B+Qwen3-30B	0.081	0.048
Rerank (top-3)	E4B+R0.6B+Qwen3-30B	0.187	0.076
Rerank (top-5)	E4B+R0.6B+Qwen3-30B	0.239	0.100
TableSeek (5-step-5s1-small)	Qwen3-30B	0.263	0.110
TableSeek (10-step-5s1-small)	Qwen3-30B	0.335	0.139

Appendix

Table 44: Results on the transposed-table version of TCR-Bench.

Explore similar work

Semantically Similar, Yet Not Answerable: Diagnosing the Semantic-Answerability Gap in Table RAG
Jul 20, 2026Jiaming Tian, Liyao Li, Wentao Ye +6Hievi-Rag Tabular Data
Zhejiang University · Bank of Hangzhou Co., Ltd. · Zhejiang Lab
FollowTable: A Benchmark for Instruction-Following Table Retrieval
May 1, 2026Rihui Jin, Yuchen Lu, Ting Zhang +7Tabular Data Retrievers
Southeast University Nanjing, China Key Laboratory of New Generation Artificial Intelligence Technology and Its Interdisciplinary Applications (Southeast University), Ministry of Education · Huawei Noah’s Ark Lab Shenzhen, China
TabScope: Question-Adaptive Scope Selection for Table Question Answering
Sep 3, 2026Yuxiang Wang, Junhao Gan, Jianzhong QiTabular Data Question
The University of Melbourne

Jul 20, 2026, cs.AIJ/K move · Enter open · S save

Semantically Similar, Yet Not Answerable: Diagnosing the Semantic-Answerability Gap in Table RAG

Jiaming Tian, Liyao Li, Wentao Ye, Haobo Wang+5

Zhejiang University · Bank of Hangzhou Co., Ltd. · Zhejiang Lab

Hievi-Rag Tabular Data Answer Semantic Gap Contextual Cues

Open paper AskPDF arXiv

In retrieval-augmented generation (RAG), semantic relevance asks whether a source matches a query in meaning, while answerability asks whether it contains sufficient information to answer the query. A Semantic-Answerability Gap (SAG) may arise in retrieval when a retriever can reach semantically relevant sources yet fail to identify those that are uniquely answerable. We uncover this gap using tables as a controlled setting, where shared schemas and entities provide strong semantic signals while localized content and row-column bindings distinguish answerable from non-answerable sources. Using TCR-Bench, a controlled sibling-table benchmark, we find that dense retrievers achieve only 18.2% top-1 target retrieval, reducing QA F1 from 0.755 with the oracle table to 0.330 with retrieved top-5 tables. Controlled diagnostics show that retrievers favor semantic volume over sufficiency, respond weakly to row-column binding disruptions, and struggle to distinguish Targets from Siblings. Explicit answerability assessment substantially improves target identification, while fine-tuning shows that answerability is learnable but difficult to transfer without compromising broad semantic retrieval.

Abstract

Explore similar work

Semantically Similar, Yet Not Answerable: Diagnosing the Semantic-Answerability Gap in Table RAG

FollowTable: A Benchmark for Instruction-Following Table Retrieval

TabScope: Question-Adaptive Scope Selection for Table Question Answering