Deep Tabular Research via Continual Experience-Driven Execution

Junnan Dong; Chuang Zhou; Zheng Yuan; Yifei Yu; Qiufeng Wang; Yinghui Li; Siyu An; Di Yin

cs.AIMar 10, 2026

Deep Tabular Research via Continual Experience-Driven Execution

Authors: Junnan Dong, Chuang Zhou, Zheng Yuan, Yifei Yu, Qiufeng Wang, Yinghui Li, Siyu An, Di Yin,

Organizations: Tencent Youtu Lab · Ruijin Hospital, Shanghai Jiaotong University

Abstract

Large language models often struggle with complex long-horizon analytical tasks over unstructured tables, which typically feature hierarchical and bidirectional headers and non-canonical layouts. We formalize this challenge as Deep Tabular Research (DTR), requiring multi-step reasoning over interdependent table regions. To address DTR, we propose a novel agentic framework that treats tabular reasoning as a closed-loop decision-making process. We carefully design a coupled query and table comprehension for path decision making and operational execution. Specifically, (i) DTR first constructs a hierarchical meta graph to capture bidirectional semantics, mapping natural language queries into an operation-level search space; (ii) To navigate this space, we introduce an expectation-aware selection policy that prioritizes high-utility execution paths; (iii) Crucially, historical execution outcomes are synthesized into a siamese structured memory, i.e., parameterized updates and abstracted texts, enabling continual refinement. Extensive experiments on challenging unstructured tabular benchmarks verify the effectiveness and highlight the necessity of separating strategic planning from low-level execution for long-horizon tabular reasoning.

Figures & tables

Figure 1: Existing Table QA pipelines (a) are limited to well-structured tables and shallow queries, and fail to handle unstructured tabular properties (b) and long-horizon analytical tasks (c), motivating our Deep Tabular Research.

Figure 2: A sketched overview for our proposed Deep Tabular Research framework for complex unstructured tabular reasoning. DTR decomposes analytical intent into meta operations, plans executable macro paths via expectation-guided search, and executes them with an experience-aware memory that records structured execution feedback and updates planning policies across iterations.

Methods	Accuracy		Analysis Depth		Feasibility		Aesthetics		Avg. Runtime (s) $\downarrow$	Total Output Tokens $\downarrow$	Avg. LLM Calls $\downarrow$
	Win Rate	Score Rate	Win Rate	Score Rate	Win Rate	Score Rate	Win Rate	Score Rate
	(Tie=0) $\uparrow$	(Tie=0.5) $\uparrow$	(Tie=0) $\uparrow$	(Tie=0.5) $\uparrow$	(Tie=0) $\uparrow$	(Tie=0.5) $\uparrow$	(Tie=0) $\uparrow$	(Tie=0.5) $\uparrow$
Table specific LLM
TableGPT2-7B	0.20	8.41	5.12	5.12	4.33	4.35	6.21	6.21	3.42	12,450	1.0
TableLLM-7B	0.15	6.22	3.84	3.84	3.10	3.12	4.55	4.55	3.15	10,820	1.0
StructGPT	0.10	4.15	2.10	2.10	1.85	1.86	2.30	2.30	4.88	8,420	1.0

Table 1: Full Comparisons over DTR-Bench to evaluate over accuracy, quality, and efficiency.

Methods	Fact Checking		Numerical Reasoning		Structure Comprehending		Data Analysis		Chart/Report Generation
Methods	EM $\uparrow$	F1 $\uparrow$	EM $\uparrow$	F1 $\uparrow$	EM $\uparrow$	F1 $\uparrow$	LLM-EVAL $\uparrow$	ROUGE $\uparrow$	PASS@1 $\uparrow$	ECR $\uparrow$
TableGPT2-7B	46.10	53.80	29.31	39.81	48.23	56.68	62.76	33.25	32.47	67.53
TableLLM-7B	38.25	44.12	22.40	31.65	41.10	49.34	55.80	28.42	18.20	42.15
StructGPT	25.40	32.15	14.55	20.80	30.25	38.60	42.33	19.50	5.12	12.44
GPT4o	43.39	51.87	27.63	36.68	42.68	52.89	65.24	33.10	10.39	25.32
DeepSeek-v3	57.21	53.42	47.05	50.61	43.31	74.63	61.40	34.76	9.09	24.68

Table 2: Comparisons over the RealHitBench dataset for five different task types.

Configuration	Meta	QDO	Exp.	Abst.	Acc.	Anal.	Feas.	Aesth.
+ Meta Info	✓	$\times$	$\times$	$\times$	34.8	27.1	25.6	38.2
+ QDO	✓	✓	$\times$	$\times$	36.2	28.8	26.7	40.5
+ Expectation	✓	✓	✓	$\times$	37.1	29.6	27.2	41.8
DTR Full	✓	✓	✓	✓	37.5	30.2	27.6	42.7

Table 3: Architecture ablation study of each components in DTR.

Prompt Strategy	Accuracy	Analysis	Code Error	Avg
		Depth	Rate	Calls
No [THINK] (Direct)	35.2	27.8	42.3%	5.8
Simple [THINK] Hint	36.4	28.9	35.6%	5.2
[THINK]+[CODE]	37.5	30.2	28.4%	4.78
Multi-stage Reflection	37.2	29.8	26.1%	5.5

Table 4: Exploration over prompting strategy for think and code.

Figure 3: LLM call budget analysis. The blue curve shows performance (left y-axis) and the purple curve shows marginal gain (right y-axis). DTR’s 4.78-call configuration red star) achieves optimal efficiency by avoiding the plateau region.

Figure 4: Path selection evolution across 10 batches. Colors changing from light blue to deep purple indicate both exploration and exploitation, respectively.

Appendix figures & tables7 assets

Supplementary material from the paper’s appendix.

Appendix

\mathbb{E}(\pi)=\hat{R}(\pi)+c\cdot P(\pi)\sqrt{\frac{\log\sum_{\pi^{\prime}}N(\pi^{\prime})}{1+N(\pi)}}

Appendix

Algorithm 1 Deep Tabular Research

User Persona	Questions	Primary Analysis Focus
Social Researcher	105	Inequality analysis (Gini coefficient, Lorenz curves), demographic patterns
Student/Researcher	88	Statistical hypothesis testing (ANOVA), effect size calculation (Cohen’s $d$ , $\eta^{2}$ )
Government Staff	62	Longitudinal policy impact assessment, equity-efficiency trade-offs
Data Analyst	61	Dimensionality reduction, cohort analysis, z-score-based anomaly detection
Business Owner	55	Profitability decomposition, BCG matrix classification, break-even analysis
Procurement Manager	53	Supplier concentration (HHI), risk-performance matrices

Appendix

Table 5: User Persona Distribution in DTR-Bench

Persona	Analysis Type	Template Example
Investor	Risk-Return Analysis	Conduct a risk-return analysis across all {Category} by examining the relationship between {NumericCol1} (returns) and {NumericCol2} (risk metrics). Identify the efficient frontier segments.
Social Researcher	Inequality Analysis	Measure the distribution of {NumericCol} across {Category} using Gini coefficient and decile ratios. Investigate structural factors driving inequality.
Data Analyst	Anomaly Detection	Use statistical methods (z-scores) to identify unusual patterns in {NumericCol} across {Category} . Investigate anomalies and assess impact.
Student/Researcher	Statistical Testing	Conduct rigorous statistical analysis of {NumericCol} across {Category} : test for normality, perform ANOVA/Kruskal-Wallis tests, and calculate effect sizes.

Appendix

Table 6: Representative Question Templates by Persona

Figure 5: Employment Distribution in 1984: Agriculture (4.5%) vs Non-Agriculture (95.5%)

Figure 6: Employment Trends 1983-1990: Employed population grew from 56.8M to 65.1M while unemployed declined from 6.3M to 3.9M

Figure 7: Scatter plot showing the relationship between unemployment rate and civilian labor force size over 16 years (1983-1998)

Table QA

Table question answering studies how AI systems retrieve, interpret, and reason over structured data to answer natural-language questions.

See all 66 papers

From Table to Cell: Attention for Better Reasoning with TABALIGN
May 14, 2026Tung Sum Thomas Kwok, Zeyong Zhang, Xinyu Wang +6Table QA Table Reasoning
University of California, Los Angeles · SimpleWay · New Jersey Institute of Technology +4
H2Table: Hierarchical Hypergraph-Enhanced Large Language Models for Complex Table Reasoning
Sep 1, 2026Jia Ling, Yangfan Wang, Chen Tang +4Table QA Hierarchical Representation Learning
Harbin Institute of Technology · AI Research Center, Midea Group (Shanghai) Co., Ltd. · Changchun University of Science and Technology +1
TabScope: Question-Adaptive Scope Selection for Table Question Answering
Sep 3, 2026Yuxiang Wang, Junhao Gan, Jianzhong QiTable QA
The University of Melbourne

May 14, 2026J/K move · Enter open · S save

From Table to Cell: Attention for Better Reasoning with TABALIGN

Tung Sum Thomas Kwok, Zeyong Zhang, Xinyu Wang, Chunhe Wang+5

University of California, Los Angeles · SimpleWay · New Jersey Institute of Technology+4

Table QA Table Reasoning Structured Reasoning Diffusion Language Models LLM Reasoning

Open paper AskPDF arXiv

Multi-step LLM reasoning over structured tables fails because planning and execution share no explicit cell-grounding contract. Existing methods constrain the planner to a left-to-right factorization at odds with table permutation invariance, and score intermediate states by generated content alone, overlooking cell grounding. We conduct a pilot study showing that diffusion language models (DLMs) produce more human-aligned and permutation-stable cell attention on tables than autoregressive models, with a 40.2% median reduction in attention-AUROC variability under row reordering. Motivated by this, we propose TABALIGN, a planned table reasoning framework that operationalizes the contract. TABALIGN pairs a masked DLM planner, whose bidirectional denoising emits plan steps as binary cell masks, with TABATTN, a lightweight verifier trained on 1,600 human-verified attention standards to score each step by its attention overlap with the plan-designated mask. Across eight benchmarks covering table question answering and fact verification, TABALIGN improves average accuracy by 15.76 percentage points over the strongest open-source baseline at comparable 8B-class scale, with a matched-backbone ablation attributing 2.87 percentage points of this gain to the DLM planner over an AR planner on a fixed reasoner. Cleaner DLM plans also accelerate downstream reasoning execution by 44.64%.

Abstract

Table QA

Explore similar work

From Table to Cell: Attention for Better Reasoning with TABALIGN

H2Table: Hierarchical Hypergraph-Enhanced Large Language Models for Complex Table Reasoning

TabScope: Question-Adaptive Scope Selection for Table Question Answering