Replacing Multi-Step Assembly of Data Preparation Pipelines with One-Step LLM Pipeline Generation for Table QA

Fengyu Li; Junhao Zhu; Kaishi Song; Lu Chen; Zhongming Yao; Tianyi Li; Christian S. Jensen

cs.DBFeb 26, 2026

Replacing Multi-Step Assembly of Data Preparation Pipelines with One-Step LLM Pipeline Generation for Table QA

Authors: Fengyu Li, Junhao Zhu, Kaishi Song, Lu Chen, Zhongming Yao, Tianyi Li, Christian S. Jensen

Organizations: Zhejiang University Hangzhou, China · Northeastern University Shenyang, China · Aalborg University Aalborg, Denmark

Abstract

Table Question Answering (TQA) aims to answer natural language questions over structured tables. Large Language Models (LLMs) enable promising solutions to this problem, with operator-centric solutions that generate table manipulation pipelines in a multi-step manner offering state-of-the-art performance. However, these solutions rely on multiple LLM calls, resulting in prohibitive latencies and computational costs. We propose Operation-R1, the first framework that trains lightweight LLMs (e.g., Qwen-4B/1.7B) via a novel variant of reinforcement learning with verifiable rewards to produce high-quality data-preparation pipelines for TQA in a single inference step. To train such an LLM, we first introduce a self-supervised rewarding mechanism to automatically obtain fine-grained pipeline-wise supervision signals for LLM training. We also propose variance-aware group resampling to mitigate training instability. To further enhance robustness of pipeline generation, we develop two complementary mechanisms: operation merge, which filters spurious operations through multi-candidate consensus, and adaptive rollback, which offers runtime protection against information loss in data transformation. Experiments on two benchmark datasets show that, with the same LLM backbone, Operation-R1 achieves average absolute accuracy gains of 8.83 and 4.44 percentage points over multi-step preparation baselines, with 79% table compression and a 2.2 $\times$ reduction in monetary cost.

Figures & tables

Figure 1. (a) Direct text-based reasoning, (b) Program-based reasoning, (c) Operation-based multi-agent framwork, and (d) Pipeline-specific model.

Figure 2. Overview of Operation-R1 , including ORPO-based offline training, parallel operator generation, operation-tree merging, and adaptive rollback for reliable tabular QA.

Method	Base Model	Train Data Size	Finetuning Method	WikiTQ	TabFact
TableGPT2 †	Qwen2.5-72B	2.36M	Full-Parameter	71.5	82.5
Table-R1-SFT (Yang et al.) †	Llama-3.1-8B-Instruct	34K	Full-Parameter	83.8	91.1
Table-R1-SFT (Yang et al.) †	Qwen2.5-7B-Instruct	34K	Full-Parameter	81.9	89.9
Table-R1-ZERO (Yang et al.) †	Llama-3.1-8B-Instruct	49K	Full-Parameter	81.2	87.6
Table-R1-ZERO (Yang et al.) †	Qwen2.5-7B-Instruct	49K	Full-Parameter	79.8	87.6
Table-R1 (Wu et al.) †	Qwen3-8B (Thinking mode)	20K	Full-Parameter	73.87	-

Table 2. Comparison of End2End Training-based QA Methods. (

\dagger

: results from the original studies due to unavailable model releases or GPU limits).

Figure 3. Efficiency-Accuracy-Cost Trade-off Comparison.

Figure 4. Training curves of Op-R1-1.7B with and without VGR strategy. Removing VGR causes a pronounced early drop in reward and slower subsequent improvement, whereas the VGR yields smoother, steadily increasing curves and higher rewards.

Method	Qwen2.5-7B		Qwen3-4B(Thk)
Method	WikiTQ	TabFact	WikiTQ	TabFact
No-Prep	50.69	71.99	77.46	92.89
Operation-R1	69.27	88.34	81.63	93.47
$-$ ORPO & VGR strategy	63.49	85.67	73.87	89.53
$-$ VGR strategy	68.67	86.35	80.57	93.28
$-$ Operation Merge	68.16	85.67	76.84	88.49

Table 3. Ablation Studies on Two Backbone Models

Dataset	Question Type	Method	Performance
TableBench	Fact checking	No Prep	57.29
		CoTable	41.67
		AutoPrep	65.62
		Operation-R1	77.08
	Numerical reasoning	No Prep.	11.59
		CoTable	29.22

Table 4. Performance comparison on additional types of questions on TableBench and FeTaQA.

Dataset	Type	Prop.	Method	Accuracy (%)
WikiTQ	CF	46.1%	No-Prep	63.29
			CoTable	56.72
			AutoPrep	60.77
			Operation-R1	77.51
	NCF	53.9%	No-Prep	39.92
			CoTable	48.28

Table 5. Performance comparison on Cell-Focused (CF) vs. Non-Cell-Focused (NCF) tasks.

Dataset	Method	Prepare	Extra	QA	Total Calls
WikiTQ	CoTable	12537	197	4344	17078
	AutoPrep	9691	8109	5248	20965
	Operation-R1	4344	134	4831	9309
TabFact	CoTable	6258	44	2024	8326
	AutoPrep	3593	3780	2318	9691
	Operation-R1	2024	56	2376	4456

Table 6. LLM invocation comparison across Operation-R1 and baselines.

Dataset	Usage (%)	Accuracy (%)
Dataset	Usage (%)	No Prep	ours	ours w/ additional $op$
Subset requiring aggregate operator ( $op=op_{\text{aggr}}$ )
TableBench-NR	52.64	5.26	37.80	69.86 (+32.06)
TableBench-FV	13.54	46.16	61.54	69.23 (+7.69)
Subset requiring clean_column operator ( $op=op_{\text{clean}}$ )
WikiTQ	13.63	45.43	52.36	56.93 (+4.57)

Table 7. Accuracy comparison on subsets of datasets where the additional operator unseen in training is required for TableQA.

Training set	Query Num.	WikiTQ Acc.	TabFact Acc.
WikiTQ	5403	69.66	88.44
TabFact	5000	69.77	88.49

Table 8. The evaluation results of Operation-R1 trained on different datasets (Accuracy %).

Appendix figures & tables8 assets

Supplementary material from the paper’s appendix.

Appendix

Benchmark	Task Type	CF (%)
WikiTQ	Short-form QA	46.1
HybridQA	Hybrid QA	40.4
SQA	Sequential QA	100
KET-QA	KB-Augmented QA	44.0
WikiSQL	NL2SQL	71.1

Appendix

Table 9. Quantitative breakdown of popular TQA benchmarks by cell-focused (CF) instances.

Figure 5. Impact of Operator Sequence Length on WikiTQ and TabFact.

Figure 6. Accuracy by Operator Type on WikiTQ and TabFact.

Dataset	No trigger	Full	Step3	Step2	Step1
WikiTQ	4001	343	7	92	45
TabFact	1715	309	2	54	37

Appendix

Table 10. Rollback trigger statistics of Operation-R1 during inference. Step1–Step3 denote the pipeline length (number of operators) when rollback is triggered.

Dataset	Strategy	Accuracy	Extra LLM Calls
WikiTQ	Full-First-Origin	69.27	431
	Binary rollback (ours)	69.66	487
	Step-by-step rollback	68.95	941
TabFact	Full-First-Origin	88.34	338
	Binary rollback (ours)	88.44	402
	Step-by-step rollback	87.35	483

Appendix

Table 11. Comparison of rollback strategies (Accuracy %).

Figure 7. Impact of parallel inference count on Operation Merge.

Figure 8. Accuracy comparison across different table lengths (measured by serialized token count).

Method	Model	WikiTQ		TabFact
Method	Model	Acc. (%)	Time (s)	Acc. (%)	Time (s)
CoTable	Qwen3-4B	57.50	8008	86.46	4743
CoTable	Qwen3-4B (Thk)	62.27	65837	89.62	25392

Appendix

Table 12. Performance comparison using Qwen3-4B under thinking and non-thinking modes.

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

Efficient Table QA via TableGrid Navigation and Progressive Inference Prompting
May 18, 2026Amritansh Maurya, Navjot Singh, Mohammed Javed +1Table QA Structured Prompting
Vision Intelligence Lab, IIIT Allahabad, Prayagraj, India · iMeDIA Lab, IIIT Allahabad, Prayagraj, India · CV:HCI Lab, Karlsruhe Institute of Technology, Karlsruhe, Germany
Evolving from Lessons: Skill-Augmented Table Graph Reasoning for Operation-wise Table Question Answering
Jun 20, 2026Guixin Su, Qiankun Pi, Mayi Xu +5Table QA
School of Computer Science, Wuhan University, Wuhan, China
TabScope: Question-Adaptive Scope Selection for Table Question Answering
Sep 3, 2026Yuxiang Wang, Junhao Gan, Jianzhong QiTable QA
The University of Melbourne

May 18, 2026J/K move · Enter open · S save

Efficient Table QA via TableGrid Navigation and Progressive Inference Prompting

Amritansh Maurya, Navjot Singh, Mohammed Javed, Omar Moured

Vision Intelligence Lab, IIIT Allahabad, Prayagraj, India · iMeDIA Lab, IIIT Allahabad, Prayagraj, India · CV:HCI Lab, Karlsruhe Institute of Technology, Karlsruhe, Germany

Table QA Structured Prompting LLM Prompting Prompt Engineering

Open paper AskPDF arXiv

Large Language Models (LLMs) have shown promising results on NLP tasks, however, their performance on tabular data still needs research attention, because Table Question-Answering (TQA) requires precise cell retrieval and multi-step structured reasoning. Existing work improves TQA either by fine-tuning or training LLMs on task-specific tabular data, but often lacks verifiable control over how the model navigates tables and derives answers. In this work, we propose a training-free TQA approach with two structured prompting frameworks: TableGrid Navigation (TGN), which iteratively navigates rows and columns via a three-module loop to locate evidence and refine answers, and Progressive Inference Prompting (PIP), which enforces columns identification for explicit progressive row selection constraint according to the query. We evaluate 17 LLMs against 6 baselines on TableBench and FeTaQa dataset. On TableBench, TGN improves over the strongest baseline by 3.8 points, and on FeTaQa, PIP achieves SOTA performance over ReAct and Chain-of-Thought. Beyond inference-time gains, PIP and TGN can also serve as supervision templates to fine-tune small models, narrowing the performance gap to much larger architectures in resource-constrained settings, offering versatile and cost-efficient solution for TQA.

Abstract

Table QA

Explore similar work

Efficient Table QA via TableGrid Navigation and Progressive Inference Prompting

Evolving from Lessons: Skill-Augmented Table Graph Reasoning for Operation-wise Table Question Answering

TabScope: Question-Adaptive Scope Selection for Table Question Answering