QUIETT: Query-Independent Table Transformation for Robust Reasoning

Gaurav Najpande; Tampu Ravi Kumar; Manan Roy Choudhury; Neha Valeti; Yanjie Fu; Vivek Gupta

cs.CLFeb 23, 2026

QUIETT: Query-Independent Table Transformation for Robust Reasoning

Authors: Gaurav Najpande, Tampu Ravi Kumar, Manan Roy Choudhury, Neha Valeti, Yanjie Fu, Vivek Gupta

Organizations: Arizona State University

Abstract

Real-world tables often contain schema inconsistencies, heterogeneous value formats, and implicit relational structures that degrade table reasoning and question answering. Existing approaches address these issues at query time, repeating normalization and restructuring for every new question and producing representations that do not generalize to unseen queries. We introduce QuIeTT, a transform-first, query-later framework that converts each raw table into a single SQL-ready canonical representation before any evaluation query is observed, under three constraints: no evaluation query access, information preservation, and single-table reuse. QuIeTT operates through a three-stage pipeline: issue probing, where synthetic queries surface structural deficiencies such as ambiguous schemas and heterogeneous formats; plan generation; and structured execution. Experiments on four benchmarks across five model families show consistent improvements over strong baselines in both direct prompting and agentic reasoning paradigms, with particularly strong gains on a manually annotated challenge set of structurally diverse unseen questions.

Figures & tables

Figure 1: QuIeTT transforms a raw table into a single SQL-ready canonical representation before any evaluation query is observed. The transformed table is reused across all queries without modification.

Figure 2: QuIeTT framework for robust table reasoning. Stage 1: Issue Probing. Synthetic queries are generated from the raw table to expose structural issues such as inconsistent formats, implicit relations, missing derived fields, and undecomposed attributes. Stage 2: Transformation. Detected issues guide a transformation plan that applies normalization, parsing, restructuring, and derived column creation, executed via code to materialize a canonical, SQL-ready table. Stage 3: QA. The transformed table is reused across all downstream queries, enabling accurate and consistent reasoning without modifying the underlying model or prompting strategy.

Expert Pair	Cohen’s Kappa	Jaccard Coeff.
$\alpha_{1,2}$	92.9	96.9
$\alpha_{2,3}$	92.0	96.6
$\alpha_{1,3}$	94.9	97.7
$\alpha_{1,2,3}$ (Fleiss’ $\kappa$ )	0.9328	0.9562

Table 1: Inter-rater agreement for challenge set annotation.

Statistic	Values	Table Dimension
Total tables	529	Mean: $(20,\,11)$
Total questions	2,500	Median: $(8,\,8)$
Questions per table	4 / 5 / 7	Min: $(1,\,1)$
(mean / median / max)		Max: $(380,\,76)$

Table 2: Challenge set composition and table structure statistics. The set comprises 2,500 questions over 529 tables. Tuple-valued entries in the Table Dimension column report (rows, columns); the wide size range reflects the structural diversity the challenge set targets.

		Gemini 2.5								GPT-OSS
Category	Method	WikiTQ		NQ-Table		SequentialQA		HiTab		WikiTQ		NQ-Table		SequentialQA		HiTab
Category	Method	F1	EM	F1	EM	F1	EM	F1	EM	F1	EM	F1	EM	F1	EM	F1	EM
Direct	CoT	74.31	72.85	71.22	66.36	61.98	57.21	75.21	71.00	69.40	66.56	66.60	63.54	60.26	58.29	53.90	50.80
	CoT + SQL	54.66	52.49	36.64	30.96	58.23	53.37	69.01	67.00	43.00	42.29	47.90	33.95	55.20	49.60	31.90	30.70
	NormTab	58.70	56.19	53.00	51.05	50.00	49.92	56.81	54.60	65.91	62.61	50.30	48.34	48.60	43.10	54.67	52.00
	E5	74.35	70.30	55.90	52.92	31.64	29.81	78.89	75.10	56.90	52.15	54.00	51.76	43.80	40.46	77.47	75.00

Table 3: Main results. QuIeTT is compared against direct prompting methods and agentic methods across four benchmarks and four model families. Top half: Gemini 2.5 and GPT-OSS; bottom half: DeepSeek V3.1 and Qwen 3. Bold indicates best result per column. Gemini 2.0 results are in Table 12 .

Dataset	Method	Gemini 2.5		GPT-OSS
		F1	EM	F1	EM
WikiTQ	CoT	67.11	65.50	63.58	60.36
	NormTab	62.28	61.90	61.71	57.23
	EEDP	67.60	66.34	65.12	61.86
	Binder	65.56	63.39	65.12	64.48
	QuIeTT	74.41	71.37	70.06	68.59

Table 4: Challenge set results

Figure 3: Table-level performance outcomes. Win, No Change, and Loss indicate whether QuIeTT improves, leaves unchanged, or degrades per-table QA performance relative to the raw table baseline.

Figure 4: Error Analysis for degraded tables.

Dataset	Gemini 2.5		GPT-OSS
Dataset	F1	EM	F1	EM
WikiTQ	53.22	52.06	51.40	51.15
NQ-Table	56.80	54.37	54.90	51.82

Table 5: QuIeTT performance without issue probing.

Dataset	Gemini 2.5		GPT-OSS
Dataset	F1	EM	F1	EM
WikiTQ	60.57	58.26	58.33	57.29
NQ-Table	63.20	60.56	61.08	58.66

Table 6: QuIeTT performance when planning and execution are merged into a single-agent transformation pipeline.

Dataset	Gemini 2.0		Gemini 2.5		DeepSeek V3.1		Qwen 3		GPT-OSS
	F1	EM	F1	EM	F1	EM	F1	EM	F1	EM
WikiTQ	37.61	36.34	38.92	35.12	27.65	26.78	27.19	23.50	28.21	24.59
NQ-Table	35.42	33.32	32.49	32.03	25.38	24.23	35.31	31.21	32.44	30.46
SequentialQA	40.71	38.90	42.97	36.32	38.98	35.33	36.67	31.12	41.98	38.48
HiTab	59.60	56.80	62.30	59.60	58.12	58.30	46.27	45.30	53.60	52.50

Table 7: QuIeTT with limited table representations (n=2).

Appendix figures & tables14 assets

Supplementary material from the paper’s appendix.

Appendix

Operator	Description
add_row_id	Adds a unique, stable row identifier column to the table.
rename	Renames columns according to a provided name mapping.
select	Selects and retains a specified subset of columns.
parse_date_text	Parses free-text date strings into a normalized date representation.
parse_number	Extracts numeric values (and optional units) from text fields.
extract_regex	Extracts structured groups from a column using a regular expression.

Appendix

Table 8: Atomic table transformation operators used in the transformation plan.

Model	Context window
DeepSeek-V3.1	Up to 128K tokens
Qwen-3-80B	Up to 128K tokens
GPT-OSS-120B	Up to 131K tokens
Gemini-2.0	Up to $\sim$ 1,048,576 tokens
Gemini-2.5	Up to $\sim$ 1,048,576 tokens

Appendix

Table 9: Documented context window limits for models used during QUIETT preprocessing. Actual rate limits depend on provider- and project-specific quota configurations.

Hyperparameter	Value
Decoding	Greedy / temperature $=0.2$
Max output tokens (Step 1: Issue)	8000
Max output tokens (Step 2: Plan)	6000
Max output tokens (Step 3: Code)	1024
Max output tokens (Step 4: Q&A)	4096
Stop sequences:	none (generation is post-processed/validated)

Appendix

Table 10: Inference hyperparameters used in the QUIETT pipeline. Step 1-3 are query-independent preprocessing; Step 4 performs downstream QA on the canonical table.

Criterion	Description
Answer Correctness	The provided answer exactly matches the value(s) supported by the table.
Answerability	The question can be answered using only the table content.
Clarity	The question is unambiguous and clearly specifies the required reasoning.
Structural Reasoning	The question requires aggregation, comparison, filtering, or reasoning over derived or normalized fields.

Appendix

Table 11: Annotation rubric used for human verification of the QUIETT challenge set.

		Gemini 2.0
Category	Method	WikiTQ		NQ-Table		SequentialQA		HiTab
Category	Method	F1	EM	F1	EM	F1	EM	F1	EM
Direct	CoT	61.76	60.67	68.70	67.50	62.33	59.87	76.11	72.00
	CoT + SQL	44.43	44.02	43.29	41.45	50.06	46.83	52.10	50.00
	NormTab	55.95	52.42	51.80	49.76	47.95	42.14	56.24	53.60
	E5	66.60	65.90	48.40	46.34	42.75	38.18	75.83	71.40

Appendix

Table 12: Gemini 2.0 results (F1 / EM). QuIeTT compared against direct prompting methods (CoT, CoT+SQL, NormTab, E5, EEDP, Binder, TableSQLify, TableReasoner) and agentic methods (Chain-of-Table, POT, GraphOTTER) across four benchmarks. Bold indicates best result per column.

Dataset	Gemini 2.5		GPT-OSS
Dataset	F1	EM	F1	EM
WikiTQ	68.34	68.14	66.72	63.57
NQ-Table	70.11	69.34	68.20	65.56

Appendix

Table 13: QuIeTT performance when a smaller model (LLaMA-14B Scout) is used for issue generation.

Dataset	Mean Issues	Median	% in 11-20
WikiTQ	16.1	15	98.5%
NQ-Table	16.1	15	99.4%

Appendix

Table 14: Issue statistics across datasets demonstrating that most tables naturally fall within the

11

20

issue range.

Transform model: Gemini 2.5
QA model $\rightarrow$	LLaMA		Mistral
	F1	EM	F1	EM
WikiTQ	59.87	57.50	65.12	62.23
NQ-Table	59.01	53.78	63.28	58.12
SequentialQA	46.72	43.12	47.24	44.84
HiTab	62.30	61.20	59.61	58.40

Appendix

Table 15: F1 / EM performance across transformation model and QA model combinations. Stronger transformations consistently improve downstream QA regardless of QA model size.

Figure 5: Distribution of detected structural issues per table across WikiTQ and NQ-Table. Nearly all tables exhibit between $11$ - $20$ issues, supporting the probe budget used by QUIETT.

Figure 6: Original hierarchical table with multi-level headers for agri-food employment distribution across two Ontario regions (2011).

Figure 7: Flattened table obtained by collapsing hierarchical column headers into composite columns combining region and worker type.

Figure 8: Relational table produced by QUIETT after decomposing composite headers into atomic attributes such as region and worker type. Each row represents a single observation suitable for SQL generation.

Figure 9: Original WebTable containing Ohio State Buckeyes football season data. The table is already flat but contains implicit structure such as game location, ranking, and match outcomes embedded in textual fields.

Figure 10: QUIETT-transformed relational representation of the original WebTable (Figure 9 ). Composite fields such as result and site are decomposed into atomic attributes including Is_Win , Team_Score , Opponent_Score , Stadium , City , and State . The resulting schema enables direct and unambiguous SQL querying.

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

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
EKLAV: Elicited Key CoT Learning via Answer only superVision
Jul 28, 2026Adarsh Singh, Aman Singh, Kushal Raj Bhandari +3Table QA CoT Distillation
Arizona State University · Rensselaer Polytechnic Institute · Scale AI
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

Jun 20, 2026J/K move · Enter open · S save

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

Guixin Su, Qiankun Pi, Mayi Xu, Wenli Li+4

School of Computer Science, Wuhan University, Wuhan, China

Table QA

Open paper AskPDF arXiv

Table Question Answering (TableQA) aims to reason over tables to answer user queries. Existing research treats all questions uniformly and evaluates solely through overall accuracy, obscuring a critical reality that LLMs excel at simple lookups yet struggle with complex operations like aggregation and arithmetic. To reveal this disparity, we introduce a novel Operation-wise TableQA task with a fine-grained question taxonomy and release two datasets named WikiTQ-ow and TabFact-ow for evaluation. As for modeling bottlenecks, existing methods flatten tables into linearized texts, disrupting inherent structures and inducing the ``lost-in-the-middle'' issue, which poses a primary barrier to complex cross-row reasoning. Moreover, they typically reason from scratch, neglecting reusable patterns shared across similar operations. To address these limitations, we propose a Skill-augmented Table Graph Reasoning (SkillTGR) framework for self-evolving structured reasoning. Specifically, SkillTGR represents tables as attributed graphs with explicit row-column-cell structures, where LLMs plan and execute dynamic chains to retrieve evidence subgraphs for graph traversal reasoning. Based on this, SkillTGR builds a hierarchical SkillBank to distill reason trajectories into abstract skills under cognitive heuristics, then hybrid retrieves both successful and failed skills for contrastive augmented table graph reasoning, thereby enabling the continual self-evolution. Extensive experiments demonstrate that SkillTGR achieves superior performance with an average of 5.91% overall and 5.06% operation-wise improvement, also reducing 19.76% token consumption and 27.64% inference latency. Our codes and data will be released upon publication.

Abstract

Table QA

Explore similar work

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

EKLAV: Elicited Key CoT Learning via Answer only superVision

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