Large Language Models (LLMs) have shown strong capabilities in table reasoning, but their effectiveness degrades as tables grow in size and complexity due to irrelevant context and difficulty localizing the evidence required for reasoning. Existing approaches typically reason over either the full table or a single reduced view, which can still obscure important row-column relationships. We introducePARTAB (Partition-Aware Reasoning overTables), a framework that constructs a structured evidence interface between the LLM and the table. PARTAB represents query-relevant evidence as semantically coherent, row-linked table regions and performs hierarchical selection over column groups and row-level partitions before composing the selected evidence for answer generation. We evaluate PARTAB on multiple table reasoning benchmarks, covering question answering, fact verification, and numerical reasoning. PARTAB consistently improves over full-table prompting and several recent table reasoning methods, achieving strong performance on WikiTableQuestions and TabFact while remaining competitive on numerical reasoning. Additional analyses show that semantic partitioning and targeted evidence selection improve evidence localization, substantially reduce the reasoning context, and provide larger benefits on complex tables. These results demonstrate the value of structured, partition aware evidence construction for scalable table reasoning.
Figures & tables
Figure 1: Comparison of reasoning behaviors under different table reasoning paradigms. (I) Full-table prompting exposes the model to the entire table, increasing irrelevant context and making evidence localization more difficult in long and wide tables. (II) Single-view pruning reduces context size but may remove required evidence, leading to incomplete reasoning and incorrect aggregation. All data shown are synthetic and for illustration only.
Figure 2: Overview of PARTAB. Given a question and table, PARTAB constructs semantically coherent table partitions and performs localized understanding over selected partitions enabling reliable reasoning over large and noisy tables. All data shown are synthetic and for illustration only .
Method
WikiTableQuestions
TabFact
TableBench
EM
Acc
NR (EM)
FC (EM)
BINDER ( Cheng et al., 2023 )
38.03
67.50
-
-
DATER ( Ye et al., 2023 )
52.00
74.70
-
-
ReAcTable ( Zhang et al., 2024b )
57.09
-
-
-
End-to-End Prompting
59.43
73.22
59.50
66.66
Chain-of-Thought (CoT)
64.31
75.99
-
-
Table 1: Downstream performance on table reasoning benchmarks. PARTAB achieves strong results across question answering, fact verification, and numerical reasoning, outperforming several recent baselines on WikiTableQuestions and TabFact while remaining competitive on TableBench.
Figure 3: Partition construction performance. Manual evaluation of semantic column-grouping quality over 100 tables.The generated groups are fully coherent in 96% of the evaluated cases.
Selection Strategy
WikiTQ
TabFact
TableBench
EM
Acc
NR (EM)
FC (EM)
Basic Selection
74.71
87.88
64.43
79.26
Similarity-based Selection
67.25
85.87
58.69
59.25
LLM-based Selection
79.31
90.48
70.33
82.71
Table 2: Comparison of part-selection strategies across table reasoning benchmarks. The best result in each column is shown in bold.
Figure 4: Distribution of the number of semantic column groups generated per table. Most tables are decomposed into three or four groups across all evaluated datasets.
LLM Backbone
WikiTQ
TabFact
TableBench
EM
Acc
NR (EM)
FC (EM)
GPT-4o-mini
79.31
90.48
70.33
82.71
Gemini-2.5-Flash-Lite
71.23
83.56
61.29
64.53
DeepSeek-v4-Flash
78.30
91.70
70.23
75.34
Table 3: Performance of PARTAB across different answer-execution backbones.
LLM Backbone
Method
WikiTQ-Hard
TabFact-Hard
TB-NR-Hard
EM
Δ
EM
Δ
EM
Δ
GPT-4o-mini
Full Table
45.80
–
55.32
–
40.47
–
PARTAB (Ours)
55.72
+9.92
80.85
+25.53
61.90
+21.43
GPT-5-mini
Full Table
60.30
–
57.14
–
57.45
–
PARTAB (Ours)
61.83
+1.53
78.26
+21.12
61.90
+4.45
DeepSeek-v4-Pro
Full Table
58.01
–
55.32
–
54.76
–
Table 4: Performance of full-table prompting and PARTAB on hard subsets containing large or wide tables. WikiTQ-Hard and TB-NR-Hard are evaluated using Exact Match (EM), while TabFact-Hard is evaluated using accuracy. Each Δ denotes the absolute improvement of PARTAB over full-table prompting with the same backbone.
Dataset
Partition Construction
Partition Selection
Groups / Table
Columns / Group
Parts / Table
Selected Parts
Reduction (%)
WikiTQ
3.13
2.10
17.56
3.88
77.9
TabFact
3.17
2.11
9.84
2.44
75.2
TableBench
3.32
2.31
12.27
3.04
75.2
Table 5: Statistics of partition construction and selection. Reduction is computed as 1−candidate partsselected parts .
Configuration
WikiTQ
EM
Δ
Full PARTAB
79.31
–
PARTAB w/o Question Analyzer
72.21
−7.10
PARTAB with Random Column Grouping
68.84
−10.47
PARTAB with Row Chunking Only
57.19
−22.12
Table 6: Component-level ablation results on WikiTQ. Each Δ denotes the absolute change in exact match relative to the full PARTAB configuration.
Figure 5: Distribution of failure types based on 100 incorrect predictions. Part-selection errors constitute the largest source of failure.
Appendix figures & tables13 assets
Supplementary material from the paper’s appendix.
Appendix
Method
LLM Calls
WikiTQ EM
Full-table (Budget-matched)
5
64.61
PARTAB
5
79.31
Appendix
Table 7: Budget-matched comparison between repeated full-table prompting and PARTAB .
Dataset
Fallback Rate (%)
Evidence Recall (%)
WikiTQ
2.70
87.0 †
TabFact
1.90
–
TableBench
0.24
–
Appendix
Table 8: Selection stability across the evaluation datasets. † Evidence recall is based on manual annotation of 100 WikiTQ examples.
Method
LLM Calls
Avg. Tokens / Query
TabFact Acc.
Full-table prompting
1
2.1k
73.22
Chain-of-Table
≤25
13.2k
85.09
TableMaster
11
3.3k
90.12
PARTAB – Basic
4
2.1k
87.88
PARTAB – Similarity
4
1.6k
85.87
PARTAB – LLM-based
5
2.3k
90.48
Appendix
Table 9: Accuracy–efficiency comparison on TabFact. Token counts cover the complete per-query pipeline. Bold values indicate the lowest token usage and the highest accuracy, respectively.
Method
Avg. Selected Parts ↓
WikiTQ (EM)
Full Table
17.56
59.43
Basic Selection
6.50
74.71
Similarity-based
5.36
67.25
LLM-based
3.88
79.31
Appendix
Table 10: Partition reduction efficiency on WikiTableQuestions. We report the average number of selected parts and Exact Match (EM) accuracy.
Dataset
Number of Rows
Number of Columns
Min
Avg
Max
Min
Avg
Max
WikiTableQuestions
5
25.65
517
3
6.39
21
TabFact
5
13.41
47
5
6.29
20
TableBench
6
16.75
176
4
6.64
20
Appendix
Table 11: Structural statistics of the evaluation datasets.
Figure 6: Illustrative example of PARTAB on a multi-step numerical reasoning task. The table is truncated for clarity; ellipses ( … ) indicate omitted columns and rows. PARTAB isolates the relevant metric columns, selects a small subset of partitions, and correctly computes the difference between the highest and lowest elevations.
Figure 7: Illustrative example of PARTAB on a comparison-based reasoning task. The table is truncated for clarity; ellipses ( … ) indicate omitted columns and rows. PARTAB selects the relevant semantic groups and partitions, enabling accurate identification of the most biodiverse country.
Figure 8: Illustrative example of PARTAB on a lookup-based reasoning task. The table is truncated for clarity; ellipses ( … ) indicate omitted rows and columns. PARTAB identifies the relevant row and extracts the country information from the athlete name field.
Figure 9: Question Analyzer Prompt Template used to infer reasoning requirements from a natural language question.
Figure 10: Column Grouping Prompt Template used to construct semantically coherent column groups for partition-aware table reasoning.
Figure 11: Group Selection Prompt Template used to identify the minimum required semantic groups for answering a question.
Figure 12: LLM-based Part Selection Prompt Template used to select the minimum required table parts from the candidate partition set.
Figure 13: Answer Execution Prompt Template used to generate the final answer from the selected table parts.
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%.
Tung Sum Thomas Kwok, Zeyong Zhang, Xinyu Wang +6
University of California, Los Angeles · SimpleWay · New Jersey Institute of Technology +4
Large Language Models (LLMs) have shown strong performance on table question answering, yet their accuracy often degrades as table size increases. We find that this degradation is not uniform across question types. Localization-sensitive questions are particularly affected by irrelevant table content, while questions requiring broader evidence may still benefit from full-table reasoning. Based on this observation, we propose a question-adaptive framework that dynamically selects between localized and full-table reasoning. The framework constructs question-specific sub-tables through operation-aware table decomposition and uses the predicted question type to determine the appropriate reasoning mode. We further introduce silver reference sub-tables for evaluating evidence selection and construct SLQA, a benchmark based on real-world long tables. Experiments on WikiTQ and SLQA show that localization is particularly effective for lookup and local reasoning questions, while adaptive selection between localized and full-table reasoning achieves the best overall performance. These results highlight that long-table QA requires deciding not only how to localize, but also when to localize. Our code and datasets will be made available upon publication of the paper.
Table reasoning remains challenging for large language models (LLMs), particularly in tasks that require multi-step inference over long and structured tables. Existing approaches predominantly rely on single-direction reasoning, which limits their ability to explore alternative hypotheses across tasks. In this work, we propose CRAFT, a unified Counterfactual Reasoning Framework that reformulates Tabular question answering and fact verification into a general bidirectional verification process. Our method explicitly constructs both declarative statements and their counterfactual variants. Evidence is then extracted from reasoning along both the original and counterfactual paths, and integrated via a weighted mechanism to arrive at the final answer. Experimental results show that our approach consistently surpasses representative baselines on table reasoning datasets such as WikiTQ and TabFact, achieving especially large improvements on complex question answering. Our framework also significantly mitigates performance gaps between different backbone LLMs. This indicates that counterfactual reasoning effectively overcomes the limitations of single-direction inference, guiding LLMs toward more discerning reasoning and establishing a more principled paradigm for structured reasoning tasks. Our code will be made publicly available upon acceptance.