A preprint reports that PARTAB scored 79.31 in exact match on WikiTableQuestions and 90.48 in accuracy on TabFact. On TableBench, it scored 70.33 in exact match for numerical reasoning and 82.71 in accuracy for fact checking. The work was evaluated on three benchmarks covering table question answering and fact verification. These are reported benchmark results from the study.
PARTAB maps a question and table to an answer through four stages: Question Analyzer, Partition Builder, Group and Part Selector, and Answer Executor. The pipeline moves from analyzing the question and building partitions to selecting groups and parts and then executing the answer.
A smaller evidence set
In an evaluated WikiTQ subset, a budget-matched comparison reported 79.31 exact match for PARTAB versus 64.61 for full-table prompting, a 14.70-point difference. It compared the two approaches under the same reported budget condition.
On hard subsets, the reported average gains varied by answer-execution backbone: 18.96 points for GPT-4o-mini, 9.03 points for GPT-5-mini and 12.65 points for DeepSeek-v4-Pro. These are separate reported averages for the three listed backbones.
Relative to candidate partitions, the reported reduction was 77.9% on WikiTQ, 75.2% on TabFact and 75.2% on TableBench. In practical terms, the answer stage used a smaller selection of table material than the candidate pool in those comparisons.
Checks inside the pipeline
A separate diagnostic reported that 96% of generated groups were fully coherent. The result relates to the grouping and selection stages named in PARTAB's four-stage pipeline.
Selected partitions achieved 87% evidence recall in the reported diagnostic. The check manually inspected 100 WikiTQ examples. The percentage is a sample-based diagnostic, not a universal estimate across tables and tasks.
In component comparisons, removing the Question Analyzer was associated with a 7.10-point lower exact-match result, while randomly constructed column groups were associated with a 10.47-point lower result. The largest reported degradation occurred with row chunking without semantic column grouping, at 22.12 points. The figures describe reported differences between versions in these ablation comparisons.
Where the method has limits
One important boundary is global completeness. The method does not explicitly enforce it for aggregation tasks, so required evidence may be missing when full-table coverage is needed. The limitation concerns tasks where the full table must be covered.
The reported analysis is descriptive: it presents averages, percentage-point differences, percentages and manual tallies, but no confidence intervals or significance tests. The 87% evidence-recall estimate came from manual review of 100 WikiTQ examples, so it should be read as a limited diagnostic.
The paper also notes that LLM-based question analysis, grouping and selection are sensitive to prompt design and model variability, with errors potentially propagating across stages. The multi-stage pipeline may add latency compared with single-pass prompting.
The document identifies itself as arXiv:2608.24082v1, dated 25 Aug 2026. The authors state that code and data will be released upon acceptance of the paper.
Paper data and sources
Original title: PARTAB: Partition-Aware Reasoning with Structured Evidence for Scalable Table Understanding
Authors: Md Mahadi Hasan Nahid, Davood Rafiei
Journal/Repository: arXiv
Status: Preprint, not yet peer-reviewed
First online: 2026-08-25
DOI: Not available
Original paper · Full text