TM TabularMath
ACL 2026 Findings · Accepted · Structured Reasoning

TabularMath: Understanding Math Reasoning over Tables with Large Language Models

A neuro-symbolic pipeline and benchmark for probing mathematical reasoning over large, imperfect, and multimodal tables.

Shi-Yu Tian1,2,* Zhi Zhou1,* Wei Dong2,* Kun-Yang Yu1,2 Ming Yang1,2 Zi-Jian Cheng1,3 Lan-Zhe Guo1,3 Yu-Feng Li1,2
1 National Key Laboratory for Novel Software Technology, Nanjing University
2 School of Artificial Intelligence, Nanjing University
3 School of Intelligence Science and Technology, Nanjing University
* Equal contribution

Why TabularMath Matters

Existing math reasoning benchmarks still under-cover realistic table use: larger structured contexts, flawed tables, and the split between text and visual table inputs.

Motivation

Mathematical reasoning is widely used to evaluate large language models, but table-based reasoning remains comparatively underexplored. In practice, systems must retrieve evidence from larger tables, remain robust when tables are incomplete or inconsistent, and generalize across both text-rendered and visually rendered inputs. TabularMath is introduced to make these pressures explicit rather than hiding them behind average-case scores.

AutoT2T: Automated Text-to-Table Generation

AutoT2T converts math word problems into validated tabular reasoning tasks without manual table annotation.

Overview of the AutoT2T pipeline

AutoT2T first formalizes the problem, then transforms it into a seed table, and finally applies controllable augmentations for difficulty and robustness testing.

Stage 1

Semantic Decoupling

Large language models extract variables, constraints, and goals from natural-language math problems, while formal solvers verify satisfiability and consistency.

Stage 2

Table Transformation

The verified formal state is rewritten into a structured seed table with explicit fields, entity rows, and tabularized assignments.

Stage 3

Table Augmentation

Row augmentation, column augmentation, order shuffling, and information modification create controllable splits for complexity, retrieval stress, and robustness.

A Structured Testbed for Table Reasoning

The benchmark is built from transformed GSM8K-style problems and covers both text-based and rendered image-based tables.

Dataset comparison for TabularMath

Compared with prior math and table reasoning datasets, TabularMath emphasizes larger tables, coupled retrieval-and-reasoning, and robustness to flawed inputs.

Release

The paper names the benchmark TabularMath. The released Hugging Face dataset is available as TabularGSM, covering easy, medium, hard, and imperfect settings for controlled evaluation.

Easy

Low variation with simpler retrieval.

Medium

Shuffled tables with harder localization.

Hard

Larger tables with row and column expansion.

Imperfect

Missing or contradictory tables for robustness.

Three Core Experimental Messages

We analyze TabularMath from three main angles: table size and structural complexity, table quality, and modality-sensitive representation.

Performance gap between average and hard problems
Angle 1

Table Size and Structural Complexity Drive Difficulty

Performance on top-complexity questions drops much more sharply than on average samples, showing that larger and more structurally demanding tables remain a major bottleneck.

Trade-off between robustness checking and normal performance
Angle 2

Table Quality Matters, Especially for Imperfect Inputs

Informing models that a table may be problematic improves discrimination, but also hurts accuracy on ordinary well-defined problems, revealing a genuine robustness trade-off.

Different text table representations
Angle 3

Modality and Representation Change the Difficulty Profile

Text-rendered tables remain easier than visual ones in general, and within text inputs, structured formats such as JSON or serialization stay more reliable than Markdown.

Retrieval versus full reasoning comparison
Discussion

Retrieval Is Not the Same as End-to-End Reasoning

When the retrieval target is made explicit, models perform much better. The harder part is deciding what to retrieve while reasoning through the full problem.

Code reasoning comparison
Discussion

Code Helps Less Than Clear Retrieval Guidance

Writing code does not automatically solve table reasoning; strong gains appear when the model is explicitly guided toward the right retrieval trajectory.

BibTeX

The citation below follows the current arXiv preprint metadata for arXiv:2505.19563.

DOI
Ready to copy.
@article{tian2025tabularmath,
	  title   = {TabularMath: Understanding Math Reasoning over Tables with Large Language Models},
	  author  = {Tian, Shi-Yu and Zhou, Zhi and Dong, Wei and Yu, Kun-Yang and Yang, Ming and Cheng, Zi-Jian and Guo, Lan-Zhe and Li, Yu-Feng},
  journal = {arXiv preprint arXiv:2505.19563},
  year    = {2025},
  doi     = {10.48550/arXiv.2505.19563},
  url     = {https://arxiv.org/abs/2505.19563}
}