Adaptive multi-format synergistic table understanding framework for large language models

Abstract

Tables serve as a foundational medium for representing structured information and are widely adopted across diverse domains such as scientific research, financial analysis, public administration, and data-intensive industrial applications. Existing large language models (LLMs)-based table understanding approaches commonly adopt a single serialized representation, which easily introduces format bias and structural information loss, accompanied by critical bottlenecks including semantic sparsity, inaccurate evidence grounding, and unreliable numerical computation. To tackle these limitations, we propose an Adaptive Multi-Format Synergistic Table Understanding framework named AFS-TU. Firstly, the framework establishes a heterogeneous representation space covering semi-structured and natural-language-based table descriptions to construct comprehensive multi-view representations for tabular data.Secondly, a meta-strategy network dynamically selects the optimal format subset tailored to the given task to generate complementary multi-view observations and enrich table structural semantics. Subsequently, a staged semantic enhancement module converts sparse tabular inputs into dense structured contexts to improve the accuracy of evidence localization. Finally, we incorporate evidence-aware confidence calibration and selective symbolic correction by generating and executing deterministic code to rectify numerical errors in low-confidence samples. Empirical evaluations on three benchmarks including WikiTableQuestions, TabFact and the challenging TableBench dataset demonstrate that AFS-TU delivers consistent and scalable performance improvements across LLMs of different architectural scales. The framework effectively alleviates the core bottlenecks in table understanding, and achieves synergistic improvements in both accurate evidence localization and precise arithmetic computation.