CFMS: A Coarse-to-Fine Multimodal Synthesis Framework for Enhanced Tabular Reasoning

Researchers introduce CFMS, a two-stage framework combining multimodal large language models with symbolic reasoning to improve tabular data comprehension for question answering and fact verification tasks. The approach achieves competitive results on WikiTQ and TabFact benchmarks while demonstrating particular robustness with large tables and smaller model architectures.

The CFMS framework addresses a fundamental limitation in current tabular reasoning systems: the gap between visual pattern recognition and symbolic logic. While Chain-of-Thought methods have advanced reasoning capabilities, they operate within purely symbolic constraints that miss structural and visual patterns inherent in tables. This research bridges that divide through a hierarchical two-stage approach that separates concerns between perception and computation.

The methodology leverages multimodal large language models to extract high-level insights from tables in the coarse stage, generating a knowledge tuple that captures multi-perspective understanding. This tuple then guides a symbolic engine in the fine stage, creating an efficient reasoning map rather than forcing the model to reason blindly over raw data. This decoupling represents a meaningful architectural innovation in how AI systems can process semi-structured information.

The competitive results on established benchmarks validate the framework's effectiveness, but the noted robustness with large tables and smaller backbone models carries particular significance for practical deployment. Many production systems operate under computational constraints, and demonstrating that this approach scales efficiently with reduced model size increases its real-world applicability. This addresses a persistent challenge in AI development: achieving strong performance without proportional increases in computational requirements.

Future work should focus on extending this framework to other semi-structured domains beyond tabular data, such as knowledge graphs or multi-document reasoning scenarios. The principle of hierarchical decoupling between perception and symbolic reasoning could generalize across multiple data modalities, potentially reshaping how reasoning systems are architecturally designed.

• →CFMS combines multimodal perception with symbolic reasoning in a two-stage framework for improved tabular question answering and fact verification.
• →The coarse stage uses MLLMs to synthesize multi-perspective knowledge tuples that guide fine-stage symbolic reasoning execution.
• →Benchmark results show competitive accuracy on WikiTQ and TabFact with particular strength handling large tables.
• →Framework demonstrates effectiveness with smaller backbone models, reducing computational requirements for deployment.
• →The hierarchical decoupling approach between perception and reasoning represents a generalizable architectural principle for semi-structured data.