SkillTGR Enhances Table Question Answering with Graph Reasoning

Traditional Table Question Answering (TableQA) systems often treat all questions uniformly, masking the fact that Large Language Models (LLMs) excel at simple lookups but struggle with complex operations like aggregation or arithmetic. This research introduces a new "Operation-wise TableQA" task and two datasets (WikiTQ-ow, TabFact-ow) to highlight and address this disparity. A key challenge for existing methods is that they linearize tables, losing their inherent structural information and leading to "lost-in-the-middle" issues for cross-row reasoning. To overcome these limitations, the paper proposes Skill-augmented Table Graph Reasoning (SkillTGR). This framework represents tables as attributed graphs, allowing LLMs to plan and execute dynamic chains to retrieve evidence subgraphs for graph traversal reasoning. SkillTGR also builds a hierarchical SkillBank, distilling successful and failed reasoning trajectories into abstract skills. By contrastively augmenting reasoning with these skills, the system enables continual self-evolution. Experiments show SkillTGR achieves superior performance, with an average of 5.91% overall and 6.03% operation-wise improvement, while also reducing token consumption by nearly 20% and inference latency by over 27%.