🤖 AI Summary
Large language models (LLMs) suffer from degraded table reasoning performance due to input-length constraints, hindering effective processing of long, wide tables.
Method: We propose Adaptive Table Filtering (ATF), a plug-and-play framework that requires no model fine-tuning. ATF dynamically identifies and retains query-relevant table regions via question-aware column semantic description generation, hierarchical clustering, and sparse–dense vector alignment scoring. Its modular design enables cross-task adaptive balancing between information preservation and structural simplification.
Contribution/Results: ATF prunes ~70% of table cells on average, significantly improving reasoning accuracy across diverse TableQA benchmarks. Only in rare cases requiring full-table structural understanding does performance marginally decline. Crucially, ATF is the first approach to jointly model LLM-driven semantic comprehension with interpretable, structured filtering—achieving strong efficiency, task-agnostic generalizability, and parameter-free transferability.
📝 Abstract
Large language models (LLMs) for table-based reasoning often struggle with large tables due to input length limits. We propose ATF (Adaptive Table Filtering Framework), a modular and question-aware filtering pipeline that prunes uninformative columns and rows using LLM-generated column descriptions, clustering, and sparse-dense alignment scores. ATF integrates seamlessly with existing models (e.g., TAPAS, TAPEX) without retraining. Experiments show that ATF reduces table cells by ~70%, boosting performance on out-of-domain TableQA tasks while causing slight performance drops on Table Fact Verification, where full-table context is more critical. These results highlight ATF's ability to adaptively balance informativeness and minimalism across tasks.