Introduction
Large Language Models (LLMs) have demonstrated impressive capabilities in reasoning over structured data. However, most existing research treats tables as static artifacts, focusing primarily on question answering (QA). In real-world scenarios, users often need to interact with tables dynamically—interpreting values, modifying existing entries, or generating new rows based on textual descriptions. To bridge this gap, we introduce iTBLS, a framework and dataset designed for interactive conversations over tabular information. By reformulating complex table operations as a series of conditional QA tasks, we achieve significant improvements in both structural integrity and content accuracy.
The Interaction Gap in Table Processing
While LLMs excel at reading tables, they frequently struggle with "writing" or "editing" them. Two primary challenges persist:
- The Interaction Gap: There is a lack of datasets capturing natural language commands for dynamic table manipulation, such as swapping rows or correcting specific errors.
- The Structural Integrity Problem: Direct generation of table syntax (e.g., LaTeX or Markdown) often leads to hallucinations where the model fails to maintain consistent column counts or row structures.
Methodology: Tables as Question-Answering
The core innovation of the iTBLS framework is the shift from direct table generation to a multi-step Question-Answering reformulation. Instead of asking an LLM to output a modified table directly, the process is decomposed into three stages:
- Interaction Identification: The system classifies the user's intent as interpretation (QA), modification (editing existing data), or generation (adding new data).
- QA Reformulation: For modifications, the model identifies the specific cells referenced and uses a programmatic backend (e.g., a pandas dataframe) to execute changes.
- Conditional Generation: When adding data, the model generates headers and then performs extractive QA for each cell, using the user's natural language input as the source text.
This approach ensures that the LLM acts as an "extractor" rather than an "architect," leaving the structural management to a robust code-based backend.
The iTBLS Dataset
The iTBLS dataset consists of 4,000 tables extracted from scientific pre-prints on arXiv, covering fields like Mathematics, Physics, and Computer Science. It includes over 12,000 interaction turns. Unlike datasets sourced from Wikipedia, iTBLS provides a rigorous testbed involving complex numerical data, scientific units, and professional terminology. Models must handle both explicit commands (e.g., "Swap row 1 and 2") and implicit semantic corrections (e.g., "Fix the error where the error rates are flipped").
Results and Performance Gains
The "iTBLS as QA" approach significantly outperforms traditional sequence-to-sequence baselines. On the iTBLS "Generate" task, Exact-Match accuracy improved from 48.94% to 73.32%. Furthermore, on standard text-to-table benchmarks like E2E and WikiTableText, the method achieved a 13% improvement in Exact-Match accuracy and a 16% improvement in BERTScore compared to previous state-of-the-art models.
Conclusion and Future Work
The iTBLS framework demonstrates that decomposing table operations into extractive QA tasks effectively mitigates structural hallucinations. Future research will focus on expanding the scope of modifications to include mathematical transformations and multi-turn "conversational wrangling," where users perform sequences of complex edits.
Summary updated