PoTable: Programming Standardly on Table-based Reasoning Like a Human Analyst
Keywords: Table-based Reasoning, Large Language Model, Symbolic Tools, Real-time Program Execution, Human Cognitive Behavior
TL;DR: We propose PoTable, a novel table-based reasoning method to program standardly like a human analyst in a real-time Python interpreter with an LLM.
Abstract: Table-based reasoning has garnered substantial research interest, particularly in its integration with Large Language Model (LLM) which has revolutionized the general reasoning paradigm. Numerous LLM-based studies introduce symbolic tools (e.g., databases, Python) as assistants in complex information understanding and arithmetic computations. However, they emphasize extensive and flexible utilization of symbolic tools, without fully considering the intrinsic logic of the reasoning process. In this study, we propose PoTable as a simple yet effective table-based reasoning method. Specifically, PoTable features a planning phase and an executing phase, implemented with an LLM-based operation planner and code generator and a Python interpreter as the real-time executor. To incorporate logical top-level guidance, we split the entire reasoning process into several distinct analysis stages with macroscopic instruction injection. As the reasoning process is structured suitably under the top-level guidance with precise and specific goals, PoTable produces superior reasoning results with highly accurate, steply commented and completely executable code. To summarize, PoTable enjoys the advantages of accuracy and explainability that make it a distinguished tabular data analyst. Extensive experiments over three evaluation datasets from two public benchmarks on two backbones demonstrate the outstanding performance of PoTable. In particular, GPT-based PoTable achieves over 4% higher absolute accuracy than runner-ups on all evaluation datasets. Our code is available at https://anonymous.4open.science/r/PoTable-6788.
Primary Area: applications to computer vision, audio, language, and other modalities
Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.
Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide.
Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors’ identity.
No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.
Submission Number: 10382
Loading