LLMs Boost the Performance of Decision Trees on Tabular Data across Sample Sizes
Keywords: Tabular data, large language models, decision trees, ensembling
TL;DR: We fuse LLMs and gradient-boosted decision trees for a single learner that performs well on tabular datasets of all sizes.
Abstract: Large language models (LLMs) perform remarkably well on tabular datasets in zero- and few-shot settings, since they can extract meaning from natural language column headers that describe features and labels. In contrast to LLMs, gradient-boosted decision trees (GBDTs) must learn the relationships among columns from scratch, increasing their data requirements. Meanwhile, LLMs are not competitive with GBDTs on medium or large datasets, and their scalability is capped by their limited context lengths. In this paper, we propose LLM-Boost, a simple and lightweight approach for fusing large language models with gradient-boosted decision trees, which enables larger datasets to benefit from the natural language capabilities of LLMs than was previously shown. While matching LLMs at sufficiently small dataset sizes and GBDTs at sufficiently large sizes, LLM-Boost outperforms both standalone models on a wide range of dataset sizes in between. We demonstrate state-of-the-art performance against numerous baselines and ensembling approaches, and we also show how to fuse GBDTs with TabPFN, a recent non-LLM model for in-context learning on tabular data. We find that this combination achieves the best performance on larger datasets. We release our code at https://anonymous.4open.science/r/LLM-Boost-21DD.
Supplementary Material: zip
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.
Reciprocal Reviewing: I understand the reciprocal reviewing requirement as described on https://iclr.cc/Conferences/2025/CallForPapers. If none of the authors are registered as a reviewer, it may result in a desk rejection at the discretion of the program chairs. To request an exception, please complete this form at https://forms.gle/Huojr6VjkFxiQsUp6.
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: 504
Loading