# TabLLM: Few-shot Classification of Tabular Data with Large Language Models

**Authors:** Stefan Hegselmann, Alejandro Buendia, Hunter Lang, Monica Agrawal, Xiaoyi Jiang, David Sontag  
**arXiv:** [2210.10723](https://arxiv.org/abs/2210.10723)  
**Date:** 2022-10-19  
**Summary page:** https://www.summarizepaper.com/en/paper/2210.10723v1/

## Key points

- The paper explores the application of large language models in zero-shot and few-shot classification of tabular data.
- The authors propose a technique where they prompt a large language model with a serialization of the tabular data in natural-language format, along with a brief description of the classification problem.
- In the few-shot setting, they fine-tune the language model using labeled examples.
- The study evaluates various serialization methods, including templates, table-to-text models, and large language models.
- This approach outperforms previous deep-learning-based methods for tabular classification on multiple benchmark datasets.
- Even in zero-shot classification scenarios, the technique achieves significant performance by leveraging prior knowledge encoded in large language models.
- The approach is competitive with strong traditional baselines like gradient-boosted trees, especially in very few shot settings.
- The potential of large language models for tabular data classification tasks is highlighted.
- The authors provide additional context about their research paper's publication date, categories, and names of authors involved.

## Summary

The paper titled "TabLLM: Few-shot Classification of Tabular Data with Large Language Models" explores the application of large language models in zero-shot and few-shot classification of tabular data. The authors propose a technique where they prompt a large language model with a serialization of the tabular data in natural-language format, along with a brief description of the classification problem. In the few-shot setting, they fine-tune the language model using labeled examples. The study evaluates various serialization methods, including templates, table-to-text models, and large language models. Surprisingly, despite its simplicity, this approach outperforms previous deep-learning-based methods for tabular classification on multiple benchmark datasets. Even in zero-shot classification scenarios, the technique achieves significant performance, demonstrating its ability to leverage prior knowledge encoded in large language models. One notable aspect is that this approach proves to be competitive with strong traditional baselines like gradient-boosted trees, especially in very few shot settings. This finding highlights the potential of large language models for tabular data classification tasks. The authors provide additional context by mentioning that their work addresses an existing summary from an XML feed related to their research paper and provide details about the publication date and categories as well as names of authors involved in it. Overall, this study presents an innovative approach to leveraging large language models for tabular data classification tasks and demonstrates its effectiveness compared to traditional methods as well as other deep learning techniques.