Harnessing LLMs Explanations to Boost Surrogate Models in Tabular Data Classification

TL;DR

The paper tackles the challenge of high resource usage and limited interpretability in LLM-based tabular data classification by proposing a three-stage in-context learning framework that uses LLM-generated explanations to guide demonstration selection and train a lightweight surrogate language model. By generating post hoc explanations for candidate demonstrations and leveraging them to filter and select informative prompts, the approach improves both predictive performance and interpretability while reducing reliance on costly LLM calls. Empirical results across four binary tabular datasets show a 5.31 percentage point average accuracy gain over strong baselines in few-shot settings, with substantial reductions in API usage when using a local SLM like Llama2-7B. The work demonstrates a practical, scalable path to harness LLM reasoning for tabular data through explanation-guided prompt design and demonstrates the value of preserving interpretability in model predictions for real-world deployment.

Abstract

Large Language Models (LLMs) have shown remarkable ability in solving complex tasks, making them a promising tool for enhancing tabular learning. However, existing LLM-based methods suffer from high resource requirements, suboptimal demonstration selection, and limited interpretability, which largely hinder their prediction performance and application in the real world. To overcome these problems, we propose a novel in-context learning framework for tabular prediction. The core idea is to leverage the explanations generated by LLMs to guide a smaller, locally deployable Surrogate Language Model (SLM) to make interpretable tabular predictions. Specifically, our framework mainly involves three stages: (i) Post Hoc Explanation Generation, where LLMs are utilized to generate explanations for question-answer pairs in candidate demonstrations, providing insights into the reasoning behind the answer. (ii) Post Hoc Explanation-Guided Demonstrations Selection, which utilizes explanations generated by LLMs to guide the process of demonstration selection from candidate demonstrations. (iii) Post Hoc Explanation-Guided Interpretable SLM Prediction, which utilizes the demonstrations obtained in step (ii) as in-context and merges corresponding explanations as rationales to improve the performance of SLM and guide the model to generate interpretable outputs. Experimental results highlight the framework's effectiveness, with an average accuracy improvement of 5.31% across various tabular datasets in diverse domains.

Figures (4)

• Figure 1: Previous approaches to context learning with LLMs, as depicted on the top side of the figure, directly utilize tabular data to prompt the LLMs for predictions. This process necessitates frequent and extensive access to large language models (LLMs), resulting in substantial computational and efficiency demands. In contrast, our method depict at the bottom, presents a more efficient approach, which uses LLM generated explanations to guide the few shot demonstrations selection and the interpretable outputs of SLM
• Figure 2: Overview of our method. We begin by randomly selecting a small number of samples from the tabular dataset to form candidate demonstration sets. These samples are then converted into text and fed into an LLM, which generates post hoc explanations. These explanations are used to guide the selection of relevant demonstrations and to facilitate the generation of interpretable predictions by the SLM.
• Figure 3: Comparison of different demonstration selection algorithms and baselines (random selection) on four datasets.
• Figure 4: Effect of filter threshold $p$ and the number of post hoc explanations $n$. The accuracy (acc) across all datasets is reported.