# Rethinking Interpretability in the Era of Large Language Models

**Authors:** Chandan Singh, Jeevana Priya Inala, Michel Galley, Rich Caruana, Jianfeng Gao  
**arXiv:** [2402.01761](https://arxiv.org/abs/2402.01761)  
**Date:** 2024-01-30  
**Summary page:** https://www.summarizepaper.com/en/paper/2402.01761v1/

## Key points

- Interpretable machine learning is a significant area of interest driven by large datasets and deep neural networks.
- Large language models (LLMs) have impressive capabilities and offer new possibilities for interpretable machine learning.
- LLMs can provide explanations in natural language, expanding the scope of patterns communicated to humans.
- Challenges include hallucinated explanations and high computational costs.
- Research priorities for LLM interpretation include analyzing new datasets directly and generating interactive explanations.
- The authors discuss the development of inherently interpretable models and post-hoc interpretability techniques.
- LLMs can offer explanations for expert human behavior and enable user-centric interactive explanations.
- Integrating LLMs into interpretative processes has transformative potential in redefining boundaries in machine learning interpretability.

## Summary

Interpretable machine learning has become a significant area of interest in recent years, driven by the proliferation of large datasets and deep neural networks. Concurrently, large language models (LLMs) have showcased impressive capabilities across various tasks, offering new possibilities for interpretable machine learning. The ability of LLMs to provide explanations in natural language expands the scope and complexity of patterns that can be communicated to humans. However, these advancements also bring challenges like hallucinated explanations and high computational costs. In this position paper, the authors delve into evaluating methods for interpreting LLMs and utilizing them for explanation. Despite their limitations, LLMs present an opportunity to redefine interpretability with a broader scope across different applications, including auditing LLMs themselves. The paper highlights two emerging research priorities for LLM interpretation: leveraging LLMs to analyze new datasets directly and generating interactive explanations. The authors emphasize the rapid growth of interpretable ML fueled by the availability of vast datasets and powerful neural network models. They discuss the development of inherently interpretable models alongside post-hoc interpretability techniques. Additionally, they explore how LLMs can offer explanations for expert human behavior and enable more user-centric interactive explanations. In conclusion, the paper underscores the transformative potential of integrating LLMs into interpretative processes to redefine boundaries in machine learning interpretability. The authors advocate for harnessing the full capabilities of LLMs to enhance explanation reliability and advance dataset interpretation for knowledge discovery. This shift towards incorporating LLMs represents a pivotal moment in shaping the future landscape of interpretable ML.