# IntentGPT: Few-shot Intent Discovery with Large Language Models

**Authors:** Juan A. Rodriguez, Nicholas Botzer, David Vazquez, Christopher Pal, Marco Pedersoli, Issam Laradji  
**arXiv:** [2411.10670](https://arxiv.org/abs/2411.10670)  
**Date:** 2024-11-16  
**Summary page:** https://www.summarizepaper.com/en/paper/2411.10670v1/

## Key points

- Dialogue systems are crucial for user interactions in various domains, requiring efficient identification of user intents.
- Intent Detection models aim to categorize user intents accurately but struggle with the diverse and dynamic nature of intents.
- Intent Discovery is emerging as a solution to discover new intents as they emerge, reducing reliance on predefined intents.
- IntentGPT is a novel training-free method leveraging Large Language Models like GPT-4 for intent discovery with minimal labeled data.
- Experimental results show that IntentGPT outperforms methods relying heavily on domain-specific data in benchmarks like CLINC and BANKING.
- Red-teaming efforts are essential to identify and mitigate biases within LLMs like GPT-4 and Llama 2-70B used in the study.
- Responsible AI research should prioritize addressing bias concerns and ensuring fair content generation practices when using AI assistants.

## Summary

, , , , 

In the rapidly evolving digital landscape, dialogue systems have become integral in facilitating user interactions across various domains, from customer service to virtual assistants. A key challenge in these dialogues is automatically identifying users' goals to efficiently address their needs. This has led to the adoption of Intent Detection models, which aim to categorize user intents accurately. However, the diverse and dynamic nature of users' intents makes it difficult to rely on a fixed set of predefined intents. As a solution, there is a growing emphasis on developing models capable of discovering new intents as they emerge, giving rise to the field of Intent Discovery. Intent Discovery has garnered significant attention in recent research endeavors as existing methods typically require extensive training data to effectively identify new intents, necessitating substantial human effort. To tackle this issue, a novel approach called IntentGPT has been introduced. IntentGPT is a training-free method designed to leverage Large Language Models (LLMs) like GPT-4 for intent discovery with minimal labeled data. The framework consists of an In-Context Prompt Generator for generating informative prompts for In-Context Learning, an Intent Predictor for classifying and uncovering user intents from utterances, and a Semantic Few-Shot Sampler that selects relevant few-shot examples and known intents to enhance prompt injection. Experimental results showcase that IntentGPT surpasses previous methods that heavily rely on domain-specific data and fine-tuning in prominent benchmarks such as CLINC and BANKING. By enabling LLMs like GPT-4 to autonomously discover new intents with limited supervision, IntentGPT offers a promising solution for enhancing dialogue systems' adaptability and performance in real-world applications. Furthermore, it's essential to acknowledge potential biases present in user utterances that may influence the model's inferred intents. Red-teaming efforts are crucial for identifying and mitigating biases within LLMs like GPT-4 and Llama 2-70B used in this study. Responsible AI research must prioritize addressing bias concerns and ensuring fair content generation practices when utilizing AI assistants for tasks such as code development enhancement or manuscript writing assistance. In conclusion, the innovative approach presented by IntentGPT signifies a significant advancement in intent discovery within dialogue systems by leveraging state-of-the-art LLMs effectively while also highlighting the importance of ethical considerations and bias mitigation strategies in AI development.