# Are LLMs All You Need for Task-Oriented Dialogue?

**Authors:** Vojtěch Hudeček, Ondřej Dušek  
**arXiv:** [2304.06556](https://arxiv.org/abs/2304.06556)  
**Date:** 2023-04-13  
**Summary page:** https://www.summarizepaper.com/en/paper/2304.06556v1/

## Key points

- Research explores effectiveness of Instructions-tuned Large Language Models (LLMs) in task-oriented dialogue scenarios
- LLMs are popular for engaging in conversations with users
- LLMs underperform compared to specialized models in explicit belief state tracking
- LLMs can guide dialogues towards successful outcomes with accurate slot values
- Access to true belief state distribution or domain-specific examples improves dialogue completion for LLMs
- Research provides insights into strengths and limitations of LLMs in task-oriented dialogue systems
- Emphasizes the need for specialized models for explicit belief state tracking

## Summary

This research by Vojtěch Hudeček and Ondřej Dušek explores the effectiveness of Instructions-tuned Large Language Models (LLMs) in completing multi-turn tasks and interacting with external databases in task-oriented dialogue scenarios. LLMs have gained significant popularity due to their ability to engage in conversations with users. The authors evaluate the performance of LLMs in explicit belief state tracking, comparing them to specialized task-specific models. The results indicate that LLMs underperform in this aspect. However, they demonstrate the capability to guide dialogues towards successful outcomes when provided with accurate slot values. Additionally, the study reveals that the ability of LLMs to achieve successful dialogue completion improves when they have access to either the true belief state distribution or examples from within the specific domain. Overall, this research provides valuable insights into the strengths and limitations of LLMs in task-oriented dialogue systems. It highlights their potential for guiding conversations effectively but emphasizes the need for specialized models for explicit belief state tracking.