Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Human vs. Agent in Task-Oriented Conversations

Zhefan Wang, Ning Geng, Zhiqiang Guo, Weizhi Ma, Min Zhang

TL;DR

This work addresses the challenge of data scarcity in task-oriented conversational systems by systematically comparing human users to LLM-based agent users under controlled, parallel conditions. It introduces a three-aspect, ten-dimension analytical framework and analyzes four representative scenarios using both automated GPT-4o-based evaluations and human annotations. Key findings reveal distinct differences in problem-solving style, question specificity, engagement patterns, and language style, while showing consistency in certain dimensions like breadth versus depth and perceived usefulness. The study provides actionable guidance to refine LLM-based user simulations, enabling more realistic offline evaluation and privacy-preserving data generation for interactive AI systems. Overall, the work lays a foundation for more human-aligned simulations and informs how to better leverage synthetic users in developing and evaluating task-oriented dialogue agents.

Abstract

Task-oriented conversational systems are essential for efficiently addressing diverse user needs, yet their development requires substantial amounts of high-quality conversational data that is challenging and costly to obtain. While large language models (LLMs) have demonstrated potential in generating synthetic conversations, the extent to which these agent-generated interactions can effectively substitute real human conversations remains unclear. This work presents the first systematic comparison between LLM-simulated users and human users in personalized task-oriented conversations. We propose a comprehensive analytical framework encompassing three key aspects (conversation strategy, interaction style, and conversation evaluation) and ten distinct dimensions for evaluating user behaviors, and collect parallel conversational datasets from both human users and LLM agent users across four representative scenarios under identical conditions. Our analysis reveals significant behavioral differences between the two user types in problem-solving approaches, question broadness, user engagement, context dependency, feedback polarity and promise, language style, and hallucination awareness. We found consistency in the agent users and human users across the depth-first or breadth-first dimensions, as well as the usefulness dimensions. These findings provide critical insights for advancing LLM-based user simulation. Our multi-dimensional taxonomy constructed a generalizable framework for analyzing user behavior patterns, offering insights from LLM agent users and human users. By this work, we provide perspectives on rethinking how to use user simulation in conversational systems in the future.

Human vs. Agent in Task-Oriented Conversations

TL;DR

This work addresses the challenge of data scarcity in task-oriented conversational systems by systematically comparing human users to LLM-based agent users under controlled, parallel conditions. It introduces a three-aspect, ten-dimension analytical framework and analyzes four representative scenarios using both automated GPT-4o-based evaluations and human annotations. Key findings reveal distinct differences in problem-solving style, question specificity, engagement patterns, and language style, while showing consistency in certain dimensions like breadth versus depth and perceived usefulness. The study provides actionable guidance to refine LLM-based user simulations, enabling more realistic offline evaluation and privacy-preserving data generation for interactive AI systems. Overall, the work lays a foundation for more human-aligned simulations and informs how to better leverage synthetic users in developing and evaluating task-oriented dialogue agents.

Abstract

Task-oriented conversational systems are essential for efficiently addressing diverse user needs, yet their development requires substantial amounts of high-quality conversational data that is challenging and costly to obtain. While large language models (LLMs) have demonstrated potential in generating synthetic conversations, the extent to which these agent-generated interactions can effectively substitute real human conversations remains unclear. This work presents the first systematic comparison between LLM-simulated users and human users in personalized task-oriented conversations. We propose a comprehensive analytical framework encompassing three key aspects (conversation strategy, interaction style, and conversation evaluation) and ten distinct dimensions for evaluating user behaviors, and collect parallel conversational datasets from both human users and LLM agent users across four representative scenarios under identical conditions. Our analysis reveals significant behavioral differences between the two user types in problem-solving approaches, question broadness, user engagement, context dependency, feedback polarity and promise, language style, and hallucination awareness. We found consistency in the agent users and human users across the depth-first or breadth-first dimensions, as well as the usefulness dimensions. These findings provide critical insights for advancing LLM-based user simulation. Our multi-dimensional taxonomy constructed a generalizable framework for analyzing user behavior patterns, offering insights from LLM agent users and human users. By this work, we provide perspectives on rethinking how to use user simulation in conversational systems in the future.

Paper Structure

This paper contains 32 sections, 3 figures, 11 tables.

Table of Contents

  1. Introduction
  2. Related Work
  3. User Simulation
  4. Task-Oriented Conversational System
  5. Previous Observations in Human vs. LLMs
  6. Conversation Protocol and Study Settings
  7. Study Overview
  8. Conversation Scenarios
  9. User Profiling
  10. Assistant LLMs
  11. User Simulation with LLM Agents
  12. Human Users
  13. Empirical Study Settings
  14. Findings: Conversation Strategy
  15. All-in-one or Step-by-step
  16. ...and 17 more sections

Figures (3)

  • Figure 1: The overview of our study, including three main steps: user profiles and task-oriented system, conversation with human users or agent users, and multi-aspect multidimensional comparison.
  • Figure 2: The distribution of all-in-one or step-by-step dimension on our constructed conversations. The dashed line indicates the 50% threshold.
  • Figure 3: Distribution of hallucination detected on our constructed conversations. We report the rate both on the conversation level and the utterance level.