Understanding the Role of Mental Models in User Interaction with an Adaptive Dialog Agent

Lindsey Vanderlyn, Dirk Väth, Ngoc Thang VuNAACL 2025 · DOI

Abstract: In human-computer interaction, a user's "mental model" represents their internal theory of how a system operates. When these models mismatch the system's actual behavior, interaction quality suffers. This paper investigates the role of mental models in task-oriented, information-seeking dialogs. We present an Adaptive Dialog Agent (ADA) that implicitly infers user intent—categorized as "orienteering" or "teleporting"—to provide tailored guidance. Evaluating ADA against static baselines using the new RDMM dataset, we find that users possess diverse and often contradictory expectations. However, implicit adaptation allows the agent to achieve a significantly higher task success rate (80.65%) and improved user perception while maintaining mental model stability.

Introduction

A mental model is a simplified representation used by individuals to explain and predict a system's behavior. In the context of conversational agents, if a user expects a system to handle natural language but the system only processes keywords, the interaction is likely to fail. While adaptive dialog agents—systems that change behavior to suit the user—offer a potential solution, they introduce the "moving target" problem. Unpredictable changes can shatter a user's mental model, leading to confusion and loss of trust. This study addresses whether an agent can adapt implicitly, inferring needs from the initial query, without causing "mental model whiplash" in complex domains like business travel reimbursement.

Methodology and the RDMM Dataset

The researchers conducted a between-subject study with 63 participants across three agent types:

  1. Adaptive Dialog Agent (ADA): An LLM-based system (GPT-4o-mini) that analyzes the first user input to determine a search strategy: "orienteering" (step-by-step guidance) or "teleporting" (direct answers).
  2. Handcrafted Dialog System (HDC): A rigid baseline providing maximum guidance.
  3. FAQ System: A baseline providing only single-turn, direct answers.

To support this research, the authors introduced the REIMBURSE Dialog Mental Models (RDMM) dataset. It contains 180 interaction logs augmented with pre- and post-interaction surveys on mental models, usability, and trust.

The Diversity of User Mental Models

A key finding of this research is the "No Consensus" reality. Users do not share a single, unified mental model for how a chatbot should function. As shown in the data, expectations vary wildly regarding input style (natural language vs. keywords) and output type (general vs. personalized answers).

Distribution of user expectations regarding chatbot interaction styles.

This diversity suggests that static, "one-size-fits-all" dialog designs are fundamentally flawed. When a user's model does not align with the system's capabilities—for instance, expecting personalized answers from a basic FAQ system—performance and satisfaction drop significantly.

Performance Evaluation

The Adaptive Dialog Agent (ADA) demonstrated superior performance across all primary metrics. By aligning with user expectations from the first turn, ADA achieved a task success rate of over 80%, compared to 57% for the FAQ and 46% for the HDC systems.

Comparison of performance metrics across Adaptive, FAQ, and Handcrafted Dialog systems.

ADA also optimized dialog efficiency, requiring approximately 10.6 turns per successful interaction. This represents a "Goldilocks" zone: longer than the often insufficient single-turn FAQ responses, but significantly shorter and less tedious than the 15 turns required by the overly-guided HDC system.

Discussion and Conclusion

The study concludes that implicit adaptation is a "free lunch" for dialog design. ADA improved task success and user perception (Trust, Reliability, and Usability) without destabilizing user mental models. While users of static systems often had to "painfully" revise their beliefs when the system failed, ADA users experienced no significant shocks to their internal theories of the system. Future work will explore expanding adaptation to linguistic style and politeness across different domains.

Summary updated

BibTeX
@inproceedings{vanderlyn2025understanding,
  author = {Lindsey Vanderlyn and Dirk Väth and Ngoc Thang Vu},
  title = {{Understanding the Role of Mental Models in User Interaction with an Adaptive Dialog Agent}},
  booktitle = {Conference of the North American Chapter of the Association for Computational Linguistics},
  year = {2025},
  doi = {10.18653/v1/2025.findings-naacl.56},
}
Figure