Does Your AI Agent Get You? A Personalizable Framework for Approximating Human Models from Argumentation-based Dialogue Traces

TL;DR

This work addresses the need for adaptive, personalized human modeling in argumentation-based Explainable AI. It introduces Persona, a framework that combines Bayesian belief updates over a probabilistic human model with a prospect-theory–inspired probability weighting function, enabling dynamic, user-specific reasoning during dialogue. Through a large human-subject study (n=184) and comprehensive experiments, Persona outperforms state-of-the-art baselines in both approximating human beliefs and estimating argument probabilities, while maintaining real-time runtimes. The results demonstrate that learning and personalizing how humans perceive probabilistic arguments can significantly improve interactive explanations and decision support in AI systems, with clear implications for trust and collaboration in real-world settings.

Abstract

Explainable AI is increasingly employing argumentation methods to facilitate interactive explanations between AI agents and human users. While existing approaches typically rely on predetermined human user models, there remains a critical gap in dynamically learning and updating these models during interactions. In this paper, we present a framework that enables AI agents to adapt their understanding of human users through argumentation-based dialogues. Our approach, called Persona, draws on prospect theory and integrates a probability weighting function with a Bayesian belief update mechanism that refines a probability distribution over possible human models based on exchanged arguments. Through empirical evaluations with human users in an applied argumentation setting, we demonstrate that Persona effectively captures evolving human beliefs, facilitates personalized interactions, and outperforms state-of-the-art methods.

Figures (5)

• Figure 1: Example dialogue trace $\mathcal{T}$ between two agents.
• Figure 2: Probability weighting function with value pairs $(s, r)$ given as $\{(0.5, 1), (0.5, 2),(0.5, 3), (0.3, 3), (0.7, 3)\}$. Lower values of $s$ imply that the average probability reflects a lower level of human confidence in the agent's argument, whereas higher values of $r$ suggest excessive distortion, either through overweighting or underweighting of the probability. Note that the relationship between confidence and probability is linear when $s = 0.5$ and $r = 1$.
• Figure 3: Comparisons of Spearman's rank correlation distributions in round four derived from the first $k$ rounds of interaction data $\mathcal{D}_k$.
• Figure 4: Comparisons of Spearman's rank correlation distributions in model estimation within $[0.75, 1]$ in round $k$$(k = 2,3,4,5)$ of human model rankings where parameters are learned from the previous $k-1$ rounds. Note that for participants with only four interactions, the results for Round 5 are identical to those of Round 4.
• Figure 5: Comparisons of Spearman's rank correlation distributions for argument beliefs across different methods.