Suggestion-Based Assistance of Suboptimal Users in Sequential Decision-Making Tasks

Niklas Dieckow, Pierre-Alexandre Murena

Keywords: Human-AI collaboration, zero-shot assistance, two-agents collaboration

Abstract:

AI agents, with their ability to plan and solve complex tasks, can assist humans by providing actionable suggestions. In such human-AI interactions, users have the choice to either accept the suggested action or reject it and decide independently, creating an asymmetric collaboration dynamic: the human user has full control over the actions taken, while the AI assistant lacks direct agency within the environment.
In this context, we show that even if the assistant has complete knowledge of the task and its optimal solutions, merely suggesting the task-optimal actions can lead to worse outcomes than leaving the user unassisted. To be effective, the assistant must also understand the user's decision-making process.
To address this challenge, we propose a zero-shot assistance method based on Bayesian estimation of a model of the user's acceptance and fallback behavior. Our approach does not require prior data and relies solely on a parametric model of typical user behavior. We validate our approach theoretically and empirically on a novel toy environment. By comparing to several baselines, we assess the stability of our solution in situations where the estimates are incorrect or the user’s policy is not expressible by the chosen model.

Area: Human-Agent Interaction (HAI)

Generative A I: I acknowledge that I have read and will follow this policy.

Submission Number: 991