Suggestion-based assistance of suboptimal users in sequential decision-making tasks

AI agents can assist humans by offering suggestions which users may accept or reject. This creates an asymmetric collaboration where the user retains full control while the AI lacks direct agency. We demonstrate that merely suggesting task-optimal actions can yield worse outcomes than unassisted performance; effective assistance requires understanding the user’s decision-making. To address this, we propose a zero-shot method based on Bayesian estimation of the user’s acceptance and fallback behavior, relying on a parametric model rather than prior data. We validate our approach theoretically and empirically on a novel toy environment, assessing its performance against baselines and stability in situations where parameters are incorrectly estimated.

Subjects

Human-AI collaboration

two-agent collaboration

zero-shot assistance

DDC Class

006.3: Artificial Intelligence

006.33: Knowledge-based Systems

600: Technology

Options

Suggestion-based assistance of suboptimal users in sequential decision-making tasks