In human-AI collaboration, it is essential for AI systems to understand and anticipate human behavior in order to coordinate effectively. Conversely, humans also form inferences about the agent’s beliefs and goals to facilitate smoother collaboration. As a result, AI agents should adapt their behavior to align with human reasoning patterns, making their actions more interpretable and predictable. This principle forms the foundation of transparent planning (MacNally et al, 2018).

A key prerequisite for transparent planning is a robust model of how humans interpret and infer goals from observed behaviors (i.e., a human model of goal recognition). Prior work has explored human goal recognition in controlled environments (Zhang et al, 2024) via manipulating factors such as timing and solvability in puzzle-like tasks. However, that work only focuses on passive settings. As an extension of goal recognition tasks, active goal recognition (AGR) offers a more realistic framework where observers are not passive but can interact with the environment to gather information actively. Therefore, there is a need to extend these models to account for human behavior in active goal recognition scenarios. Such an extension is critical for building AI systems that can both interpret and be interpreted by humans in interactive, real-world environments.