World Model as Tool: An Empirical Study on Agent Foresight Governance

Cheng Qian, Emre Can Acikgoz, Bingxuan Li, Xiusi Chen, Yuji Zhang, Bingxiang He, Qinyu Luo, Dilek Hakkani-Tür, Gokhan Tur, Yunzhu Li, Heng Ji

Keywords: Agent, World Model, Foresight Governance

Abstract: Vision-language agents are increasingly being applied to problems that require reasoning about how the world will unfold, rather than acting solely on short-term inference. Generative world models present a potential solution by serving as external simulators that let agents preview possible consequences before making decisions. In this paper, we study whether existing agents can actually use such world models as tools to improve their reasoning. Across a range of agentic tasks and visual question answering settings, we find that some agents almost never choose to simulate (under 1%), often mishandle the resulting rollouts (around 15%), and can show unstable or even worse performance (by as much as 5%) when simulation is provided or required. Additional attribution analysis suggests that the main limitation is not simulation quality itself, but the agents’ ability to judge when simulation is useful, interpret forecasted outcomes, and incorporate that foresight into later reasoning. These results point to the need for mechanisms that support more calibrated and strategic use of world models, ultimately enabling more dependable anticipatory cognition in future agent systems.

Email Sharing: We authorize the sharing of all author emails with Program Chairs.

Data Release: We authorize the release of our submission and author names to the public in the event of acceptance.

Submission Number: 129