AI World Models Exhibit Sudden Collapse at Critical Points.

This paper investigates a phenomenon termed "world-model collapse" in long-horizon language agents, likening it to a physical phase transition where a small change in conditions leads to a dramatic shift in state, much like water boiling. The researchers observed that for certain parameter settings, minor adjustments to factors such as state load or the agent's planning horizon can cause an abrupt and significant deterioration in the agent's implicit understanding of its environment. Through extensive grid searches across various task parameters—including state cardinality, dependency density, horizon, branching, observation mode, and mutation rate—the study identified a distinct phase diagram. This diagram illustrates a "solved plateau" where agents perform well, a narrow "transition band," and a "collapse floor" where performance drastically drops. Detailed per-step traces revealed the underlying mechanism: the agent's internal fidelity to the world state fails *before* its ability to select valid actions. This indicates that the agent isn't merely making poor choices; rather, it's operating from a fundamentally corrupted or inaccurate internal representation of the world. The study also found that while stronger models can shift the critical boundary, they do not eliminate the qualitative nature of this sudden transition. These findings highlight world-model collapse as a measurable and significant bottleneck for developing reliable long-horizon AI agents.