World-Model Collapse as a Phase Transition
For researchers building long-horizon language agents, this work formalizes world-model collapse as a measurable bottleneck, showing it is a qualitative phenomenon that stronger models cannot eliminate.
The paper identifies a phase transition in long-horizon language agents' world models, where small parameter changes cause sudden collapse from accurate to corrupted world-state fidelity, independent of model strength. A grid search over task parameters reveals a consistent phase diagram with a narrow transition band.
Water looks unchanged as it warms, then at a critical point it boils. We ask whether long-horizon language agents show an analogous transition in their implicit world models. In some parameter settings, changing state load by a small amount, or adding a single step of horizon, leaves behavior nearly unchanged; near a critical boundary, the same small change causes a sudden world collapse. We study this effect in a deterministic task family with exact per-step gold state. A large grid search over state cardinality, dependency density, horizon, branching, observation mode, and mutation rate reveals a phase diagram: a solved plateau, a narrow transition band, and a collapse floor. Per-step traces show the mechanism: world-state fidelity fails before action validity, so the agent is not merely choosing a bad action; it is acting from a corrupted world. Stronger models translate the critical boundary but do not remove the qualitative transition. These results make world-model collapse a measurable bottleneck for long-horizon agents.