Search papers, labs, and topics across Lattice.
This study investigates the phenomenon of world-model collapse in long-horizon language agents, drawing an analogy to the phase transition observed in boiling water. By conducting a comprehensive grid search across various parameters, the authors identify a critical boundary where minor changes can lead to a sudden degradation in the agent's world model, resulting in a collapse of its behavior. The findings reveal that this collapse occurs due to a failure in world-state fidelity before action validity, establishing a measurable bottleneck for the performance of long-horizon agents.
Small changes in parameters can trigger a dramatic collapse in language agents' world models, revealing a critical phase transition that impacts their long-term decision-making.
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.
