World-Model Collapse as a Phase Transition

2026-06-30

AI Summary

This study investigates whether long-horizon language agents exhibit abrupt, phase-transition-like collapse behaviors in implicit world models due to minor parameter perturbations. By constructing a family of deterministic tasks with precisely defined ground-truth states at each step, the authors systematically analyze how state capacity, dependency density, and horizon length influence model stability. They reveal, for the first time, a sharp phase transition boundary in language agent world models: performance shifts discontinuously from stable solving to sudden collapse, driven by distortions in world state representation rather than action errors. Through large-scale grid searches and trajectory analyses, they map a phase diagram featuring a solvable plateau, a narrow transition band, and a collapsed floor. Notably, stronger models merely shift the critical boundary without eliminating the fundamental phase transition, establishing world model collapse as a key bottleneck for long-horizon language agents.

Abstract

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.

Problem

• world-model collapse
• phase transition
• long-horizon agents
• implicit world models
• critical boundary

Innovation

• world-model collapse
• phase transition
• long-horizon agents
• implicit world models
• critical boundary