⌁ Awesome World Model Hub Papers · Datasets · Projects
中文 GitHub
← Back to papers

World-Model Collapse as a Phase Transition

arXiv 2026 34.7 theory

TLDR

Long-horizon language agents exhibit a sudden world-model collapse at a critical boundary, analogous to a phase transition.

Reasoning

The paper systematically investigates world-model collapse in deterministic tasks with exact state, revealing a phase diagram and mechanism where world-state fidelity fails before action validity. Strengths include clear experimental design and identification of a measurable bottleneck; weaknesses are the synthetic setting (no real-world data) and limited generality beyond deterministic tasks.

Read-first score

Read-first score 34.7, weighted from topical fit, citation, graph, method, reproducibility, and recency signals. Original total remains 35.

Recency 6%
100

Uses a gentle age decay so recent papers surface without erasing older foundations. 2026

Topical relevance 29%
50

Uses existing LLM keyword relevance scores normalized to 0-100. world model,world simulator,generative world model,interactive world model,video world model,world dynamics prediction,model-based reinforcement learning world model

Methodology quality 18%
50

Screens visible abstract and analysis fields for experiment, dataset, baseline, metric, and limitation evidence. markers=result

Reproducibility 18%
30

Screens links and visible text for paper, code, dataset, artifact, and repository signals. pdf=True; code=False; dataset=False; markers=none

Citation impact 18%
0

Uses OpenAlex-shaped citation metadata as a bibliometric attention signal, separate from paper quality. cited_by_count=0

Citation velocity 12%
0

Citation velocity estimates citations per publication-year to reduce old-paper bias. velocity=0.00

Field roles

Frontier

Rank sensitivity

Stability: volatile; rank range: 154.

Keyword Scores

world model
10
world dynamics prediction
8
model-based reinforcement learning world model
6
world simulator
5
generative world model
3
interactive world model
2
video world model
1

Deep Analysis

Innovations

  • Conceptualizing world-model collapse as a phase transition in long-horizon language agents.
  • Demonstrating that small parameter changes near a critical boundary cause sudden collapse of the agent's world model.
  • Mapping a phase diagram with solved plateau, narrow transition band, and collapse floor via grid search over multiple factors.
  • Revealing that world-state fidelity fails before action validity, indicating the agent acts from a corrupted world.
  • Showing that stronger models shift the critical boundary but do not eliminate the qualitative transition.

Methodology

The authors study a deterministic task family with exact per-step gold state, performing a large grid search over state cardinality, dependency density, horizon, branching, observation mode, and mutation rate. They analyze per-step traces to examine the mechanism of collapse.

Key Results

A phase diagram emerges with a solved plateau, a narrow transition band, and a collapse floor. World-state fidelity degrades before action validity, so agents act from a corrupted world model. Stronger models shift the critical boundary but do not remove the transition.

Resources

Paper

Tags