AI Model Simulates Thucydides Trap Using Fuzzy Cognitive Maps
A new arXiv preprint (2605.17903) introduces a method for automatically generating feedback causal fuzzy cognitive maps (FCMs) from text. The technique uses large-language-model agents to break text into overlapping chunks, then convex mixing of these chunk FCMs produces a representative cyclic FCM knowledge graph. The mixing structure enables Bayesian inference to create 'de-chunked' posterior-like FCMs. The approach is demonstrated on Graham Allison's essay on the 'Thucydides Trap'—the conflict model between a dominant power (United States) and a rising power (China). The resulting FCM dynamical systems predict outcomes of such geopolitical tensions. The method scales efficiently using sparse causal chunk matrices.
Key facts
- arXiv:2605.17903v1
- Uses large-language-model agents to generate fuzzy cognitive maps from text
- Text is broken into overlapping chunks
- Convex mixing of chunk FCMs creates a cyclic FCM knowledge graph
- Bayesian inference produces de-chunked posterior-like FCMs
- Demonstrated on Allison's Thucydides Trap model
- Thucydides Trap describes conflict between dominant and rising powers
- Dominant power: United States; rising power: China
Entities
Institutions
- arXiv
Locations
- United States
- China