New Taxonomy for Agentic World Models in AI

A recent publication on arXiv presents a classification system for agentic world modeling, introducing a framework based on 'levels x laws.' This classification outlines three levels of capability: L1 Predictor (local transitions in one step), L2 Simulator (rollouts conditioned on multiple actions), and L3 Evolver (self-directed model updates). Additionally, it categorizes four governing-law domains: physical, digital, social, and scientific. The authors contend that as AI evolves from generating text to engaging in goal-driven interactions, understanding environmental dynamics becomes essential. This framework seeks to harmonize the varied interpretations of 'world model' found within different research fields.