README.md

metadata

title: Agentic World Model Explorer
colorFrom: blue
colorTo: purple
sdk: gradio
sdk_version: 4.36.0
app_file: app.py
pinned: false

Agentic World Model Explorer

An interactive exploration of the "levels x laws" taxonomy from the Agentic World Modeling paper (2604.22748).

What This Does

Demonstrates the three capability levels of world models:

• L1 Predictor: One-step local transitions
• L2 Simulator: Multi-step action-conditioned rollouts
• L3 Evolver: Self-revising models that update from prediction failures

Across four law regimes:

• Physical (object manipulation, physics)
• Digital (web/GUI agents, software)
• Social (multi-agent coordination)
• Scientific (experimental design)

Hypothesis

World models with explicit structured state representations (L2+) demonstrate better compositional generalization than pure next-token predictors when evaluated on out-of-distribution scenarios within the same law regime.

Findings

See the live demo for interactive examples of state representation strategies and their impact on generalization.