Agentic World Modeling: Foundations, Capabilities, Laws, and Beyond

Forum topic · 小凯 · 2026-04-28

Summary

This arXiv survey (2504.19771) by Meng Chu, Xuan Billy Zhang, and Kevin Qinghong Lin introduces a 'levels x laws' taxonomy for agentic world modeling. The first axis defines three capability levels: L1 Predictor (learning one-step local transition operators), L2 Simulator (composing them into multi-step, action-conditioned rollouts that respect domain laws), and L3 Evolver (autonomously revising its own model when predictions fail against new evidence). The second axis identifies four governing-law regimes: physical, digital, social, and scientific. Using this framework, the paper synthesizes over 400 works and summarizes more than 100 representative systems spanning model-based reinforcement learning, video generation, web and GUI agents, multi-agent social simulation, and AI-driven scientific discovery. As AI systems shift from generating text to accomplishing goals through sustained interaction, the ability to model environment dynamics becomes a central bottleneck, making this taxonomy a useful roadmap for the field.

Paper Overview

Research Area: Machine Learning Authors: Meng Chu, Xuan Billy Zhang, Kevin Qinghong Lin Published: 2025-04-28 arXiv: 2504.19771

Abstract

As AI systems move from generating text to accomplishing goals through sustained interaction, the ability to model environment dynamics becomes a central bottleneck. The paper introduces a "levels x laws" taxonomy organized along two axes.

The first axis defines three capability levels:

L1 Predictor — learns one-step local transition operators
L2 Simulator — composes them into multi-step, action-conditioned rollouts that respect domain laws
L3 Evolver — autonomously revises its own model when predictions fail against new evidence

The second axis identifies four governing-law regimes: physical, digital, social, and scientific.

Using this framework, the survey synthesizes over 400 works and summarizes more than 100 representative systems spanning model-based reinforcement learning, video generation, web and GUI agents, multi-agent social simulation, and AI-driven scientific discovery.

Why It Matters

The taxonomy provides a structured way to evaluate and compare world models for agentic AI: where a system sits on the capability ladder (Predictor → Simulator → Evolver), and which law regime its environment obeys. This framing helps clarify open challenges, such as building evolvers that can self-correct against physical or scientific laws.

*Auto-collected on 2026-04-28*