Can World Models Enable Agents to Foresee the Future? A Counterintuitive Answer from a New Paradigm Study

Background

World models receive the current environment state, simulate the next state under physical laws, and output a prediction, while agents observe the current state and select actions to achieve a goal. From this perspective the two form a naturally complementary closed loop, providing the theoretical basis for using world models to empower agent decision‑making.

Tool‑making Paradigm

The authors treat the world model as a third‑party foresight tool. In the proposed paradigm an agent can, at each step, decide whether to invoke the world model to simulate the consequences of a candidate action before executing it. Figure 1 illustrates this loop, where the agent optionally calls the world model for foresight in a dense‑room escape scenario.

Tasks and Evaluation Modes

The study evaluates two task families:

Agentic Task : agents operate in simulated environments (e.g., box pushing, object picking, navigation) requiring multi‑step reasoning.

Visual Question‑Answering (VQA) Task : agents answer spatial reasoning questions from images, using world‑model rollouts (WAN2.1) to obtain 3‑D foresight.

Three experimental modes are defined:

World Model Invisible Mode : the agent is unaware of the world model and never calls it.

Normal Mode : the agent knows the world model exists and may call it voluntarily (the main setting).

World Model Forcing Mode : the system forces the agent to call the world model at every step.

Key Findings

Finding 1: Adding perfect foresight does not reliably improve performance; in many cases it degrades results because agents treat the foresight signal as noise.

Finding 2: Most models rarely invoke the world model, showing a low call rate (often <0.1 calls per episode), especially large models such as GPT‑5 which never call it.

Finding 3: Call‑rate varies across model families; smaller models tend to call more often (cognitive offloading), but higher call frequency does not guarantee better performance.

Foresight Governance Framework

To explain successful versus failed integration, the authors propose a three‑stage governance pipeline:

Foresight Formulation (What to ask) : the agent decides when and what to request from the world model.

Simulation Generation (What to simulate) : the world model produces accurate, high‑quality simulations.

Interpretation & Integration (How to use) : the agent interprets the simulation results and incorporates them into the next action.

Implications

The study concludes that the dominant bottleneck is the stability of foresight governance rather than the raw fidelity of the world model. Future research should focus on developing mechanisms for agents to (1) assess when foresight is worthwhile, (2) formulate precise simulation requests, and (3) reliably integrate simulation evidence into multi‑step decision loops.