Technical Report / Preprint

World Intelligence Agent

Unifying World Modeling and Action Learning for Digital Agents

Darwin Agent Team

A unified approach to understanding how actions change the world and learning how digital agents should act.

70 pages · PDF · 12.9 MiB · Original report

First page of the World Intelligence Agent report, showing its title, abstract, and evaluation overview. — Preview of the report's first page

About this report

World Intelligence Agent (WIA) brings together world modeling and action learning for digital agents. The report introduces WIA-Corpus and WIA-Bench, presents the WIA-WM family of world models, and explores simulation-based training and inference lookahead. It then extends the world-model interface toward action generation with WIA-WAM.

World modeling

Predicting the observations induced by actions, connecting perception, planning, and policy learning.

Data and evaluation

WIA-Corpus and WIA-Bench support training and evaluation across seven digital domains.

Action learning

Simulation-based learning and the WIA-WAM extension connect world models to policy decision-making.