Daily papers

Tuesday, 10 November 2026 · 4 min read

WorldFoundry provides unified infrastructure for evaluating generative world models

WorldFoundry defines generative world models as systems with persistent state, controllable dynamics, and closed-loop interaction, and offers open-source infrastructure integrating 270 models and 74 benchmarks for reproducible evaluation across video, 3D, 4D, and embodied systems.

Based on the paper WorldFoundry: Towards Unified Infrastructure for World Intelligence by Juanxi Tian, Jiadong Hou, Baiang Li, Xin Lin and 12 others

Challenges in defining and evaluating generative world models

Recent generative models increasingly synthesize not just static observations but evolving environments that respond to interventions, requiring a notion of world intelligence that includes perception, state inference, prediction, and interaction.

However, existing models vary widely in representations, interfaces, and evaluation protocols, making it difficult to compose capabilities, reproduce results, or measure progress toward true world modeling.

Visual plausibility alone is insufficient to demonstrate a model's world intelligence because it may lack persistent state, physical dynamics, or causal consistency under interventions, which are essential for closed-loop world modeling.

Sign in to read the whole paper

You have read the opening. The rest is free with a Bezen account, and no card is needed.

In the rest of this paper

  • How the method works, step by step, with a flow chart
  • The figure from the paper, and how to read it
  • How it was tested, what was found, and where it falls short
WorldFoundry provides unified infrastructure for evaluating generative world models · Bezen