Tongji’s “Boundless” World Model Wins Open‑Source #1 and Overall #2 in WorldArena

WorldArena: The Hardcore Testbed for World Models

WorldArena provides a unified, public, multi‑dimensional evaluation suite for world models, focusing on six core aspects—visual quality, motion quality, content consistency, physical compliance, 3D spatial accuracy, and controllability—further broken down into 16 metrics that assess a model’s ability to predict future states, respond to action sequences, and understand physical interactions.

Why “Boundless” Stands Out

In Track‑1 (video quality), the open‑source model BLM scored 64.54, ranking first among open‑source entries, while the closed‑source model BWM‑Fast scored 67.87, placing second overall, just 0.39 points behind the leader. The model delivers high visual fidelity, consistent lighting, stable object contacts, and accurate long‑term spatial structure, directly addressing common drift and physical distortion issues in earlier world models.

Six Embodied Tasks Covered

Spatial Re‑arrangement: Preserves object identity and layout, maintains stable stacking contacts, and predicts adaptive gripper control.

Hinge Interaction: Captures opening/closing dynamics under hinge constraints, keeping geometry coherent during rotation.

Fine‑grained Manipulation: Dynamically captures affordances, aligns contact points with functional regions, and maintains consistent interaction outcomes.

Dual‑arm Coordination: Models coordinated arm motions, ensures object consistency, and avoids collisions during close‑handed hand‑offs.

Long‑range Placement: Sustains scene consistency over long sequences, prevents occlusion‑induced drift, and achieves stable placement in confined spaces.

Out‑of‑Distribution Generalization: Handles unseen initial scenes and object appearances while preserving physical consistency throughout the action sequence.

Efficiency‑Focused Technical Strategy

Rather than scaling data and compute indiscriminately, the Tongji team enhanced data value through augmentation and introduced a DiT‑based architecture with first‑frame guidance, dynamic memory, and dual‑path action modulation. These innovations improve training efficiency while delivering world‑class performance, proving that superior world models need not rely solely on massive resources.

Open‑Source and Closed‑Source Dual Track

The open‑source BLM provides a reproducible foundation that lowers entry barriers for research and development, whereas the closed‑source BWM‑Fast pushes performance limits, validating the technical direction under intense competition. This dual approach fosters both community standards and industrial relevance.

Implications for Embodied Intelligence

World models serve as the predictive core for robots to understand environments, plan actions, and reduce real‑world trial costs. The “Boundless” results indicate systematic advances in perception accuracy, action response, physical consistency, and long‑term reasoning, moving the field toward practical deployment.

Open‑Source Resources

https://github.com/boundless-large-model/boundless-world-model
https://huggingface.co/BLM-Lab/Boundless-World-Model

The project continues to evolve, aiming to build a more open, reproducible, and real‑world‑oriented world‑model ecosystem.