ZhiYuan’s GE 2.0 Wins WorldArena World Model Championship – How It Achieved Bare‑Bones Victory

WorldArena is the most authoritative leaderboard for world‑model research, featuring a rigorous evaluation suite that combines 16 detailed core metrics with three real‑application tasks to assess perception accuracy, physical‑law understanding, 3‑D spatial cognition, and action prediction.

In the latest CVPR 2026 competition, ZhiYuan entered its self‑developed model Genie Envisioner‑Sim 2.0 (GE 2.0) without any special task‑specific design. The team performed only a basic fine‑tune on the leaderboard data, yet the “light‑weight” entry still secured the overall champion, confirming GE 2.0’s strong general‑purpose adaptability.

GE 2.0’s functional matrix is the first to fully cover long‑sequence generation, multi‑view generation, ontology‑state generation, near‑real‑time inference and reward discrimination, establishing a complete technical capability loop for a world simulator.

On long‑sequence inference, GE 2.0 exhibits remarkable stability: when generating continuous video clips of 40–50 seconds, the visual quality degrades far less than the industry baseline and remains superior to the baseline’s performance within its first 10 seconds.

The model’s reliability is validated through extensive closed‑loop evaluations. Results show strong correlation with real‑world outcomes across multiple tasks. The team conducted case‑by‑case rollout comparisons and presented quantitative evidence via a confusion matrix, underscoring GE 2.0’s trustworthiness as a strategy evaluator.

With a reward model‑driven feedback loop, GE 2.0 automatically filters high‑quality rollout data and feeds it back to the policy model. Experiments demonstrate that this mechanism yields significant performance gains for the policy model on several tasks.

Overall, GE 2.0 advances beyond the common industry issue of “visual‑heavy, physics‑light” models. It delivers stable, physically accurate, and deployable simulation for embodied AI, enabling robots to accumulate experience in virtual environments and transfer it efficiently to real‑world operations.