Fudan University and Meituan's Longcat team have open-sourced a new interactive world model benchmark called WBench. It's designed to poke and prod video generation models on their understanding of physics, spatiotemporal consistency, and interactive control — the underlying skills that separate a convincing video from a glitchy mess.
https://twitter.com/Meituan_LongCat/status/2059658634829996047
The benchmark includes 289 test cases and 1,058 rounds of interaction, covering both first-person and third-person perspectives. It combines navigation control, subject actions, event editing, and viewpoint switching. WBench standardizes the input interface across text instructions, 6-DoF poses, and discrete actions, making it possible to compare different control paradigms head-to-head. The evaluation pipeline uses 22 automatic metrics, and the scoring results achieve a Spearman rank correlation of at least 0.94 with human blind-test win rates.
The tests reveal some uncomfortable truths. While video quality is becoming increasingly saturated, interactive control is almost decoupled from rendering, consistency, and physics. Camera motion control doesn't guarantee subject consistency: models like HY-World 1.5 and Matrix-Game 3.0, which excel at navigation, still struggle with identity loss and view drift in third-person perspectives. Meanwhile, physical correctness correlates strongly with rendering quality, but shows near-zero correlation with control ability. Open-source world models lead in several dimensions: HY-World 1.5 tops the navigation control charts, LingBot-World dominates consistency, and Matrix-Game 3.0 ranks first in action-based navigation.
Multi-round interaction tells a more sobering story. Every model's performance degrades as interaction rounds increase, with navigation control dropping fastest due to accumulated spatial bias. Explicit geometric control can help mitigate drift — HY-World 1.5's multi-round stability far exceeds that of text-driven Kling 3.0. Scene dynamics and subject rigidity form the underlying difficulty structure of the benchmark. First-person, static scenes with rigid robot subjects are easy to pass, while intense motion and non-rigid subjects like animals remain long-term challenges due to deformation and speed complexity.