June 2, 2026

LongCat releases WBench, a benchmark for interactive video world models

LongCat has introduced WBench, a new benchmark designed for evaluating interactive video world models, distinguishing itself by focusing not just on visual quality but on a model’s ability to maintain a controllable world during multi-turn interactions. WBench assesses twenty models using 289 test cases across five dimensions: video quality, setting adherence, interaction adherence, consistency, and physics compliance. The benchmark reveals a notable gap between aesthetically pleasing video generation and effective world simulation, indicating that no model excels in all areas simultaneously. Particularly, it finds that navigation capabilities are almost entirely disconnected from visual quality and other performance metrics, challenging models to handle interactions that demonstrate control and coherent memory over multiple actions.

Astra: Astra is an open-source video model from Tsinghua University evaluated in the WBench benchmark. It completed the set of 20 models analyzed for world simulation capabilities.
WBench: WBench is a comprehensive multi-turn benchmark designed for evaluating interactive video world models across navigation, subject actions, event edits, and perspective switches. It was released by the Meituan LongCat Team to shift assessment from single-clip visual quality to sustained controllability, memory, and physical plausibility over multiple turns. The framework includes automated metrics validated against human judgments and supports text-, camera-, and action-conditioned models.
Genie 3: Genie 3 is a web-based video world model from Google evaluated in WBench. It served as an example of closed-source beta systems in the benchmark.
LTX 2.3: LTX 2.3 is an open-source video model from Lightricks that participated in WBench testing. It contributed to comparisons of camera- and text-conditioned approaches.
LongCat: LongCat refers to the Meituan LongCat Team, a research group developing AI models with a focus on video generation and world simulation. The team released WBench as a new benchmark to rigorously test interactive video world models beyond visual appeal. Their work emphasizes separating initial scene setup from subsequent user-driven interactions in evaluation.
Wan 2.7: Wan 2.7 is a video generation model available through Alibaba APIs. It was assessed in WBench for its performance in interaction adherence and physical plausibility.
Jiamu Li: Jiamu Li is a researcher at Fudan University and part of the Meituan LongCat Team who co-developed WBench. His role includes advancing the unified navigation protocol bridging different model control interfaces.
Siyu Ren: Siyu Ren is a researcher from Fudan University and the Meituan LongCat Team involved in the WBench project. She contributed to the creation of the benchmark dataset and metrics for testing long-horizon controllability in video world models.
YUME 1.5: YUME 1.5 is an open-source video model from Shanghai AI Lab included in WBench. It contributed to findings on the independence of navigation from other capabilities.
Kling 3.0: Kling 3.0 is a commercial video generation model offered via API by Kling AI. It was one of the top-performing systems evaluated in the WBench benchmark across multiple dimensions of world model capability.
Cosmos 2.5: Cosmos 2.5 is an open-source video model from NVIDIA evaluated in the WBench benchmark. It was assessed for its handling of multi-turn interactions and physics compliance.
Hengrui Hu: Hengrui Hu is a researcher at Fudan University and the Meituan LongCat Team who co-authored the WBench paper. He helped develop the evaluation framework that assesses models on setting adherence, interaction adherence, consistency, and physics compliance. His work supports systematic diagnosis of failures in current video generation systems.
Kairos 3.0: Kairos 3.0 is an open-source video model from SenseTime evaluated in WBench. It was included in comparisons across quality, interaction, and physics dimensions.
Xuezhi Cao: Xuezhi Cao is a researcher at Fudan University and the Meituan LongCat Team involved in WBench. He assisted in building the evaluation code and model submission pipeline for the benchmark.
Ziwen Wang: Ziwen Wang is a researcher from Fudan University and the Meituan LongCat Team who participated in the WBench release. His contributions include aspects of the interactive leaderboard and dataset resources.
MatrixGame2: MatrixGame2 is an open-source video model from Skywork that participated in WBench testing. It helped demonstrate performance variations among game-oriented models.
MatrixGame3: MatrixGame3 is an open-source model from Skywork that took part in the WBench evaluation. It represented action-conditioned systems in the testing.
HY-GameCraft: HY-GameCraft is an open-source model from Tencent assessed in the WBench benchmark. It contributed to the leaderboard of 20 evaluated systems.
HY-Video 1.5: HY-Video 1.5 is an open-source video model developed by Tencent. It featured in the WBench benchmark with noted results in consistency metrics.
HY-World 1.5: HY-World 1.5 is an open-source world model from Tencent. It was evaluated in WBench demonstrating strong navigation capabilities relative to other models.
Happy Oyster: Happy Oyster is a web-based video model from Alibaba evaluated in WBench. It represents web-product style systems tested for interactive world simulation.
Henghui Ding: Henghui Ding is a researcher at Fudan University and the Meituan LongCat Team serving as a corresponding author on the WBench paper. He oversaw the integration of automatic metrics with human judgment validation for the benchmark.
Kaining Ying: Kaining Ying is a researcher at Fudan University affiliated with the Meituan LongCat Team. He is the lead author of the WBench paper introducing the benchmark for video world model evaluation. His contributions center on the design of multi-turn interaction cases and diagnostic metrics.
Seedance 1.5: Seedance 1.5 is a video generation model from ByteDance available via API. It was included in the WBench evaluation across quality and setting adherence dimensions.
Xunliang Cai: Xunliang Cai is a researcher affiliated with Fudan University and the Meituan LongCat Team on the WBench project. He contributed to the overall framework that enables fair comparison across diverse video model families.
Fantasy-World: Fantasy-World is an open-source model from Amap included in the WBench evaluation. It helped illustrate varying strengths across different model families.
Fengjiao Chen: Fengjiao Chen is a researcher affiliated with Fudan University and the Meituan LongCat Team on the WBench effort. She supported the design of the five-dimensional evaluation covering visual quality through physical plausibility.
LingBot-World: LingBot-World is an open-source video world model developed by Ant Group. It participated in the WBench evaluation highlighting strengths in consistency among tested systems.
LongCat-Video: LongCat-Video is an open-source video model from Meituan that participated in the WBench evaluation. It was tested alongside other systems for controllability and memory.
InSpatio-World: InSpatio-World is an open-source video world model from InSpatio evaluated in WBench. It was part of the broad assessment of navigation and consistency.
Infinite-World: Infinite-World is an open-source video world model from Meituan evaluated in WBench. It was assessed for multi-turn performance in the benchmark.

Key Finding: No evaluated video world model excels simultaneously across visual quality, setting adherence, interaction adherence, consistency, and physical plausibility.
Benchmark Design: WBench separates initial world setup from subsequent user actions to isolate failures in rendering, scene control, memory, or physics during extended interactions.
Capability Insight: Navigation performance shows near-zero correlation with visual quality, consistency, or physics compliance in current video generation systems.

AI tools built by Emerald Force

Built and supported by Emerald Force.

You might also like