Kunlunx Robot's World Model Wins WorldArena 2.0 Image Quality Champion, Secures Global Runner-up

The benchmark, co-built by international top universities and research institutions, is seen as an "international arena and authoritative touchstone" for embodied world models. Track 1 evaluates basic capabilities such as accurate perception, correct prediction, and physically plausible rendering. This time, 77 models from around the world participated, and GeWu was the only one to rank in the top four in four of the six evaluation dimensions.

According to the report, the model showed significant leads in physics adherence and controllability. The company attributes this to its causal architecture, which differs from mainstream modality-based designs. The Kunlun World Model splits into three Transformer towers for intent, intervention, and consequence, with a united causal attention mechanism that makes action decisions blind to future visual outcomes while rendering can see both.

Ablation experiments demonstrated the necessity of this design. When any causal pathway was cut, the model degraded from causal simulation to a "copy mode," producing physically impossible results, such as objects moving without corresponding actions. In the interaction quality dimension, the model scored 82.40, a direct reflection of the causal architecture. The company also developed a chunk-wise autoregressive strategy with a bio-inspired memory sampling mechanism, keeping long-sequence video generation consistent.

Hao Zhihui, head of model research at Kunlunx, said in the report that the architecture should respect the causal structure of the physical world. "Intent drives action, action leads to consequence. This causal chain is the first principle." He added that aligning inductive bias with the problem essence improved metrics, training efficiency, and generalization.

Kunlunx Robot, founded with the vision of building general-purpose AI robots, follows a "body + brain" dual-driven strategy. The team will continue validating and expanding the model in more complex tasks, applying it to behavior planning and simulation-based reinforcement, aiming to evolve it from a visual renderer into a physical intelligence foundation.