EVEWorld cuts world-model 'laziness' rate 85.7% in robot videos
Researchers at China Merchants Group's Advanced Technology Research Institute and Nanyang Technological University (NTU) propose EVEWorld, a post-training supervision framework that enforces physical consistency of manipulated objects in world models: Instance-Guided Recovery (IGR) builds object-duplication perturbations and trains the model to restore the correct count, while Temporal Instance Alignment (TIA) matches and transfers features across adjacent frames in selected layers of a video diffusion Transformer. The authors define unjustified object duplication or disappearance as "Model Laziness" and measure it with a Model Laziness Rate (MLR). According to the paper, EVEWorld lowers MLR on DreamGenBench from 11.11% to 1.59%, an 85.7% relative drop, while Qwen-IF and Gemini-IF scores rise 6.35 and 7.28 points; on WorldArena 1.0, persistently anomalous videos fall from 43 to 21 of 157 shared evaluable samples (about 51.1% relative), and transferred to FlowWAM it cuts RoboTwin MLR from 47.89% to 35.21%. The authors note the method's limitation: its local pairwise supervision does not explicitly model long-horizon state dependencies or complex multi-object interactions.
Object-count and identity consistency is a precondition for using world models in robot prediction and planning, so this kind of post-training supervision suggests physical consistency deserves its own place in training and evaluation.
EVEWorldInstance-Guided Recovery (IGR)Temporal Instance Alignment (TIA)