Frozen Flows Forget: Diagnosing and Restoring Lost Motion in a Latent-flow World Model
Latent world models that integrate a flow in a frozen self supervised latent space train stably and cheaply, yet silently lose the property manipulation depends on most: motion. The pretrained flow never moves the manipulated object; retraining it with latent-only losses only trades stillness for teleport-like motion. We trace the failure to the training signal, not the represe
Reading in Plain English mode — toggle below to switch
Latent world models that integrate a flow in a frozen self supervised latent space train stably and cheaply, yet silently lose the property manipulation depends on most: motion. The pretrained flow never moves the manipulated object; retraining it with latent-only losses only trades stillness for teleport-like motion. We trace the failure to the training signal, not the represe
This paper addresses a critical limitation in latent world models, specifically their inability to accurately model motion, which is crucial for agentic behavior and planning. Understanding how to restore motion in frozen latent spaces is vital for developing more robust and efficient AI agents that can interact dynamically with their environments.