ODEWorld: A Continuous Predictive Architecture via Physical-Time Flow
ODEWorld treats future prediction as continuous-time integration inside a compressed latent space.
The paper introduces Physical-Time Flow, which learns a latent velocity field governed by an ordinary differential equation. That lets ODEWorld predict at arbitrary temporal resolutions and run backward in time, unlike most discrete-time world models. The authors say the design also avoids representation collapse and preserves image reconstruction quality over long horizons. Experiments are reported on video generation and robotic control. HF Daily Papers' note
The paper introduces Physical-Time Flow, which learns a latent velocity field governed by an ordinary differential equation. That lets ODEWorld predict at arbitrary temporal resolutions and run backward in time, unlike most discrete-time world models. The authors say the design also avoids representation collapse and preserves image reconstruction quality over long horizons. Experiments are reported on video generation and robotic control. HF Daily Papers' note
score 5