Temporally Invertible Latent World Models via Flow Matching

Mar 1, 2026
  • Introduced temporal cycle consistency in ODE dynamics.
  • Improved H16 cosine similarity from 0.247 to 0.87.
  • Enabled reverse goal-conditioned planning with 67% success versus 0% baseline.
World Models Generative Modeling Reinforcement Learning
Abhijit Chunduru
Abhijit Chunduru
Master’s Student

ML Systems and Trustworthy AI at UMass Amherst