Liouville-Regularized Flow-Matching World Models
- Proposed Liouville (squared divergence) and Jacobian Frobenius pointwise regularizers for flow-matching latent world models, reducing long-horizon latent drift by 17 orders of magnitude at H=128 on DMControl Walker-Walk versus an unregularized baseline.
- Formally proved cycle-consistency is structurally vacuous for rectified flows in the autonomous limit.
- Confirmed empirically that Liouville variants achieve cycle-error floors 3 orders of magnitude lower than naive cycle regularization across all solver tolerances.
Abhijit Chunduru
ML Systems Researcher | M.S. Computer Science
Distributed training, LLM post-training, and scientific ML at Argonne National Laboratory and UMass Amherst