FlowWM
Summary
FlowWM is a passive stochastic visual world model that performs conditional flow matching over future trajectories in frozen DINOv3 feature space. It is the closest current empirical precedent in this wiki for replacing a deterministic future representation with multiple sampled semantic future trajectories.
Role In The Wiki
FlowWM anchors the feature-space-flow branch of probabilistic world modeling. It demonstrates that flow matching can operate in a high-dimensional pretrained semantic feature space when the velocity head, timestep schedule, and auxiliary objectives are adapted to that space.
The model is JEPA-adjacent rather than a JEPA implementation: DINOv3 is frozen, no JEPA encoder/target objective is trained, and the predictor is passive rather than conditioned on actions or control inputs.
Evidence Boundary
FlowWM provides strong controlled evidence for sample validity and mode coverage on Bouncing Shapes, plus downstream detection and depth improvements on FuturePerception. It does not establish calibrated mode probabilities, raw-versus-latent superiority, action-conditioned rollouts, planning utility, rare-tail preservation, or numeric time-series transfer.
FuturePerception’s AP(N) is a best-of- oracle metric. It shows that useful future proposals are present among the samples, not that the model can assign correct probabilities or select the right rollout before the future is observed.
Official Artifacts
The verified 2026-07-19 repository snapshot contains training and benchmark code but no released checkpoints or GitHub releases.
Evidence
Relation To Foundation TSFM Agenda
Use the source-level agenda mapping in flowwm-2026 rather than duplicating verdict rows here.
For time-series work, FlowWM is a mechanism and benchmark precedent: generate structured future latent trajectories with flow matching, supervise endpoint properties through one-step projection, and test mode precision and recall on an enumerable branching system. The missing step is to move from passive visual features to calibrated multivariate time-series trajectories under typed actions, control inputs, interventions, and exogenous variables.