AdaJEPA

Summary

AdaJEPA is an adaptive JEPA-style latent world model for closed-loop model predictive control. Instead of freezing the world model after offline training, it performs lightweight test-time adaptation after each executed action chunk, using the observed next transition as a self-supervised latent-prediction target before replanning.

Role In The Wiki

AdaJEPA is the online-adaptation branch of the local JEPA/world-model line. Compared with LeWorldModel, it changes the deployment loop rather than only the offline training objective. Compared with stable-worldmodel, it is a model/update method rather than an evaluation platform. Compared with SkyJEPA, it is demonstrated on PushT/PushObj and PointMaze distribution shifts rather than an outdoor quadrotor platform.

Evidence

Official Artifacts

Artifact caveat: at ingest time, the project URL resolved to the Agentic Learning AI Lab site rather than a dedicated AdaJEPA page, and no official code repository or official X announcement was verified. Local artifact-discovery notes are stored at papers/adajepa-2026/official_artifacts_snapshot.md.

Relation To Foundation TSFM Agenda

Use the source-level agenda mapping in adajepa-2026 rather than duplicating verdict rows here.

At the entity level, AdaJEPA is a named method for this interface:

observation + action/control-input history
  -> latent world-model rollout
  -> MPC action choice
  -> observed transition
  -> lightweight self-supervised update
  -> replan

For the foundation time-series agenda, the transferable lesson is that world models may need online calibration from their own closed-loop experience. The current evidence is still outside numeric time series and does not solve the full always-on streaming, intervention, safety, or benchmark protocol requirements.