acceptodds
Withdrawn from ICLR 2027

MetaDyn-JEPA: Learning to Predict across Physical Worlds with Interaction Evidence

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Physical interaction data often span systems with different underlying dynamics: the same action from a similar observed configuration can produce different outcomes. A short observation history may therefore be insufficient to predict the consequences of an action. We propose MetaDyn-JEPA (MetaDyn), a support-conditioned joint-embedding world model that uses additional interactions from the same physical system to inform prediction. MetaDyn encodes directed action–response transitions and aggregates them into a reusable world code that conditions multi-step latent prediction without test-time weight updates. Treating transitions as the units of evidence allows MetaDyn to aggregate support from disconnected fragments without relying on fragment-level sequence summaries. Across PokeWorld and MuJoCo, same-world support improves prediction over no-support baselines, while replacing it with wrong-world support degrades performance. MetaDyn remains stable under evidence-preserving repartitioning without fragment augmentation, and generally benefits from increasing same-world support toward the training length. In multi-dynamics Push-T, same-world support improves latent and physical prediction, and MetaDyn achieves higher reach-and-hold success than LeWM and LeWM + AdaJEPA. Overall, these results show that world-specific interaction evidence can provide a useful conditioning signal for prediction across varying dynamics.

Primary area:
learning on time series and dynamical systems
Keywords:
World Models, Joint-Embedding Predictive Architectures

Related papers

Discussion (0)

Sign in to comment.