No Gaussian Required: Contrastive Inverse Dynamics for JEPA World Models
Researchers have proposed a new approach to prevent a problem called 'collapse' in world models, where the model's internal representation becomes too simple and loses information about the environment. The new method, called Action-Contrastive Masked Transition Modeling (AC-MTM), uses a training-only inverse-dynamics head that identifies the action that produced each latent transition among other actions in the batch. This approach does not require a target network, stop-gra
Researchers have proposed a new approach to prevent a problem called 'collapse' in world models, where the model's internal representation becomes too simple and loses information about the environment. The new method, called Action-Contrastive Masked Transition Modeling (AC-MTM), uses a training-only inverse-dynamics head that identifies the action that produced each latent transition among other actions in the batch. This approach does not require a target network, stop-gradient, pretrained encoder, or reconstruction objective, and is shown to be effective on several tasks, including the OGBench Visual Scene task where it outperforms previous methods by 20-24 points.
---
Why it matters: This matters because it provides an alternative solution to preventing collapse in world models, which can have a significant impact on the performance of AI systems. By avoiding the need for additional components and objectives, AC-MTM can make world modeling more efficient and effective.
Source: https://arxiv.org/abs/2608.17542
This article was originally published at: https://arxiv.org/abs/2608.17542