New JEPA method uses contrastive inverse dynamics to stabilize world models

A new research paper introduces Action-Contrastive Masked Transition Modeling (AC-MTM) as an alternative to Gaussian-based regularization in Joint-Embedding Predictive Architectures (JEPAs). This method uses an inverse dynamics head during training to identify the action that produced a latent transition, preventing collapse without requiring a constant encoder. AC-MTM demonstrates competitive performance on standard tasks and significantly outperforms Gaussian regularization on the OGBench Visual Scene task, suggesting that distribution-free contrastive signals can effectively stabilize world models. AI

IMPACT Introduces a novel, distribution-free method for training world models, potentially improving performance on complex visual tasks.

RANK_REASON Research paper introducing a new method for world models in AI. [lever_c_demoted from research: ic=1 ai=1.0]

Read on arXiv cs.AI →

• Action-Contrastive Masked Transition Modeling
• Action-NCE
• Assran et al.
• Bardes et al.
• JEPAs
• LeWorldModel
• OGBench Visual Scene
• SIGReg
• Yann Le Cun

• paper
• model release

AI-generated summary · Google Gemini · from 1 sources. How we write summaries →