V-JEPA-LM: Video-Language Models That Read Dynamics Beyond Static, Text-Aligned Representations
November 12, 2026
V-JEPA-LM addresses the gap between video dynamics and language alignment by introducing counterfactual supervision to video-language models. While frozen V-JEPA encoders achieve 87% accuracy on speed linear probes, standard VLMs currently perform at chance levels. This method forces the language model to utilize the encoder's predictive dynamics rather than relying on static, text-aligned attributes.