Thinking in Meaning Space, Not Token Space.
VL-JEPA allows an AI to understand the relationship between vision and language without converting everything into discrete text tokens. It treats text as a continuous embedding target. This solves the Symbol Grounding Problem, where LLMs know the "word" but not the "world."
By using Selective Decoding, VL-JEPA only generates text when necessary, staying in the abstract "Meaning Space" for the majority of its reasoning.
Select a text prompt to measure its distance from the video embedding. In a trained JEPA, embeddings are static points; their proximity is determined during training to ensure semantic compatibility results in low 'Energy' (short distance).