Megadose The AI curve, compressed.

V-JEPA-LM: Video-Language Models That Read Dynamics Beyond Static, Text-Aligned Representations

· ArXiv · AI/CL/LG ·
V-JEPA encodes speed well enough for 87% probe accuracy, yet a language model using those features answers at chance.

The authors train the language model to use those dynamics through transformed videos whose answers change with the transformation. At a matched training budget, this counterfactual supervision adds 4.5 points on average and 38 points on speed questions in natural benchmarks. Their 4B model averages 72.1 across eleven video benchmarks, compared with 61.1 for the best open 4B model and 64.8 for the best open 8B model. Training uses one stage and 4.53 million public samples.

ArXiv · AI/CL/LG's note

score 5

Categories: Research