V-JEPA-LM uses counterfactual supervision to enable language models to read direction, velocity, order, and state changes in video encoders, rather than merely retrieving text-aligned static attributes. The model is trained in a single stage using 4.53M public samples and 32 visual tokens per frame. Its 4B model achieves an average score of 72.1 across 11 video benchmarks, outperforming the best open-source 4B model, which scores 61.1.
Source:Hugging Face 社区论文精选 · Read Original
Leave a Reply