Where V-JEPA 2.1's Dense Features Hold Up (and Where They Don't) (poissonlabs.ai)

A recent robustness study on Meta’s V-JEPA 2.1, which was released in March 2026, reveals significant insights into its dense feature representations across four model sizes (ranging from 80M to 2B parameters). The study found that while V-JEPA 2.1's features are effective in predicting failures under temporal corruption (like frame drops and occlusion), they show no correlation with performance under image-noise corruption. Additionally, the robustness did not consistently improve with model size, as the largest 2B model was less robust than the 1B variant in several scenarios. This non-monotonic behavior defies assumptions and indicates that scaling may not guarantee better performance. The findings underscore practical implications for deploying V-JEPA 2.1 in robotics, particularly in complex environments. For instance, in industrial cable insertion and drone infrastructure inspection, where visual conditions vary drastically, model selection must be empirical rather than relying on size alone. The research emphasizes that V-JEPA 2.1's orientation sensitivity further complicates feature stability, suggesting potential challenges in its application to real-world tasks. Ultimately, this study provides crucial insights for AI/ML practitioners considering the implications of model architecture and size on deployment efficacy.