Benchmarks and Results

A consolidated reference of quantitative results across the JEPA family, organized by task category. All numbers are from the respective papers; error bars are reported where available in the original work. Results without error bars should be interpreted as point estimates from single training runs unless noted otherwise.

Video understanding

Action recognition (top-1 accuracy)

SSv2 (Something-Something v2) is the key motion-understanding benchmark — it requires temporal reasoning, not just appearance. V-JEPA 2.1's ViT-G holds the JEPA SOTA.

Action anticipation (Epic-Kitchens-100, Recall@5)

V-JEPA models surpass task-specific baselines by large margins (+12 points on action R@5 for V-JEPA 2 over PlausiVL).

Video question answering (with LLM alignment)

SOTA at the 8B parameter scale across multiple VQA benchmarks.

Short-term object anticipation (Ego4D)

+35% relative improvement over prior SOTA.

Image understanding

Classification (ImageNet, top-1 accuracy)

JEPA approaches but doesn't quite match DINOv2 on ImageNet — expected since DINOv2 is trained specifically for image classification.

Dense understanding (linear probe)

The Dense Predictive Loss in V-JEPA 2.1 is transformative for dense tasks — +20-28 mIoU improvement in segmentation over V-JEPA 2.

Video object segmentation (J&F-Mean)

3D understanding

Point cloud classification (ModelNet40)

ScanObjectNN (real-world 3D)

3D-JEPA achieves strong results with half the pretraining epochs of competing methods.

Few-shot learning

Reasoning

Visual QA (CLEVRER)

Object-level masking gives +21% absolute on counterfactual reasoning.

Robot control

Manipulation success rate

Planning efficiency

Robotic navigation (Tartan Drive)

Trajectory prediction

Hand manipulation (EgoDex)

ThinkJEPA outperforms both pure JEPA and pure VLM approaches, especially on long-horizon rollouts.

See also

• world-models-and-planning — robot control and planning details
• training-recipes — what training setup produces these results
• vision-transformers — how model scale affects performance