Research map / Vision-language and multimodal / Human and embodied modeling
VL-JEPA: Joint Embedding Predictive Architecture for Vision-language
Delong Chen, Mustafa Shukor, Théo Moutakanni, Willy Chung, Lei Yu, Tejaswi Kasarla, Allen Bolourchi, Yann LeCun, Pascale Fung
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| predict target embeddings instead of tokens (strictly matched training) | VL-JEPA vs VLM, 15M samples · zero-shot video captioning | CIDEr: 7.1 -> 14.8 | Improved |
| predict target embeddings instead of tokens (strictly matched training) | VL-JEPA vs VLM, 15M samples · zero-shot video classification | top-5 accuracy: 27.2 -> 41.0 | Improved |
| add caption pretraining stage before SFT (ablation) | VL-JEPA · ablation, classification and retrieval | accuracy / recall@1: -21.7 classification, -17.3 retrieval without it | Improved |
| embedding-guided selective decoding instead of uniform decoding (ablation) | VL-JEPA · EgoExo4D streaming | decoding operations at matched CIDEr: ~2.85x fewer decodes (0.35 Hz matches 1 Hz) | Improved |
| average pooling of embeddings before decoding (ablation) | VL-JEPA selective decoding · EgoExo4D streaming | CIDEr | Improved |
| Y-Encoder learning rate multiplier 1.00 instead of 0.05 (ablation) | VL-JEPA · ablation, classification | accuracy: 27.3 -> 23.7 | Hurt |
| Y-Encoder learning rate multiplier 0.00 (frozen) instead of 0.05 (ablation) | VL-JEPA · ablation, classification | accuracy: 27.3 -> 20.0 | Hurt |
| cosine, L1 or L2 loss instead of InfoNCE (ablation) | VL-JEPA · ablation, classification and retrieval | accuracy / recall@1: classification 23.3 -> 16.5 (cosine), 13.5 (L2) | Hurt |
| cosine loss instead of InfoNCE (ablation) | VL-JEPA · ablation, VQA | accuracy: 44.3 -> 46.6 | Improved |
| more predictor layers (Layer 0-16 vs 8-16) (ablation) | VL-JEPA · ablation, VQA | accuracy: 42.5 -> 45.5 | Improved |
| causal attention instead of bi-directional attention in predictor (ablation) | VL-JEPA · ablation, VQA | accuracy: 42.5 -> 40.6 | Hurt |
| random initialization instead of Llama-3 initialization of predictor (ablation) | VL-JEPA · ablation, VQA vs classification | accuracy: VQA 42.5 -> 40.6, classification 27.3 -> 28.1 | Mixed |
Closest papers by meaning
- Bridging Vision and Language Spaces with Assignment Prediction (ICLR 2024 · similarity 0.90)
- VLM2Vec: Training Vision-Language Models for Massive Multimodal Embedding Tasks (ICLR 2025 · similarity 0.89)
- Vision‑Language‑Vision Auto‑Encoder: Scalable Knowledge Distillation from Diffusion Models (NeurIPS 2025 · similarity 0.89)
- VL-SAE: Interpreting and Enhancing Vision-Language Alignment with a Unified Concept Set (NeurIPS 2025 · similarity 0.89)
- From Pixels to Words -- Towards Native Vision-Language Primitives at Scale (ICLR 2026 · similarity 0.88)
- ViSpec: Accelerating Vision-Language Models with Vision-Aware Speculative Decoding (NeurIPS 2025 · similarity 0.88)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
