Research map / Large language models / Language, safety and interpretability
LLM-JEPA: Large Language Models Meet Joint Embedding Predictive Architectures
Hai Huang, Yann LeCun, Randall Balestriero
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 |
|---|---|---|---|
| add JEPA loss (lambda=1,k=1) to next-token fine-tuning | Llama-3.2-1B-Instruct · NL-RX-SYNTH | accuracy: 57.29 -> 71.46 | Improved |
| add JEPA loss to fine-tuning of Llama-3.1-8B | Llama-3.1-8B-Instruct · NL-RX-SYNTH | accuracy: 35.77 -> 63.57 | Improved |
| add JEPA loss to fine-tuning of OLMo-2-0425-1B | OLMo-2-0425-1B-Instruct · NL-RX-SYNTH | accuracy: 87.09 -> 87.52 | No clear effect |
| add JEPA loss to pretraining from scratch | Llama-3.2-1B-Instruct · NL-RX-SYNTH pretraining | accuracy: 54.38 -> 60.59 | Improved |
| add JEPA loss to fine-tuning on GSM8K | Llama-3.2-1B-Instruct · GSM8K | accuracy: 32.36 -> 36.36 | Improved |
| add JEPA loss to fine-tuning on NQ-Open | Llama-3.2-1B-Instruct · NQ-Open | accuracy: 20.12 -> 21.59 | Improved |
| replace cosine similarity with l2-norm in JEPA loss (ablation) | LLM-JEPA · NL-RX-SYNTH | accuracy: 71.46 -> 2.22 | Hurt |
| replace cosine similarity with MSE (ablation) | LLM-JEPA · NL-RX-SYNTH | accuracy: 71.46 -> 70.64 | Hurt |
| append predictor token instead of prepending (ablation) | LLM-JEPA · NL-RX-SYNTH | accuracy: 68.07 -> 71.46 | Improved |
| predict Code from Text instead of Text from Code (ablation) | LLM-JEPA · NL-RX-SYNTH | accuracy: 65.70 -> 71.46 | Improved |
| use cosine JEPA loss instead of InfoNCE (ablation) | LLM-JEPA · NL-RX-SYNTH | accuracy: 34.40 -> 71.46 | Improved |
| add loss dropout LD=0.75 at matched compute (first row) (ablation) | LLM-JEPA · NL-RX-SYNTH | accuracy: 19.85 -> 32.46 | Improved |
Later papers that compared with this method
1 later accepted paper reports LLM-JEPA as a baseline in their results tables; 1 of them report a better number on most of the results they share with it.
- Semantic Tube Prediction: Beating LLM Data Efficiency with JEPA (ICML 2026 · better on 2 of 2 shared results)
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- M3-JEPA: Multimodal Alignment via Multi-gate MoE based on the Joint-Embedding Predictive Architecture (ICML 2025 · similarity 0.88)
- Large Language Diffusion Models (NeurIPS 2025 · similarity 0.88)
- EmbedLLM: Learning Compact Representations of Large Language Models (ICLR 2025 · similarity 0.87)
- A Paradigm Shift in Machine Translation: Boosting Translation Performance of Large Language Models (ICLR 2024 · similarity 0.87)
- Semantic Tube Prediction: Beating LLM Data Efficiency with JEPA (ICML 2026 · similarity 0.87)
- LLaDA-V: Large Language Diffusion Models with Visual Instruction Tuning (CVPR 2026 · similarity 0.87)
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.
