jepa

The paper introduces AC-MTM, a contrastive inverse dynamics method to prevent encoder collapse in JEPA world models, achieving improved performance on multi-object tasks without Gaussian constraints.

The paper establishes a theoretical connection between probabilistic Joint-Embedding Predictive Learning (JEPA) and Hidden Markov Models (HMMs), providing a state-space interpretation and introducing Markov-Chain JEPA for enhanced consistency.

Introduces Action-Conditioned Predictive Consistency (ACPC), a diagnostic for JEPA world models that measures how clean and perturbed observations diverge under action-conditioned rollouts, with theoretical bounds on prediction error and planner cost. Experiments on visual control tasks validate the diagnostic across models like LeWM and PLDM.

BioM-JEPA introduces a joint-embedding predictive architecture that learns single-cell representations by predicting graph-connected gene blocks instead of individual genes, showing improved efficiency and downstream performance in perturbation-response tasks.

This paper introduces NodeJEPA, a joint-embedding predictive architecture for node-level graph self-supervised learning that predicts latent representations of masked structure-aware ego-subgraphs, avoiding reconstruction and hand-crafted augmentations. The method is evaluated on node classification benchmarks and shows competitive performance.

SJEPA introduces a reconstruction-free JEPA framework that learns hybrid symbolic-neural latent dynamics, aiming for the simplest adequate predictive representation. Experiments show it discovers simpler symbolic dynamics with lower rollout error than post-hoc fitting, while controlling symbolic-neural allocation under grammar misspecification.

This paper introduces HP-JEPA, a hierarchical partitioning framework for multi-resolution graph joint-embedding predictive learning, which outperforms the fixed-resolution Graph-JEPA baseline on most graph classification and regression benchmarks.

Introduces DENSEWORLD, a 1,000-hour dataset of crowded Global South urban scenes, and FactorJEPA, a JEPA variant that factorizes future prediction into layout, agents, and interactions, improving accuracy and robustness under occlusion and heterogeneity.

This paper introduces Collective-State JEPA (CS-JEPA), a recurrent joint-embedding predictive architecture that lets every robot in a swarm predict the same future collective state from local observations and limited messages. The method shows label-efficient improvements in prediction error and inter-robot agreement, plus planning-relevant value estimation.

This paper analyzes why deterministic JEPA-style latent prediction works for images but not for text, attributing the failure to high conditional variance in language where masked contexts admit multiple valid completions whose representations lack a coherent center.

INTACT is an end-to-end unified JEPA that learns the intent-to-action mapping directly, enabling search-free world model control. It achieves 95.33% direct macro success rate across four visual-control tasks with zero test-time search and ~300x lower planning latency.

An analysis of Yann LeCun's bet that intelligence starts with world models via JEPA, not language, supported by AMI Labs' $1.03 billion funding. The article explains why next-pixel prediction fails and how JEPA predicts in latent space to avoid blurry futures.

Induction Labs introduces imagination models, a new foundation model architecture that learns from internet-scale video. Their first model, Photon-1, learns to use a computer by watching 18 years of screen recordings without action labels, achieving better results at 30× lower pretraining cost than Gemini 3.1 Flash.

Explains SIGReg, a novel regularizer for JEPA that prevents representation collapse by forcing embeddings to follow an isotropic Gaussian distribution, with theoretical guarantees and a clean training loop.

Yann LeCun discusses his new lab AMI Labs in Paris, his focus on world models as an alternative to LLMs, and his vision for AI that can plan and act in the physical world, backed by $1B investment and using JEPA architecture.

AV-JEPA extends LeJEPA to audio-visual self-supervised learning, achieving cross-modal alignment in latent space without decoders, contrastive negatives, or complex losses, and obtains competitive classification on VGGSound and AudioSet.

This paper establishes a formal correspondence between the SIGReg objective used in Joint-Embedding Predictive Architectures (JEPAs) and the Active Inference framework, showing that under certain conditions the training objective becomes an exact variational free energy. It organizes non-contrastive regularisers into an entropy-estimator hierarchy and proves theoretical results including exact information bottleneck and preserved surprise bounds for SIGReg.

The author reproduces a Joint-Embedding Predictive Architecture (JEPA) world model and trains it on Super Mario Bros, finding that while it predicts frames well, it fails at long-term planning.

A step-by-step annotated implementation and explanation of Joint Embedding Predictive Architectures (JEPA) for self-supervised learning, covering I-JEPA, V-JEPA, and LeJEPA.

This paper explores applying JEPA-style predictive learning to JA4-derived network fingerprints, building a Transformer-based model (JA4-JEPA) trained on JA4, JA4H, JA4S, and JA4X subfields. The model achieves strong performance on protocol-family classification, suggesting JEPA objectives can work for compact network fingerprint representations.