V-JEPA 2
PulseAugur coverage of V-JEPA 2 — every cluster mentioning V-JEPA 2 across labs, papers, and developer communities, ranked by signal.
5 day(s) with sentiment data
-
New Market-1T Dataset Enables Deeper Financial World Modeling
Researchers have introduced Market-1T, a massive dataset comprising nearly one trillion observations of U.S. equities from 2008 to 2025 at 1 Hz resolution. This dataset, along with a rigorous evaluation protocol, facili…
-
Video world models struggle to track hidden objects, study finds
Researchers have investigated the ability of video world models to retain information about objects that are no longer visible. Experiments using V-JEPA 2 revealed that these models struggle to maintain knowledge of sta…
-
JEPA-Anything framework unifies world model creation across 7 domains
Researchers have introduced JEPA-Anything, a novel domain-agnostic framework designed to build world models. This framework extends joint-embedding predictive architectures (JEPAs) by incorporating Orthogonal Predictive…
-
New research enhances World-Action Models for robotics and AI
Recent research explores advancements in World-Action Models (WAMs) for robotics and AI, focusing on improving prediction accuracy, action generation, and inference efficiency. Several papers introduce new methods like …
-
New MoE-JEPA model sets state-of-the-art in synthetic image detection
Researchers have developed MoE-JEPA, a novel dual-stream architecture for detecting synthetic and manipulated images. This model enhances a V-JEPA 2 backbone with a Residual Mixture-of-Experts mechanism and a noise stre…
-
New CALIPER benchmark reveals limitations in AI visual physics understanding
Researchers have developed a new benchmark called CALIPER to more accurately assess the physical reasoning capabilities of pretrained visual models. Traditional methods using clean, static scenes fail to distinguish bet…
-
New ARC-Bench protocol reveals critical flaws in frozen JEPA world models
Researchers have developed ARC-Bench, a new evaluation protocol designed to assess the action ranking capabilities of frozen latent world models. The study found that these models, which plan by scoring candidate action…
-
New AI models advance robot manipulation with improved memory and planning
Researchers have developed new methods for robot manipulation that improve performance in long-horizon tasks. The 2AM system separates task memory from the action model, allowing for more precise control and achieving a…
-
Video foundation models analyzed for spatiotemporal understanding
Researchers have analyzed two video foundation models, V-JEPA 2 and VideoMAE-v2, to understand their spatiotemporal representations. The study found that both models effectively encode camera motion and exhibit moderate…
-
Meta AI's V-JEPA 2: A New Approach to World Models
Meta AI has developed V-JEPA 2, a novel world model that differs from typical generative models. Unlike models focused on generating new data, V-JEPA 2 prioritizes understanding and predicting the underlying structure o…
-
China Merchants AI Lab debuts flexible object manipulation tech
The Lion Mountain Artificial Intelligence Laboratory, under China Merchants Group's Advanced Technology Research Institute, has unveiled its full-stack self-developed capabilities in embodied intelligence and flexible o…
-
JEPA predictors prove portable for occluded feature completion
Researchers have demonstrated that the predictor component of Joint-Embedding Predictive Architectures (JEPAs), typically discarded after training, can be repurposed as a transferable operator for occluded feature compl…
-
VideoRAE leverages VFM features for improved video generation
Researchers have introduced VideoRAE, a novel representation autoencoder designed to enhance video generative models. This system leverages features from frozen Video Foundation Models (VFMs) like V-JEPA 2 and VideoMAEv…
-
VideoRAE enhances generative video models using frozen foundation features
Researchers have introduced VideoRAE, a novel representation autoencoder designed to enhance generative video modeling. Unlike traditional methods that focus on pixel-level reconstruction, VideoRAE leverages multi-scale…
-
New PULS system anticipates video anomalies using semantic world-model pipeline
Researchers have developed PULS (Predictive Unified Latent Space), a novel pipeline for continuous video anomaly detection that moves beyond reactive methods. PULS consists of a KSD Bridge, which translates physical ten…
-
New Cross4D-JEPA method distills 2D models for 4D point cloud understanding
Researchers have introduced Cross4D-JEPA, a novel self-supervised learning method for understanding dynamic 4D point clouds. This approach distills knowledge from 2D image or video foundation models, such as DINOv2 and …
-
New research advances 3D scene graph generation for robotics and AR
Three new research papers introduce advanced methods for generating 3D semantic scene graphs, which are crucial for understanding and interacting with 3D environments. DeWorldSG utilizes world-model priors and probabili…
-
AI predicts joint forces from video, bypassing invasive methods
Researchers have developed a novel pipeline capable of predicting in vivo joint contact forces from monocular video without invasive measurements or subject-specific models. This system utilizes parametric body meshes a…
-
TrAction uses sparse trajectories for efficient action recognition
Researchers have developed TrAction, a novel transformer architecture for action recognition using sparse point trajectories instead of dense video. This method aims to reduce biases found in traditional models that rel…
-
Lion Rock Lab unveils LiOS to connect cloud AI models with robots
Researchers at the China Merchants Group Lion Rock Artificial Intelligence Lab have developed LiOS, a system designed to bridge the gap between cloud-based AI models and real-world robotic applications. This system addr…