@LiorOnAI: Most models behind agents don't learn while they're running. You train them, freeze the weights, and deploy them. Every…
Summary
Introduces AdaJEPA, an adaptive world model that continuously learns and updates its latent representation during deployment, enabling agents to adjust their plans based on real-world observations without retraining or memory tricks.
View Cached Full Text
Cached at: 07/06/26, 02:02 AM
Most models behind agents don’t learn while they’re running.
You train them, freeze the weights, and deploy them.
Everything else gets built around them: prompts, tools, retrieval, memory, routing, guardrails.
That works well for chat because every interaction has a clean boundary. The user asks a question, the model answers, and the loop ends.
AdaJEPA updates the world model inside the control loop instead.
The agent takes an action, sees what actually happened, nudges the latent model toward reality, then replans.
No retraining run. No bigger context window. No memory trick.
Just make the model a little less wrong before the next decision.
Ying Wang✈️ ICML (@yingwww_): Warm take: Your world model should never stop learning
Introducing AdaJEPA, an adaptive WM that plans, acts, and adapts in a closed loop. Every action leads to a new observation, and every transition refines the latent representation and prediction.
📝:
Similar Articles
@yingwww_: Warm take: Your world model should never stop learning Introducing AdaJEPA, an adaptive WM that plans, acts, and adapts…
AdaJEPA introduces an adaptive latent world model that continuously updates during test-time via closed-loop model predictive control, significantly improving planning success under distribution shift.
@mengyer: New research: AdaJEPA perceives, plans, and adapts in a closed loop. It is always learning!
Introduces AdaJEPA, an adaptive world model that continuously learns from perception, planning, and action in a closed loop.
Agent-World: Scaling Real-World Environment Synthesis for Evolving General Agent Intelligence
Agent-World introduces a self-evolving training framework for general agent intelligence that autonomously discovers real-world environments and tasks via the Model Context Protocol, enabling continuous learning. Agent-World-8B and 14B models outperform strong proprietary models across 23 challenging agent benchmarks.
@LiorOnAI: Most world models predict what happens next. Sora predicts pixels, JEPA compresses observations. NEO tries to figure ou…
NEO is a new type of world model that learns to discover reusable building blocks of explanation from raw observations without supervision or language, selected as an ICML 2026 oral presentation.
@LeRobotHF: VLA-JEPA just dropped in LeRobot What makes this model special is that it does not just learn what action to take from …
VLA-JEPA, a new model integrating a JEPA world model for action-relevant dynamics, has been released in LeRobot. It enables pretraining on human videos and achieves strong performance with minimal fine-tuning, running real-time on NVIDIA DGX Spark.