Researchers from New York University, in collaboration with Yann LeCun's startup AMI, have developed AdaJEPA, a world model capable of continuous learning. Unlike previous models that freeze parameters after pre-training, AdaJEPA adapts its encoder and predictor in real-time through interaction with its environment. This adaptation, achieved via a plan-execute-observe-update-replan loop, allows the model to continuously refine its understanding of the world, leading to improved planning success rates even in out-of-distribution scenarios. AI
IMPACT Enables more robust and adaptable AI agents by allowing world models to continuously learn and self-correct during deployment.
RANK_REASON The cluster describes a new research paper detailing a novel approach to world models in AI. [lever_c_demoted from research: ic=1 ai=1.0]
AI-generated summary · Google Gemini · from 1 sources. How we write summaries →
