New technique enables continuous learning of AI models during inference

A collaboration between researchers and NVIDIA has produced a new approach for training artificial intelligence models. Called "End-to-End Test-Time Training," the technique allows models to update their parameters while processing long contexts during inference. The core idea is that AI models should function as continuous learners, similar to human researchers, compressing accumulated experience into intuition.