Researchers introduced TTT-E2E, a method that lets language models learn from context during inference by updating their weights. The method requires more complex and slower training, but enables stable accuracy and constant processing time even with long inputs. Learn more in The Batch: https://hubs.ly/Q04bCKzF0