TroL: Traversal of Layers for Large Language and Vision Models | Papers ...