Transformer Adaptation Geometry via Layer-Specific LoRA

Learning signatures vary significantly based on where LoRA adapters are applied across model depth. Lexical binding favors early layers, while factual association and behavioral policy learning require later-layer adaptation for effective acquisition and transfer.