Canonical-basis realignment for Transformer LLMs: every hidden axis becomes independently measurable and controllable
Summary
The article introduces CBLL, a lossless canonical basis for Transformer LLMs that makes each hidden axis independently measurable and controllable. It presents cross-layer alignment, causal axis ablations, and architecture-general findings across RMSNorm and LayerNorm families, with open-source data and scripts to reproduce results.