New Analysis Reveals Transformer Residual Stream Geometry Patterns.

The paper proposes a novel geometric analysis method for understanding the internal workings of transformer models, specifically focusing on the residual streams. This approach measures the "relative displacement" of representations between consecutive layers and employs orthogonal Procrustes analysis to decompose each layer transition into a rigid rotation and a non-rigid residual component. Applying this framework to six instruction-tuned models across tasks like code generation and cross-lingual translation, the researchers found consistent depth regularities. Relative displacement varies significantly with layer depth, often being larger in early and late layers and quieter in the middle. Rotation magnitude remains relatively constant, while the Procrustes residual and angle concentration show depth-modulated patterns. These findings provide descriptive geometric insights into how transformers process information, highlighting model-dependent yet condition-stable depth curves.