J-space analysis quantifies layer influence on output token probabilities
July 15, 2026
The J-lens pipeline measures how nudges to a model's residual stream at layer l affect the final output layer via a response matrix J. This method produces a 4,096-dimensional dictionary of steering vectors, allowing for layer influence comparison using Centered Kernel Alignment (CKA).
HOW THIS AFFECTS YOU
●
researcherYou can use this framework to quantitatively compare how different model architectures distribute information across layers.