A new paper on arXiv, titled "Representations in Motion: Layer-wise Rotation and Trajectory Divergence in LLMs," takes up the idea that activation dynamics—how representations change during inference—might be more revealing about language model behavior than static readouts from individual layers.
The authors frame their investigation around this premise, suggesting that the movement of representations through the network could offer a richer signal than examining any single layer in isolation. The abstract indicates the work centers on layer-wise rotation and trajectory divergence, though the full findings are not detailed in the available text.
If the premise holds, this line of research could shift how researchers analyze LLMs, moving from snapshot-style inspection toward tracking the evolution of internal states over the course of a forward pass.