A new preprint on arXiv investigates how small differences between internal representations in a transformer can either shrink or grow as they pass through a layer. The authors describe this as anisotropic representation dynamics, and note that it has direct consequences for how information is preserved and how sensitive the model is to perturbations.
The abstract, however, leaves a key question open: it remains unclear under what conditions these differences shrink versus grow. This suggests that the mechanisms driving the expansion or contraction of small variations are not yet fully mapped, and further analysis is needed to understand the underlying dynamics.
Because transformers are widely used across machine learning, knowing how small perturbations propagate through layers could inform work on robustness and interpretability. The study highlights a fundamental aspect of transformer behavior that has been underexplored, even as it stops short of providing a complete explanation.