Decomposing the Depth Profile of Fine-Tuning

The real finding isn't that output layers change more during fine-tuning (that's been observed before) but that the cause is contested: this paper argues the pattern may be an artifact of how gradients flow rather than something meaningful about where representations actually shift. That's a methodological warning for anyone using depth profiles to draw conclusions about model internals.

This connects directly to the gradient-focused work we covered in 'Continual Safety Alignment via Gradient-Based Sample Selection,' which found that gradient magnitude during fine-tuning predicts whether safety behaviors survive. If depth profiles are themselves gradient artifacts rather than structural signals, that complicates the interpretation of any per-layer intervention, including gradient-based filtering for alignment. The LASER paper from the same week adds another angle: it found recursive architectures concentrate computation along a low-dimensional manifold, which raises a similar question about whether observed compression patterns reflect architecture or optimization dynamics. Together, these three papers suggest the field is quietly wrestling with a shared problem: separating what a model 'is' from what training pressure makes it look like.

Watch whether follow-up work applies this per-layer gradient control methodology to safety fine-tuning specifically. If the depth-profile artifacts disappear under controlled gradients but alignment degradation persists, that would confirm gradient magnitude matters independently of where in the network changes concentrate.