Disentangling Representation Evolution in Transformers through Directional Decomposition
Researchers decompose transformer updates into parallel and perpendicular components, linking representation geometry to editing robustness, compression diagnosis, and training interventions.
The paper studies transformer representation evolution as functional geometry, decomposing learned updates into parallel and perpendicular components across attention/MLP and value-aggregation spaces. Targeted edits reveal a space-dependent asymmetry: exclude-self value-space parallel manipulation is markedly more robust than residual-space and perpendicular counterparts. Full-aggregate parallel suppression during from-scratch pretraining lowers validation-loss trajectories and improves downstream averages, with the value-space variant strongest. Code is released on GitHub.
- Decomposes learned updates into parallel and perpendicular components
- Value-space parallel edits prove most robust for targeted model interventions
- Perpendicular error separates compression methods more clearly than parallel error
- Parallel suppression during pretraining lowers validation loss and improves downstream performance
From the original article
1 min read at huggingface.co
Transformer representations evolve through learned additive transformations that either preserve their current direction or redirect it. We study this evolution as a functional geometry, decomposing learned updates into parallel and perpendicular components. Across pretrained models, we find substantial parallel components beyond the residual identity path.
By Shwai He, Haichao Zhang, Shen Yan · huggingface.co
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.