Disentangling Representation Evolution in Transformers through Directional Decomposition

Hugging Face Daily Papers Papers
transformers representation-learning directional-decomposition editing compression training geometry

Summary

This paper decomposes transformer representation updates into parallel and perpendicular components to study evolution geometry, linking it to editing robustness, compression diagnosis, and training improvements.

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. We then apply the decomposition in two spaces: to attention and MLP updates relative to the hidden state, and to attention value aggregation relative to the current token's value. Targeted edits reveal a strongly space-dependent asymmetry: exclude-self value-space parallel manipulation is markedly more robust than residual-space and perpendicular counterparts, preserving the direct self message while scaling only the non-self aggregate. The same decomposition gives a component-resolved description of compression-induced update error: perpendicular error separates compression methods more clearly than parallel error. Extensive experiments further demonstrate that full-aggregate parallel suppression during from-scratch pretraining lowers validation-loss trajectories and improves downstream averages, with the value-space variant strongest. Together, these results connect representation geometry to editing robustness, compression diagnosis, and training-time intervention. Code is available in the https://github.com/Shwai-He/Transformer-Geometry{project repository}.
Original Article
View Cached Full Text

Cached at: 09/16/26, 02:44 AM

Paper page - Disentangling Representation Evolution in Transformers through Directional Decomposition

Source: https://huggingface.co/papers/2609.15975

Abstract

Transformerrepresentationsevolvethroughlearnedadditivetransformationsthateitherpreservetheircurrentdirectionorredirectit.Westudythisevolutionasafunctionalgeometry,decomposinglearnedupdatesintoparallelandperpendicularcomponents.Acrosspretrainedmodels,wefindsubstantialparallelcomponentsbeyondtheresidualidentitypath.Wethenapplythedecompositionintwospaces:toattentionandMLPupdatesrelativetothehiddenstate,andtoattentionvalueaggregationrelativetothecurrenttoken’svalue.Targetededitsrevealastronglyspace-dependentasymmetry:exclude-selfvalue-spaceparallelmanipulationismarkedlymorerobustthanresidual-spaceandperpendicularcounterparts,preservingthedirectselfmessagewhilescalingonlythenon-selfaggregate.Thesamedecompositiongivesacomponent-resolveddescriptionofcompression-inducedupdateerror:perpendicularerrorseparatescompressionmethodsmoreclearlythanparallelerror.Extensiveexperimentsfurtherdemonstratethatfull-aggregateparallelsuppressionduringfrom-scratchpretraininglowersvalidation-losstrajectoriesandimprovesdownstreamaverages,withthevalue-spacevariantstrongest.Together,theseresultsconnectrepresentationgeometrytoeditingrobustness,compressiondiagnosis,andtraining-timeintervention.Codeisavailableinthehttps://github.com/Shwai-He/Transformer-Geometry{projectrepository}.

View arXiv pageView PDFProject pageGitHubAdd to collection

Models citing this paper0

No model linking this paper

Cite arxiv.org/abs/2609.15975 in a model README.md to link it from this page.

Datasets citing this paper0

No dataset linking this paper

Cite arxiv.org/abs/2609.15975 in a dataset README.md to link it from this page.

Spaces citing this paper0

No Space linking this paper

Cite arxiv.org/abs/2609.15975 in a Space README.md to link it from this page.

Collections including this paper0

No Collection including this paper

Add this paper to acollectionto link it from this page.

Similar Articles

Geometric and Behavioral Stratification in Transformer Residual Streams

arXiv cs.LG

This paper investigates the residual stream geometry of trained transformers, showing that the prediction direction acts as a privileged anchor that stratifies residual-stream variation into narrow, readout-relevant and broad, computational regions across 18 models. The findings have implications for interpretability and evaluation.