Signals 4 · free daily AI digest

Disentangling Representation Evolution in Transformers through Directional Decomposition

Paper recorded by Signals 4 on 2026-09-14 in cs.LG. Abstract reproduced from arXiv; link to the original below.

Published 2026-09-14 on arXiv · recorded by Signals 4 on 2026-09-15

Category: cs.LG · 机器学习 · first seen 2026-09-15

Abstract

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 s

Read on arXiv →

#46 most recent of 215 cs.LG papers we have recorded · ↑ newer: A Chosen Future Can Still Be Rewritten: Causal Writability in Video Mo · ↓ older: Mind2Dialogue: Training Human-Aware Language Models by Simulating User
Cite this page: Disentangling Representation Evolution in Transformers through Directional Decomposition: the #46 most recent of 215 cs.LG papers we have recorded (as of 2026-09-14). Source: Signals 4 (Signals API) — https://data.jiangzhang.ca/signals4/t/papers/disentangling-representation-evolution-in-transformers-through-directional-decom.html
Free to quote with attribution to “Signals 4 (Signals API)”. Machine-readable: papers.json
Related: More cs.LG papers · arXiv signals · All papers · Today in AI
Get 4 AI signals a day by email — free.
Subscribe free → See all plans →
Get 4 AI signals a day by email — free
All models · All repos · By company · Daily editions