一作ICML 2026Spotlight· Deep Learning / Large Language Models
Shared Semantics, Divergent Mechanisms: Unsupervised Feature Discovery by Aligning Semantics and Mechanisms
Hyunjin Cho · Youngji Roh · Jaehyung Kim
首次获取该候选人的完整档案及 StartLux 评估共 20 积分；每位用户单独解锁，之后查看免费。主动刷新 20 积分。
实时查找电子邮箱，费用不超过20积分；未找到邮箱不扣费。
Hyunjin Cho · Youngji Roh · Jaehyung Kim
Youngji Roh · Hyunjin Cho · Jaehyung Kim
Large Language Models (LLMs) exhibit highly anisotropic internal representations, often characterized by massive activations, a phenomenon where a small subset of feature dimensions possesses magnitudes significantly larger than the rest. While prior works view these extreme dimensions primarily...