The paper argues and partially tests that linear concept representations in LLMs originate in the input space and propagate through the leading singular directions of activation difference matrices.
Microsoft coco: Common objects in context
1 Pith paper cite this work. Polarity classification is still indexing.
1
Pith paper citing it
fields
cs.CV 1years
2025 1verdicts
REJECT 1representative citing papers
citing papers explorer
-
Spectral Principal Paths: A Spectral Perspective on Linear Representation Formation in LLMs
The paper argues and partially tests that linear concept representations in LLMs originate in the input space and propagate through the leading singular directions of activation difference matrices.