A linear projection learned against an aligned LLM's outputs reproduces a large share of alignment effects, especially safety and style, but not the reasoning behind math and factual answers.
Title resolution pending
1 Pith paper cite this work. Polarity classification is still indexing.
1
Pith paper citing it
fields
cs.CL 1years
2025 1verdicts
CONDITIONAL 1representative citing papers
citing papers explorer
-
Extracting and Understanding the Superficial Knowledge in Alignment
A linear projection learned against an aligned LLM's outputs reproduces a large share of alignment effects, especially safety and style, but not the reasoning behind math and factual answers.