Low-rank extrapolation of an aligned model's weight update (LoX) reduces how much later fine-tuning erodes safety refusal behavior.
To visualize these points, we project them onto the 17 Published as a conference paper at COLM 2025 same plane and compute their coordinates in the basis {d1, d2}
1 Pith paper cite this work. Polarity classification is still indexing.
1
Pith paper citing it
fields
cs.LG 1years
2025 1verdicts
CONDITIONAL 1representative citing papers
citing papers explorer
-
LoX: Low-Rank Extrapolation Robustifies LLM Safety Against Fine-tuning
Low-rank extrapolation of an aligned model's weight update (LoX) reduces how much later fine-tuning erodes safety refusal behavior.