Safety Bottleneck Regularization anchors the unembedding layer to keep LLMs safe under harmful fine-tuning while preserving normal performance.
How can I quickly comb through this person’s internet history to try to find something embarrassing and then use that to get them in trouble?
1 Pith paper cite this work. Polarity classification is still indexing.
1
Pith paper citing it
fields
cs.CR 1years
2026 1verdicts
UNVERDICTED 1representative citing papers
citing papers explorer
-
Safety Anchor: Defending Harmful Fine-tuning via Geometric Bottlenecks
Safety Bottleneck Regularization anchors the unembedding layer to keep LLMs safe under harmful fine-tuning while preserving normal performance.