A missingness-aware sampling method that selects safety-critical fine-tuning examples using hidden-representation gaps reduces attack success rates after task-specific fine-tuning.
Fine-tuning aligned language models compromises safety, even when users do not intend to!
1 Pith paper cite this work. Polarity classification is still indexing.
1
Pith paper citing it
citation-role summary
background 1
citation-polarity summary
fields
cs.CL 1years
2026 1verdicts
CONDITIONAL 1roles
background 1polarities
support 1representative citing papers
citing papers explorer
-
DataRx: Missingness-Aware Sampling for Safer Large Language Model Task-Specific Fine-Tuning
A missingness-aware sampling method that selects safety-critical fine-tuning examples using hidden-representation gaps reduces attack success rates after task-specific fine-tuning.