SDeflection, a preference-trained defense that makes LLMs deflect harmful requests into benign answers, cuts ASR of a logit-manipulation attack from 89-95% to 8-35% on Llama-2, Llama-3, and Mistral.
Training language models to follow instructions with human feedback
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.CR 1years
2025 1verdicts
CONDITIONAL 1roles
background 1polarities
background 1representative citing papers
citing papers explorer
-
Strategic Deflection: Defending LLMs from Logit Manipulation
SDeflection, a preference-trained defense that makes LLMs deflect harmful requests into benign answers, cuts ASR of a logit-manipulation attack from 89-95% to 8-35% on Llama-2, Llama-3, and Mistral.