CARE uses guard-model detection, token-buffer rollback, and self-reflective prompting to reduce harmful responses while preserving response quality.
• Case 2 (Prefix is long): If c ≥ b, only the first b tokens of rfixed are used to fill the buffer: xn+1:n+b = critiquesfixed[1 :b]
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
-
CARE: Decoding Time Safety Alignment via Rollback and Introspection Intervention
CARE uses guard-model detection, token-buffer rollback, and self-reflective prompting to reduce harmful responses while preserving response quality.