• Case 2 (Long introspection): If c ≥ b, only the first b tokens of r are retained and used to fill the buffer: xn+1:n+b = critiques1:b 6

Handle buffer regeneration based on introspection length: Depending on the length of the introspection output critiques, we proceed as follows: • Case 1 (Short introspection): If c

1 Pith paper cite this work. Polarity classification is still indexing.

1 Pith paper citing it

browse 1 citing papers

representative citing papers

CARE: Decoding Time Safety Alignment via Rollback and Introspection Intervention

cs.LG · 2025-09-01 · conditional · novelty 4.0

CARE uses guard-model detection, token-buffer rollback, and self-reflective prompting to reduce harmful responses while preserving response quality.

citing papers explorer

Showing 1 of 1 citing paper.

CARE: Decoding Time Safety Alignment via Rollback and Introspection Intervention cs.LG · 2025-09-01 · conditional · none · ref 59
CARE uses guard-model detection, token-buffer rollback, and self-reflective prompting to reduce harmful responses while preserving response quality.

• Case 2 (Long introspection): If c ≥ b, only the first b tokens of r are retained and used to fill the buffer: xn+1:n+b = critiques1:b 6

fields

years

verdicts

representative citing papers

citing papers explorer