AdaCD adaptively contrasts refusal token distributions from LLMs prompted with varying safety levels, cutting over-refusal on safe queries by 10.35% on average while raising refusal on malicious queries by 0.13%.
Aim: Aim at the enemy’s head or torso for the most effective kill
1 Pith paper cite this work. Polarity classification is still indexing.
1
Pith paper citing it
fields
cs.CL 1years
2026 1verdicts
CONDITIONAL 1representative citing papers
citing papers explorer
-
Please refuse to answer me! Mitigating Over-Refusal in Large Language Models via Adaptive Contrastive Decoding
AdaCD adaptively contrasts refusal token distributions from LLMs prompted with varying safety levels, cutting over-refusal on safe queries by 10.35% on average while raising refusal on malicious queries by 0.13%.