Suppressing the double-newline token immediately after <think> substantially increases substantive answers to sensitive prompts in DeepSeek-R1 distillations without training.
Title resolution pending
1 Pith paper cite this work. Polarity classification is still indexing.
1
Pith paper citing it
fields
cs.CL 1years
2025 1verdicts
CONDITIONAL 1representative citing papers
citing papers explorer
-
Derailing Non-Answers via Logit Suppression at Output Subspace Boundaries in RLHF-Aligned Language Models
Suppressing the double-newline token immediately after <think> substantially increases substantive answers to sensitive prompts in DeepSeek-R1 distillations without training.