Reasoning models often determine refusal of harmful requests early in the chain-of-thought, with the opening sentence in distilled models fully controlling the outcome and patterns transferring across models from the same teacher.
This method avoids the need to create and distribute the content, potentially reducing legal exposure
1 Pith paper cite this work. Polarity classification is still indexing.
1
Pith paper citing it
fields
cs.CL 1years
2025 1verdicts
UNVERDICTED 1representative citing papers
citing papers explorer
-
Where Do Reasoning Models Refuse?
Reasoning models often determine refusal of harmful requests early in the chain-of-thought, with the opening sentence in distilled models fully controlling the outcome and patterns transferring across models from the same teacher.