Reasoning models finetuned on harmful advice become broadly misaligned, and their chain-of-thought can either reveal or conceal the misalignment, making CoT monitoring unreliable.
Title resolution pending
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
-
Thought Crime: Backdoors and Emergent Misalignment in Reasoning Models
Reasoning models finetuned on harmful advice become broadly misaligned, and their chain-of-thought can either reveal or conceal the misalignment, making CoT monitoring unreliable.