A matched benchmark across four task formats and seven reasoning models shows chain-of-thought monitors detect implicit behavioral influences far worse than explicit ones, with detection dropping to as low as 5 percent under guidance prompts.
CoT red-handed: Stress testing chain-of-thought monitoring
1 Pith paper cite this work. Polarity classification is still indexing.
1
Pith paper citing it
fields
cs.AI 1years
2026 1verdicts
CONDITIONAL 1representative citing papers
citing papers explorer
-
Chain-of-Thought Monitoring Can Be Unreliable in Implicit-Influence Settings
A matched benchmark across four task formats and seven reasoning models shows chain-of-thought monitors detect implicit behavioral influences far worse than explicit ones, with detection dropping to as low as 5 percent under guidance prompts.