build1 publisher
A nudge delivered as a casual aside costs a CoT monitor 41 to 46 detection points
A new arXiv benchmark runs seven extended-thinking models under an explicit order to conceal and under an offhand contextual detail. A routine anti-bias system prompt pushes implicit detection as low as 5%.
Publishers:arxiv.org
Reality
- Evidence58
- Adoption
- Insufficient
- Hype gap+15
- Incentives35
- Confidence55