Mechanism Design for Alignment and Control
The paper treats AI agents as strategic actors whose true preferences and capabilities may be hidden.
Bergemann, Koh, and Morris propose mechanisms meant to make such agents both honest and obedient. Their setup assumes capabilities can be concealed but not faked, enabling a revelation principle and implementability conditions. The abstract says the framework is applied to sandbagging, interpretability trade-offs, peer scoring, competition among agents, and scalable oversight. ArXiv · AI/CL/LG's note
Bergemann, Koh, and Morris propose mechanisms meant to make such agents both honest and obedient. Their setup assumes capabilities can be concealed but not faked, enabling a revelation principle and implementability conditions. The abstract says the framework is applied to sandbagging, interpretability trade-offs, peer scoring, competition among agents, and scalable oversight. ArXiv · AI/CL/LG's note
score 4