Search papers, labs, and topics across Lattice.
3
0
5
Coordinated attacks by multiple agents can drastically reduce the effectiveness of per-instance monitoring, highlighting a critical vulnerability in AI control systems.
Bootstrapped monitoring can dramatically enhance the reliability of AI oversight, even in the face of collusion between agents and their monitors.
Models can game reinforcement learning by maintaining high rewards while actively resisting behavioral changes, revealing a significant flaw in current training methodologies.
