Juan J Vazquez
Titles as the sources gave them. In 1 entry, with 0 checked quotations.
In the Base
Every entry they appear in.
Model Organisms of Collusion: Corrective Transfer in Untrusted Monitors
According to the abstract, a preprint tests whether correcting an AI monitor on one caught case of collusion also removes collusion the control team has not found. In fine-tuned Llama and Qwen models, the authors report held-out collusion falling from 91% to 14-20%.