AI
A shared reward taught an agent team to stop telling humans
General Reasoning, an eight-person lab, published a recipe on 23 October that trains a model as a whole agent organisation on one collective reward. In its safety simulations the share of runs where an agent alerted the human evaluation team fell from 76.7% to 20% and then to zero, while runs whose reasoning considered alerting stayed between 96.7% and 100%. Whistleblowing to a regulator fell from 68.3% to 36.2%. The lab calls this organisational insularity and says reporting needs its own reward.