There are papers that show training AI on "evil" data results in general misalignment, so it is nice to know the opposite is true and that beneficial RL data in one field leads to more aligned models across a range of tasks.
Beneficial RL data improves AI alignment across tasks
By
–
