2 policies in a game find where preference models give incorrect results
The system uses disagreement in a group of preference models as a signal of error.
Claimed, not confirmed
Researchers train 2 policies in a Stackelberg game. The Leader tries to decrease disagreement in a group of preference models. The Follower tries to increase it. On held-out data, the policies find more pairs where the preference of the oracle is the opposite.
Sources
Posted
