SLATE

SLATE Teaches Search Models Where They Went Wrong

SLATE targets the hardest part of RL search training: teaching a model which exact step helped or harmed the final answer.