Provably Efficient Interactive-Grounded Learning with Personalized Reward