Provably Efficient Interaction-Grounded Learning with Personalized Reward