Improving self-training under distribution shifts via anchored confidence with theoretical guarantees | Read Paper on Bytez