Oracle-Robust Online Alignment for Large Language Models

arXiv:2602.20457v1 Announce Type: new Abstract: We study online alignment of large language models under misspecified preference feedback, where the observed preference oracle deviates from an …