AI News

CleanRM: calibrating sample-wise noise in reward models using a few clean preferences and optimal transport

ABC2 min read
Research