rlhffactbullishFusing hidden-state representations from heterogeneous reward models provides a more expressive basis for preference evaluation than scalar-score methodsComputation and Language28 Jul 2026http://arxiv.org/abs/2607.20862v1