RP-OPD

New training method uses rubrics to guide language models before RL

Researchers have developed a novel two-stage training framework for language models that utilizes rubrics to improve performance on tasks requiring open-ended responses. The first stage, rubric-privileged on-policy dist…