JulesRoussel01

In open RLVR, “im­prove­ment” de­pends on the in­stru­ment — a small GRPO testbed sep­a­rat­ing what train­ing op­ti­mizes, mea­sures, and teaches