RiVER trains LLMs with reinforcement learning and no ground-truth answers

Category: Paper
Status: developing
Sources: arXiv 2606.27369
Summary: A preprint introduces RiVER (Ranking-induced VERifiable framework), which applies reinforcement learning to score-based tasks without ground-truth solutions by using deterministic execution feedback as continuous-valued reward. The authors identify and address two failure modes of group-relative RL on continuous rewards: scale dominance, where uncalibrated score magnitudes across instances distort updates, and a calibration issue in reward normalization.
Why it matters: Removing the ground-truth requirement extends RLVR-style post-training to optimization and coding tasks where only an executable scorer exists, not a labeled answer.
Follow-up: Watch for code release and independent reproduction beyond the paper's own evaluation.