Researchers have developed a new framework called Co-RL that enables unsupervised reasoning in reinforcement learning. Unlike traditional self-rewarding RL methods, Co-RL uses multiple decoupled models that share no parameters and are optimized through RL using rewards derived from their peers. This approach increases diversity and reduces correlated errors that can lead to homogenized responses and training collapse. The study shows that Co-RL outperforms base models and prior label-free approaches in both text-only and multimodal domains, without requiring ground-truth labels. --- Why it matters: This matters because it addresses a significant challenge in AI research: the need for large amounts of labeled data to train models. Co-RL's ability to achieve state-of-the-art performance without relying on human annotations could have far-reaching implications for areas like natural language processing and computer vision, where data labeling is often time-consuming and expensive. Source: https://arxiv.org/abs/2608.17253