EvoRubrics: Dynamic Rubrics as Rewards via Adversarial Co-Evolution for LLM Reinforcement Learning

EvoRubrics introduces a co-evolutionary reinforcement learning framework that dynamically generates rubrics for evaluating a Policy LLM, enhancing training effectiveness in open-ended tasks. By allowing the Rubric Generator to adapt its criteria in real time based on the evolving policy, EvoRubrics mitigates issues of reward saturation and improves discriminative power, outperforming both static and existing dynamic rubric methods across various benchmarks. This approach demonstrates that self-supervised co-evolution can yield rich learning signals, offering a novel avenue for practitioners to optimize LLM training without reliance on external supervision.