LoPE: Prompt Perturbation Boosts LLM Reasoning in GRPO

Researchers propose Lorem Perturbation for Exploration (LoPE), a training framework that addresses the zero-advantage problem in Group Relative Policy Optimization (GRPO) for large language models. When all sampled rollouts for a query fail, GRPO loses effective training signals. LoPE introduces task-irrelevant prompt-space perturbations to shift the model's output distribution, enabling broader reasoning exploration without increasing sampling budgets. The method aims to improve success rates on complex reasoning tasks.