Copy Less, Ground More: Overcoming Repetitive Copying in Long-Context Reasoning via Evidence-Aware Reinforcement Learning

Large language models that generate step-by-step reasoning traces are prone to a critical failure mode known as repetitive copying, where they excessively copy input text into their reasoning traces instead of solving problems productively. This issue arises when these models are applied to long-context settings, a key area of development for complex tasks. To address this, researchers have proposed evidence-aware reinforcement learning, which enables models to focus on relevant information and avoid redundant copying. By incorporating this approach, models can improve their ability to reason effectively in long-context settings. The development of more advanced language models using reinforcement learning has significant implications for both their capabilities and potential risks, particularly in terms of security¹. As a result, practitioners must consider the potential consequences of these developments on the security landscape, making this an important area of study for those concerned with the safe and effective deployment of large language models.