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

TL;DR - The paper identifies repetitive prompt copying as a widespread long-context reasoning failure caused by poor evidence grounding. Its evidence-aware RL method, GEAR, improves benchmark scores by up to 4.6 points while producing shorter, less repetitive reasoning traces.

• Repetitive copying worsens as context length increases and correlates with incorrect answers.
• GEAR rewards overlap with relevant evidence and penalizes copying from irrelevant distractors.
• An automated pipeline creates evidence-annotated training data from arbitrary documents.
• Improvements are consistent across model scales and benchmarks, with larger gains on longer contexts.