GEAR: Grounding Rewards Fix 'Repetitive Copying,' a Long-Context Reasoning Failure Mode That Worsens With Length

arXiv 2607.19345·medium signal

The paper identifies repetitive copying — models pasting large spans of input into their reasoning traces instead of solving the problem — as a distinct failure mode across frontier long-context LLMs that gets worse as context grows. Root cause is insufficient grounding in task-relevant evidence: models that fail to focus on key evidence copy indiscriminately and score significantly worse. GEAR shapes rewards with an accuracy signal plus a grounding reward for key-evidence overlap and a penalty for distractor overlap, using an automated pipeline to build evidence-annotated training data. Gains reach +4.6 average points over accuracy-only RL, larger at longer contexts, with shorter thinking traces.