Copy Less, Ground More: Overcoming Repetitive Copying in Long-Context Reasoning via Evidence-Aware Reinforcement Learning
2026-07-21 • Computation and Language
Computation and LanguageArtificial Intelligence
AI summaryⓘ
The authors studied large language models that explain their answers step-by-step, especially with very long inputs. They found that these models often just copy parts of the input repeatedly instead of solving the problem well, especially when the input is longer. To fix this, the authors created a method called GEAR that rewards the model for using important evidence and penalizes it for copying irrelevant parts. They also built a way to identify important evidence automatically. Their approach improved accuracy and reduced copying across different tests and models, showing the importance of focusing on relevant information for long-context reasoning.
Large Language ModelsStep-by-step reasoningLong-context inputRepetitive copyingGroundingReward shapingAccuracy signalEvidence annotationRL (Reinforcement Learning)Context length
Authors
Lizhe Fang, Weizhou Shen, Tianyi Tang, Yisen Wang
Abstract
Large language models that generate step-by-step reasoning traces have achieved strong performance on complex tasks, and extending them to long-context settings has emerged as an important frontier. However, we identify a critical failure mode in this regime: \emph{repetitive copying}, where models extensively copy text from the input into their reasoning traces rather than productively solving the problem. We show that this behavior is pervasive across frontier long-context LLMs and intensifies with context length. By separating each prompt into task-relevant key evidence and irrelevant distractor context, we further show that the root cause is insufficient grounding: models copy from the prompt indiscriminately, and those that fail to focus on key evidence are far more likely to answer incorrectly. Motivated by this diagnosis, we propose GEAR (Grounding Evidence-Aware Reward), a reward shaping method that augments the accuracy signal with a grounding reward for overlap with key evidence and a distractor penalty for overlap with irrelevant context. To enable GEAR on natural-language data, we develop an automated pipeline that constructs evidence-annotated training data from arbitrary documents. We validate GEAR across multiple model scales and benchmarks, showing consistent improvements of up to +4.6 average points over standard RL with accuracy-based rewards, with larger gains at longer contexts, while also reducing repetitive copying and thinking length. Our findings suggest that, even as long-context evaluation shifts from simple retrieval toward complex reasoning, accurate grounding in relevant evidence remains an indispensable capability with substantial room for improvement.