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

Background and Context

As large language models demonstrate increasingly sophisticated step-by-step reasoning capabilities in complex tasks, extending these abilities to long-context scenarios has emerged as a critical frontier in artificial intelligence research. However, recent investigations have uncovered a significant, previously overlooked failure mode known as "repetitive copying." In practical applications, many state-of-the-art long-context models exhibit a tendency to mechanically copy large portions of input text into their reasoning traces rather than engaging in genuine logical deduction. This inefficient behavior not only squanders computational resources but also severely compromises the model's ability to answer complex questions accurately. The study highlights that this copying behavior intensifies significantly as the context length increases, suggesting a fundamental limitation in how current architectures handle extended information streams.

Deep analysis reveals that the root cause of this phenomenon lies in the model's insufficient "grounding" capability. Specifically, models struggle to effectively distinguish between key evidence and irrelevant distractors within the prompt. Models that fail to focus on critical evidence are not only more prone to repetitive copying but also exhibit a substantially higher probability of generating incorrect answers compared to their more focused counterparts. This diagnosis provides a new entry point for improving long-context reasoning: rather than simply piling up context length, systems must reinforce the model's ability to locate and utilize core information. The inability to filter noise from signal in long documents is identified as the primary bottleneck preventing reliable reasoning in extended contexts.

Deep Analysis

To address these structural deficiencies, the research team proposes GEAR (Grounding Evidence-Aware Reward), a novel reward shaping method designed to optimize model reasoning behavior through reinforcement learning. GEAR expands the traditional reward signal, which relies solely on answer accuracy, into a composite framework comprising two additional dimensions. The first is a grounding reward, which encourages the model to overlap its reasoning steps with key evidence relevant to the task. The second is a distractor penalty, which actively suppresses the model's tendency to copy irrelevant context. This dual mechanism ensures that the model is incentivized to engage with substantive information while being penalized for superficial text replication.

Implementing GEAR in natural language data required the development of an automated data construction pipeline. This system automatically extracts and annotates key evidence and distractor contexts from arbitrary documents, transforming them into high-quality training data. During the reinforcement learning phase, models receive multi-dimensional guidance, balancing accuracy, grounding rewards, and distractor penalties. This approach forces the model to develop sensitivity to key information, enabling it to consciously filter and cite relevant evidence during long reasoning chains. The pipeline effectively bridges the gap between raw document data and structured reinforcement learning targets, allowing for scalable training without manual annotation bottlenecks.

Comprehensive validation across multiple model scales and benchmarks demonstrates GEAR's consistent and significant performance improvements. On average, GEAR outperforms standard reinforcement learning methods based solely on accuracy by 4.6 points. The gains are particularly pronounced in long-context scenarios, underscoring the method's unique advantage in handling complex, lengthy texts. Ablation studies further isolate the contributions of each component: the grounding reward significantly enhances the model's utilization of key information, while the distractor penalty effectively reduces redundancy in reasoning traces. Notably, models trained with GEAR generate significantly shorter reasoning chains, indicating that they have learned more efficient, focused strategies rather than masking logical flaws with excessive steps.

Industry Impact

The implications of this research extend across the open-source community, industrial applications, and future academic inquiry. For the open-source ecosystem, GEAR provides an automated evidence annotation pipeline that serves as a reproducible tool for other researchers. This lowers the barrier to entry for long-context reinforcement learning, enabling broader experimentation and innovation. By standardizing the process of converting arbitrary documents into evidence-annotated training data, the community can more rapidly iterate on reasoning models without being constrained by the high costs of manual data preparation.

In industrial settings, the reduction of repetitive copying and the shortening of reasoning lengths translate directly into lower inference costs and faster response times. For enterprise applications that must process massive volumes of documents, these efficiency gains are critical. The ability to generate accurate answers with fewer computational steps allows for higher throughput and reduced latency, making long-context reasoning more viable for real-time decision support systems. Furthermore, the improved reliability of grounding ensures that automated systems can be trusted with high-stakes tasks where hallucination or irrelevant information extraction could have serious consequences.

As long-context evaluation shifts from simple information retrieval to complex reasoning, the ability to accurately ground responses in relevant evidence becomes an indispensable core skill. This finding redirects the focus of future model optimization away from mere context expansion toward a deeper exploration of information structure understanding. The success of GEAR demonstrates that enhancing the model's structural comprehension is more impactful than simply increasing parameter count or context window size. This shift in focus is likely to influence the development of next-generation architectures, prioritizing precision and efficiency over brute-force context handling.

Outlook

The success of GEAR lays a solid foundation for building more efficient and reliable long-context reasoning systems. By proving that targeted reward shaping can correct fundamental reasoning flaws, the study opens new avenues for optimizing model behavior in complex environments. Future research is likely to build upon this framework, exploring more sophisticated methods for evidence extraction and dynamic reward adjustment. The automated pipeline developed in this study serves as a template for creating specialized training data for various domains, potentially accelerating the adoption of long-context reasoning in legal, medical, and technical fields.

Moreover, the emphasis on reducing redundant reasoning steps suggests a path toward more interpretable and auditable AI systems. Shorter, more focused reasoning traces are easier for human experts to review and validate, addressing one of the major hurdles in deploying AI for critical applications. As models become more proficient at grounding their responses, the gap between machine-generated reasoning and human expert analysis may narrow, fostering greater trust and collaboration between humans and AI systems.

Ultimately, this research marks a pivotal moment in the evolution of large language models. It shifts the paradigm from passive context consumption to active information synthesis. By addressing the pervasive issue of repetitive copying, GEAR not only improves performance metrics but also enhances the fundamental reliability of AI systems. As the field moves forward, the principles of evidence-aware reward shaping are expected to become standard practice, ensuring that future models are not just larger, but smarter and more grounded in the realities of the data they process.