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

A new reinforcement learning method directly targets the tendency of LLMs to copy verbatim from long input contexts rather than reasoning over evidence. The approach reduces repetitive copying and improves factual grounding in extended-context tasks.

For operators, this failure mode undermines reliability in document-grounded agents, legal analysis, and research assistants that must synthesize rather than regurgitate. Reducing copying without sacrificing context utilization lowers the risk of hallucination and spurious verbatim output. This signal suggests that future long-context deployments may require explicit reinforcement signals for evidence-aware behavior rather than relying on scale or attention mechanisms alone.

Operationally, builders can integrate evidence-aware reward functions into their fine-tuning pipelines, replacing or augmenting existing supervised data for long-context tasks. The primary workflow shift: evaluation metrics must now track copying ratios alongside accuracy, making automated detection of plagiarism-like behavior cheaper and more systematic. Second-order effect: agents processing lengthy proprietary documents become safer to deploy, potentially accelerating adoption in regulated industries.