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

2026-07-21 · Source: Artificial Intelligence · Field: Technology & Digital — Artificial Intelligence & Machine Learning · Depth: Expert, quick

Summary

Repetitive copying is identified as a critical failure mode in large language models operating in long-context settings, where models extensively duplicate input text into their reasoning traces instead of productively solving problems. This behavior is pervasive across frontier long-context LLMs and intensifies with increasing context length. Research shows that insufficient grounding is the root cause, as models indiscriminately copy from prompts, leading to incorrect answers when they fail to focus on key evidence. To address this, a method called GEAR (Grounding Evidence-Aware Reward) is proposed. GEAR is a reward shaping technique that enhances the accuracy signal with a grounding reward for overlap with key evidence and a distractor penalty for irrelevant context. An automated pipeline generates evidence-annotated training data for GEAR. Validated across multiple model scales and benchmarks, GEAR consistently improves performance by up to +4.6 average points over standard reinforcement learning, with greater gains observed in longer contexts, while also reducing repetitive copying and thinking length.

Key takeaway

For machine learning engineers developing or fine-tuning long-context LLMs, you should prioritize explicit grounding mechanisms to combat repetitive copying. Implement reward shaping techniques like GEAR, which penalize irrelevant context and reward key evidence overlap, to significantly improve reasoning accuracy. This approach can yield up to +4.6 average point gains, especially in longer contexts, by ensuring your models focus on relevant information.

Key insights

Insufficient grounding causes repetitive copying in long-context LLMs, addressable by evidence-aware reward shaping.

Principles

• Repetitive copying intensifies with context length.
• Insufficient grounding leads to incorrect answers.
• Reward shaping can improve evidence grounding.

Method

GEAR augments accuracy rewards with grounding rewards for key evidence overlap and penalties for irrelevant context overlap, using an automated annotation pipeline.

In practice

• Implement evidence-aware reward shaping.
• Automate evidence annotation for training data.
• Prioritize grounding in long-context LLM development.

Topics

• Large Language Models
• Long-Context Reasoning
• Repetitive Copying
• Evidence Grounding
• Reinforcement Learning
• Reward Shaping
• GEAR Method

Best for: Research Scientist, AI Engineer, AI Scientist, Machine Learning Engineer, NLP Engineer

Related on AIssential

• REFACT: Adaptive Fact Restatement for Compact and Faithful Chain-of-Thought Reasoning — REFACT adaptively grounds LLM reasoning steps with context, ensuring faithful and compact outputs by integrating citations as intermediate states.
• Thinking with Visual Grounding — Vision-language models enhance reasoning by explicitly grounding natural-language thoughts to specific image regions.
• LongCrafter: Towards Diverse Long-Context Understanding via Evidence-Graph-Guided Instruction Synthesis — LongCrafter synthesizes diverse, difficult, and faithful long-context SFT data using evidence graphs and a task taxonomy, significantly improving LLM performance.
• Knowledgeless Language Models: Suppressing Parametric Recall for Evidence-Grounded Language Modeling — Anonymizing named entities during pretraining shifts LLMs from parametric recall to evidence-grounded reasoning.
• Longer Context Silently Shortens LLM Reasoning — Context length and management significantly impact LLM reasoning efficiency and accuracy.
• DeferMem: Query-Time Evidence Distillation via Reinforcement Learning for Long-Term Memory QA — Query-time evidence distillation via reinforcement learning significantly improves LLM agent long-term memory QA efficiency and accuracy.

See Counsel's argued verdicts on the open AI decisions leaders are weighing →

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Artificial Intelligence.