GEAR: Fixing Repetitive Copying in Long-Context LLM Reasoning via Grounding Rewards

Forum topic · 小凯 · 2026-07-23

Summary

This paper (arXiv:2507.17091) identifies a critical failure mode in long-context reasoning with large language models: repetitive copying, where models copy large amounts of input text into their reasoning traces instead of solving the problem. The authors show this behavior is widespread across frontier long-context LLMs and worsens as context length grows. Decomposing prompts into task-relevant key evidence and irrelevant distractor context reveals the root cause: insufficient grounding, as models copy indiscriminately and those failing to focus on key evidence answer incorrectly far more often. The proposed solution, GEAR (Grounding Evidence-Aware Reward), is a reward-shaping method that grants grounding rewards for overlap with key evidence and distraction penalties for overlap with irrelevant context, complementing standard accuracy-based reinforcement learning. The authors also build an automated pipeline to construct evidence-annotated training data from arbitrary documents. Across multiple model sizes and benchmarks, GEAR improves average performance by up to +4.6 points over accuracy-only RL, with larger gains at longer contexts, while reducing repetitive copying and shortening thinking length.

Paper Overview

Field: NLP
Authors: Lizhe Fang, Weizhou Shen, Tianyi Tang
arXiv: 2507.17091

Key Points

LLMs that generate step-by-step reasoning traces perform strongly on complex tasks, and extending them to long-context settings is an important frontier. However, a critical failure mode emerges: repetitive copying, where models extensively copy input text into their reasoning traces rather than productively solving the problem.
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, the authors trace the root cause to insufficient grounding: models copy from the prompt indiscriminately, and those that fail to focus on key evidence are far more likely to answer incorrectly.
GEAR (Grounding Evidence-Aware Reward) is proposed: a reward-shaping method that gives grounding rewards for overlap with key evidence and distraction penalties for overlap with irrelevant context, sharpening the accuracy signal used in RL.
To apply GEAR on natural language data, the team developed an automated pipeline that constructs evidence-annotated training data from arbitrary documents.
Experiments across multiple model sizes and benchmarks show GEAR improves over standard accuracy-based RL by up to +4.6 points on average, with larger gains at longer contexts, while reducing repetitive copying and thinking length.
Conclusion: as long-context evaluation shifts from simple retrieval to complex reasoning, accurate grounding on relevant evidence remains an essential capability with substantial room for improvement.

Abstract (Original)

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: *repetitive copying*, where models extensively copy text from the input into their reasoning traces rather than productively solving the problem...

--- *Auto-collected on 2026-07-23*