---
title: "Copy Less, Ground More: Overcoming Repetitive Copying in Long-Context Reasoning via Evidence-Aware Reinforcement Learning"
source: "arXiv AI/CL/LG（论文）"
source_tier: "t2"
url: "https://arxiv.org/abs/2607.19345v1"
permalink: "https://feed.astralor.com/items/itm_d72529776569c2e45570e4e9"
published: "2026-07-21T17:59:21.000Z"
score: 79
channel: "ai"
tags: ["Large Language Models","Reinforcement Learning","Evidence-Aware Reward","Long Context"]
---

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

## AI commentary

> This paper is a must-read for anyone working on long-context LLMs—it exposes a subtle but critical failure mode where models cheat by copying rather than reasoning. GEAR is lightweight and automatable, ideal for teams aiming to reduce inference cost while improving accuracy in production.

## At a glance

This paper identifies a pervasive failure mode in long-context LLMs: repetitive copying of input text into reasoning traces, which intensifies with context length. By separating prompts into key evidence and distractor context, it shows the root cause is insufficient grounding on key evidence. The authors propose GEAR, a reward shaping method with evidence overlap reward and distractor penalty, achieving consistent improvements of up to +4.6 average points over standard RL across multiple benchmarks, with larger gains at longer contexts, while reducing copying and thinking length.

## Key points

- Identifies repetitive copying failure mode in long-context LLMs, intensifying with context length.
- Root cause is insufficient grounding on key evidence, leading to indiscriminate copying.
- Proposes GEAR with accuracy, key evidence overlap reward, and distractor penalty.
- Validated across multiple model scales and benchmarks, achieving up to +4.6 average points gain, larger at longer contexts.

## Full text

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.

---

Verify against the original before citing; curated by Astral Feed: https://feed.astralor.com/items/itm_d72529776569c2e45570e4e9