# 长上下文推理模型陷入复制陷阱，新方法提升准确率4.6%

- 来源：Rohan Paul (@rohanpaul_ai)
- 发布时间：2026-07-23 13:52
- AIHOT 分数：31
- AIHOT 链接：https://aihot.virxact.com/items/cmrx3hk6v005bro69qh6ydnu9
- 原文链接：https://x.com/rohanpaul_ai/status/2080169034935591350

## AI 摘要

研究发现，即使是强大的推理模型在处理长输入时也会养成复制大段文本的“懒惰”习惯，且输入越长问题越严重，这会消耗token预算并降低准确性。复制无关内容会导致错误，而只复制关键行则有助于正确回答。研究者提出一种基于证据感知的强化学习奖励机制，通过奖励模型关注关键文本、惩罚复制无关内容，将准确率提升最多4.6个百分点。

## 正文

Many assumed that once long-context models could reason， the old challenge of simply finding the right text was behind us.

This paper shows that is wrong. Even strong reasoning models slip into a lazy habit of copying big chunks of the input into their own thinking， and the habit gets worse the longer the input is， quietly eating their token budget and dragging accuracy down.

Copying itself is fine； the damage comes from copying the wrong stuff. Models that copy the few lines that actually matter tend to get the answer right， while models that copy the surrounding filler tend to get it wrong.

They developed a simple reward that stops long-context language models from mindlessly copying their input， lifting accuracy by up to 4.6 points.

The fix rewards the model for engaging with the small set of text that actually matters and penalizes it for copying the irrelevant filler around it.

- arxiv. org/abs/2607.19345

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