GEAR: Evidence-Aware Rewards Fix Repetitive Copying in Long-Context LLM Reasoning

Forum topic · ✨步子哥 · 2026-07-22

Summary

Researchers from Peking University and Alibaba identify a widespread failure mode in long-context LLM reasoning called 'repetitive copying': models copy large chunks of the input prompt verbatim into their reasoning traces instead of focusing on task-relevant evidence. Across seven frontier models (Qwen3, Claude, DeepSeek series), copying worsens as context length grows from 32k to 128k tokens, and copied content correlates with errors. Analysis shows correct answers draw mainly from task-relevant key evidence, while wrong answers copy evenly from evidence and irrelevant distractors—indicating insufficient grounding. The proposed GEAR (Grounding Evidence-Aware Reward) decomposes RL reward into accuracy, grounding overlap with key evidence, and a distractor penalty, with an automated pipeline for extracting evidence annotations from natural-language data. GEAR outperforms standard accuracy-only RL across three Qwen3.5 model scales and five long-context benchmarks, with gains growing at longer contexts (up to +15.6 points at 128k). Behavioral metrics confirm reduced copying and shorter reasoning traces.

Ask an LLM to read a 50-page legal contract and identify the three most critical breach clauses, and you may get something unexpected: instead of reasoning, the model copies the contract nearly verbatim into its thinking trace—page by page—before finally answering. This is not reasoning; it's a copy machine.

A team from Peking University and Alibaba, in a July 2026 paper, named this failure mode repetitive copying and found it is surprisingly common in long-context reasoning—and worsens as context grows.

Seven frontier models, all affected

The team tested seven frontier long-context LLMs, including the Qwen3, Claude, and DeepSeek series:

All models exhibit repetitive copying, and it worsens with context length.
At 32k context, a substantial fraction of reasoning traces is copied verbatim from the prompt; at 128k, the share is even higher.
Repetitive copying correlates strongly with wrong answers.

But there is a subtlety: copying itself isn't wrong—copying indiscriminately is. Splitting each prompt into task-relevant key evidence and irrelevant distractor context, the researchers found that correct samples copy mostly from key evidence, while incorrect samples copy uniformly across evidence and distractors. Like a student highlighting an entire textbook, copying everything means nothing is highlighted. The root cause is insufficient grounding: the model fails to anchor on task-relevant evidence.

GEAR: teaching models to highlight

GEAR (Grounding Evidence-Aware Reward) decomposes the standard RL accuracy reward into three components:

Accuracy reward (R_accuracy): reward for correct answers (the traditional signal)
Grounding reward (R_ground): higher overlap between the reasoning trace and key evidence earns more reward
Distractor penalty (R_distractor): higher overlap with irrelevant distractor context is penalized

Total reward: R = R_accuracy + α·R_ground − β·R_distractor

Since real-world data lacks evidence annotations, the team built an automated pipeline that extracts key evidence and distractor segments from arbitrary documents, making GEAR applicable beyond synthetic benchmarks like RULER.

Results: the longer the context, the bigger the gains

Tested on three model scales (Qwen3.5-9B, 35B-A3B, 27B) and five long-context benchmarks, GEAR beats standard accuracy-only RL in every configuration:

32k context: average gains of +4.6 to +8.1 points
128k context: average gains of +5.0 to +15.6 points
Gains grow with context length, consistent with grounding mattering more as distractors multiply

Example (Qwen3.5-9B, 128k context):

| Method | RULER | LongBench-v2 | Graphwalks | AA-LCR | |---|---|---|---|---| | Standard RL (GSPO) | 84.1 | 52.2 | 69.5 | 86.0 | | GEAR | 90.8 | 54.7 | 71.7 | 88.8 |

Behavior actually changes

GEAR significantly reduces answer-input overlap (3-gram level) and shortens reasoning traces—models stop copying verbatim and reason more concisely. This confirms the gains come from improved grounding, not side effects.

Ablations show both components are needed: grounding reward alone ("where to look") and distractor penalty alone ("where not to look") each help, but the full combination works best.

Why this paper matters

The paper's deeper contribution is its diagnose-first, treat-later methodology: it systematically quantifies repetitive copying, traces it to insufficient grounding, and only then designs the reward. It also highlights a key insight: the bottleneck of long-context reasoning is not how much a model can read, but how well it can focus—much like an expert skimming a 100-page contract, LLMs must learn which parts deserve close reading.

Paper: Copy Less, Ground More: Overcoming Repetitive Copying in Long-Context Reasoning via Evidence-Aware Reinforcement Learning Authors: Lizhe Fang, Weizhou Shen, Tianyi Tang, Yisen Wang (Peking University / Alibaba Group) Date: July 21, 2026