GEAR: Evidence-Aware Reward Fixes Repetitive Copying in LLM Long-Context 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, wasting token budget and correlating strongly with wrong answers. Testing seven frontier models (Qwen3, Claude, DeepSeek series), they find copying worsens as context grows from 32k to 128k tokens. Crucially, successful answers copy mostly task-relevant key evidence, while failed answers copy indiscriminately from both evidence and irrelevant distractors—the issue is poor grounding, not copying itself. The proposed GEAR (Grounding Evidence-Aware Reward) decomposes RL reward into accuracy, a grounding reward for overlap with key evidence, and a distractor penalty, using an automated pipeline to extract evidence labels from natural-language data. Across three Qwen3.5 model scales and five long-context benchmarks, GEAR beats standard RL by +4.6 to +8.1 points at 32k and +5.0 to +15.6 at 128k, while reducing copying behavior and reasoning length. Ablations confirm both reward components are complementary.

Ask an LLM to read a 50-page legal contract and identify the three most critical breach clauses, and you may see its reasoning trace simply transcribe the document page by page before answering. This is not reasoning—it's a copy machine.

The Problem: Repetitive Copying

A team from Peking University and Alibaba (July 2026) names this failure mode repetitive copying. They tested seven frontier long-context LLMs, including Qwen3, Claude, and DeepSeek series:

All models exhibit repetitive copying, and it worsens as context length grows. At 32k context, a substantial share of reasoning traces is verbatim prompt copying; at 128k, the share is higher.
Repetitive copying correlates strongly with wrong answers. But there is a subtlety: copying itself is not the error—indiscriminate copying is.

The team split each prompt into task-relevant key evidence and irrelevant distractor context, then measured what models copied:

Correct answers: copied content came mainly from key evidence.
Wrong answers: copied content spread evenly between evidence and distractors.

It's like highlighting a textbook: strong students highlight only exam points; weak students highlight everything. The root cause is insufficient grounding—models fail to anchor on task-relevant evidence.

GEAR: Grounding Evidence-Aware Reward

GEAR decomposes the standard RL accuracy reward into three components:

R_accuracy: reward for a correct answer (standard practice)
R_ground: reward proportional to overlap between the reasoning trace and key evidence
R_distractor: penalty proportional to overlap with irrelevant distractor context

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

Since natural-language datasets lack 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 Gain

GEAR was tested on three Qwen3.5 scales (9B, 35B-A3B, 27B) and five long-context benchmarks, outperforming standard accuracy-only RL in all configurations:

32k context: average gains of +4.6 to +8.1 points
128k context: average gains of +5.0 to +15.6 points

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

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

Behavioral measurements confirm the mechanism: GEAR significantly reduces both answer-input overlap (3-gram level) and thinking length—models stop transcribing and reason more concisely.

Ablations

Removing either component degrades results: grounding reward alone or distractor penalty alone each improve over standard RL, but the full GEAR is best. The grounding reward tells the model "where to look"; the distractor penalty tells it "where not to look."

Takeaways

The paper's strength is its diagnose-first approach: it quantifies repetitive copying, traces it to insufficient grounding, and only then designs the fix. More broadly, the bottleneck of long-context reasoning is not how much a model can read, but how well it can focus—mirroring how human experts skim long documents by locating key clauses rather than reading every word.

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