gear-grounding-evidence-aware-reward

name	gear-grounding-evidence-aware-reward
version	1.0.0
description	Grounding Evidence-Aware Reward for Long-Context Reasoning to Reduce Repetitive Copying
trigger_words	["gear","grounding evidence-aware reward","long-context reasoning","repetitive copying"]
arxiv_id	2607.19345

GEAR: Grounding Evidence-Aware Reward

Overview

GEAR (Grounding Evidence-Aware Reward) is a reward shaping method that addresses repetitive copying in long-context reasoning by encouraging models to focus on task-relevant evidence while avoiding irrelevant distractors.

Key Components

1. Evidence-Distractor Separation

• Separates prompts into task-relevant key evidence and irrelevant distractor context
• Identifies that insufficient grounding is the root cause of repetitive copying
• Shows that models failing to focus on key evidence are more likely to answer incorrectly

2. Reward Shaping

• Grounding Reward: Augments accuracy signal with reward for overlap with key evidence
• Distractor Penalty: Applies penalty for overlap with irrelevant context
• Combined reward guides models toward productive reasoning rather than copying

3. Automated Evidence Annotation Pipeline

• Constructs evidence-annotated training data from arbitrary documents
• Enables GEAR application on natural-language datasets without manual annotation

Implementation Steps

1. Data Preparation: Use automated pipeline to annotate evidence in your training data
2. Reward Function Design: Implement combined reward with grounding and distractor components
3. RL Training: Train your long-context LLM using GEAR instead of standard accuracy-only rewards
4. Evaluation: Measure both task performance and copying behavior reduction

Benefits

• Up to +4.6 average points improvement over standard RL
• Larger gains at longer contexts
• Reduces repetitive copying behavior
• Decreases thinking length (more efficient reasoning)
• Improves grounding in relevant evidence

Use Cases

• Long-context question answering
• Document-based reasoning tasks
• Legal or medical document analysis
• Any scenario requiring complex reasoning over long inputs

References

• Paper: Copy Less, Ground More: Overcoming Repetitive Copying in Long-Context Reasoning via Evidence-Aware Reinforcement Learning
• Related work: reinforcement learning, long-context LLMs, reward shaping, evidence grounding