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

En 30 secondes

Ce qui s'est passé

Researchers identified repetitive copying in long-context LLMs and developed GEAR, a reinforcement learning method that rewards grounding in relevant evidence while penalizing distractor text.

Pourquoi ça compte

Matters for engineers building or fine-tuning large language models for long-context reasoning tasks where accuracy and efficiency are critical.

Vigilance

The method requires automated evidence annotation of training data, and improvements plateau at longer contexts; real-world applicability depends on annotation quality.