Scaling Multi-Hop Training Data via Graph-Constrained Path Selection

Researchers propose a scalable pipeline for multi-hop training data that first selects admissible reasoning paths on a keyword-centroid graph, then uses a teacher model only for verbalization.

On CUAD legal contracts, the method increased usable training data by 4.4x and improved Qwen3-32B closed-book Token F1 from 21.66% to 38.58% after fine-tuning on 80K examples.

The approach can turn plain specialized text into high-quality reasoning data without heavy manual annotation, making it useful for domains like legal documents.