The research presented a novel approach to multi-hop question answering by incorporating context-augmented knowledge graph reasoning. The framework utilizes the Qwen3-14B model and focuses on improving its ability to handle complex reasoning tasks that require connecting multiple facts. The core of the method involves creating a context graph (CG) for each primary KG triple by adding supporting triples extracted from the same source text chunk. This creates two distinct supervision settings: KG-grounded supervision, which relies solely on the target KG triple or path, and CG-grounded supervision, which incorporates the target triple along with the associated context triples.
Training was conducted using supervised fine-tuning (SFT) with both supervision settings. This resulted in two model variants: KGModel and CGModel. To further enhance the model’s factual foundation, an LLM-judged, history-aware adaptive repair pipeline was implemented. This pipeline identifies unresolved one-hop failures, continuously fine-tunes the model on targeted repair examples, and manages problematic noisy triples, leading to 100% accuracy on cleaned one-hop validation sets.
Following the initial SFT, reinforcement learning (RL) was employed using lower-hop question-answer items. The RL process was initialized from the repaired SFT checkpoints, yielding larger and more stable gains in performance. Evaluation was conducted on harder multi-hop tasks, including 3-hop, 4-hop, and 5-hop questions across Gastroparesis and Diabetes knowledge graphs. Context-augmented supervision consistently improved multi-hop performance compared to KG-only supervision.
The research utilized GraphMERT for KG extraction and focused on evaluating the framework’s effectiveness using Qwen3-14B. The repair pipeline and RL component were key to achieving high accuracy and generalization across different hop lengths.
Source: https://arxiv.org/abs/2609.12230