Context-Augmented KG Training Boosts Multi-Hop QA Accuracy on Disease Graphs
A new arXiv paper tackles the difficulty of multi‑hop question answering by enriching knowledge‑graph (KG) supervision with surrounding context. The authors extract disease‑specific KGs for gastroparesis and diabetes using the GraphMERT framework, then attach supporting triples from the same source text to each primary triple, forming a context graph (CG). Two training regimes are compared:…
Key points
- Context‑augmented supervision adds supporting triples to each KG fact, improving multi‑hop QA performance.
- A repair pipeline achieves 100% accuracy on cleaned one‑hop validation sets before RL fine‑tuning.
- Qwen3‑14B models trained with CG supervision outperform KG‑only models on 3‑ to 5‑hop disease queries.
To address persistent one‑hop errors, the researchers introduce a history‑aware repair pipeline that automatically flags unresolved failures, fine‑tunes on targeted repair examples, and isolates noisy triples. This step drives the models to 100 % accuracy on cleaned one‑hop validation sets. Finally, reinforcement learning on lower‑hop questions further improves performance on harder three‑, four‑, and five‑hop tasks, with the CG‑augmented models consistently outperforming KG‑only baselines.
The work demonstrates that providing richer contextual supervision and iterative repair can substantially raise the reliability of LLMs on complex reasoning over structured biomedical knowledge.
The headline, key points and digest above were generated by Digest AI's editorial model from the linked sources. Automated summaries can contain errors: the sources are the record. Spotted a mistake? Tell us.
More in Research
All →- R2VC Boosts Fact‑Checking Accuracy by 13.74% on FEVER Using Modular Retrieval and Calibration · 1 src
- Latent-Attention Masked Autoencoders Boost Multimodal Cardiac AI on 1.2M Hospital Stays · 1 src
- Language Models Fall Short for Quantitative Decision Tasks, Proposing Large Quantitative Models · 1 src
- GTA agent and GT Bench improve LLM graph reasoning accuracy · 2 src
- New Theory Links Memory Compression to Factual Hallucination in LLMs · 1 src
Comments
via GitHub Discussions