Two separate research teams taught AI to follow useful connections instead of staring at a messy pile of information.

What are we talking about?

• Knowledge graph: a map made of things and the links between them, like “city is inside country.”
• Framework: a reusable recipe for building or testing a computer system.
• Seizure: a sudden burst of unusual electrical activity in the brain.
• EEG: a recording of brain signals made with sensors placed on the head.
• Diagnosis: a clinician’s process for working out what may be causing a health problem.

These are two different projects

HG-RAG helps a language model fetch connected facts from a hierarchy. NeuroGRIP helps a seizure-detection model check a map built from EEG signals against medical knowledge. They share a graph idea, but one does not perform the other’s job.

Paper 1: HG-RAG follows the family tree of facts

Ordinary retrieval-augmented generation, or RAG, often searches flat text chunks for words with similar meanings. That can miss a fact whose importance comes from its position. A city’s parent country, for example, may be the right answer even when its text does not look very similar to the question.

HG-RAG starts at a named item, walks through parent, neighbor and sometimes child links, then turns the selected subgraph into structured context. The author tested it with Mistral 7B on synthetic worlds containing planets, countries and cities. Across those controlled tests, it beat a flat vector-search baseline on hierarchy, relationship and multi-step questions, especially as the worlds grew.

Paper 2: NeuroGRIP checks an EEG brain map

Some seizure-detection models treat EEG sensors as dots and learned connections as lines. Those lines can be noisy or hard to explain. NeuroGRIP first builds a medical knowledge graph from clinical guidelines and literature, using a large language model to extract entities and relationships.

The framework compares local parts of the EEG graph with medical knowledge. It scores a proposed connection using semantic similarity, relation matching and source reliability, then removes connections with too little support. In experiments on the TUSZ and CHB-MIT benchmark datasets, the authors report consistent improvements across several graph models and more compact, interpretable graphs.

What changed?

Why it matters

• Connected evidence can be more useful than a bag of similar text.
• HG-RAG offers a reproducible way to test hierarchy-aware retrieval.
• NeuroGRIP aims to make a model’s brain map easier to inspect, not just produce a label.
• Both designs show where evidence came from and how it was selected.

What it does NOT prove / caveats

• Both papers are arXiv preprints, not final proof that the methods will work everywhere.
• HG-RAG was tested on invented worlds, uses a fixed 15-node context cap and needs a named entity in the question. The same Mistral model also answered and graded open-ended responses, which can bias scores.
• NeuroGRIP reports benchmark results, not a prospective clinical trial or approval for patient care. Its extracted knowledge can inherit errors or bias from sources and language models.
• NeuroGRIP is decision-support research. It does not replace a neurologist or the full process used to diagnose epilepsy.

End-to-end recap

1. Start with a hard question or a noisy EEG graph.
2. Find the small part of a connected knowledge map that matters.
3. Keep links supported by structure or medical evidence.
4. Give the cleaner evidence to the AI model.
5. Measure the result on controlled benchmarks.
6. Test much more before trusting it in new worlds or real clinics.