A reading of WISE: Workflow for Intelligent Scientific Knowledge Extraction by Sajratul Yakin Rubaiat and Hasan M. Jamil, University of Idaho (2025)
Every morning we sit down with a piece of research that connects to the problem we are working on. Today's paper is one of the most resonant we have encountered.
WISE — Workflow for Intelligent Scientific Knowledge Extraction — is a system built by researchers at the University of Idaho to address a specific challenge: how do you extract deep, accurate, non-redundant knowledge from an exponentially expanding web of interconnected sources, when the tools available — search engines, general-purpose LLMs — keep falling short at the edges of complex queries?
We are grateful this paper exists. Not because it validates anything we are doing, but because it illuminates the same class of problem from a completely different angle — scientific literature extraction — and does so with rigor, transparency, and results that speak for themselves. Rubaiat and Jamil published their methodology openly. They showed their math. They ran honest comparisons against established baselines and reported what they found without inflation. That is exactly the kind of work the research community needs more of, and exactly the kind of work we want to learn from.
We work on federated knowledge graph architecture for the American business economy, public law, and civic infrastructure. The problems we encounter are not the same as theirs. But reading their work carefully, we found ourselves recognizing the shape of familiar challenges described in a new language. That kind of cross-domain resonance is rare and worth paying attention to.
Start with a single gene. HBB. One authoritative source knows about it. That source links to 24 others. One of those links to hundreds more. Within a few traversal steps you have an exponentially expanding tree of interconnected knowledge — and no intelligent way to navigate it without either drowning in volume or stopping too early and missing what matters most.
Traditional search engines return a list and step back. The researcher does the rest manually.
General-purpose LLMs offer synthesized answers — but those answers are constrained by context window. Even the most capable models available can only hold so much simultaneously. The researchers found that GPT-4o, with its 128,000 token context window, could realistically process around eight sources at once before running out of room. This is not a criticism of any model. It is an honest accounting of a structural ceiling that affects every LLM-based retrieval system operating at scale.
The result, in domains where completeness matters — medicine, biology, materials science, social research — is answers that are confident but incomplete. The rare condition gets missed. The edge case goes unrecorded. The nuanced connection between distant sources never gets made.
WISE was built to do what a skilled expert researcher actually does: identify the most valuable leads, discard what is already understood, follow depth where it is warranted, stop when further exploration stops returning meaningful new knowledge, and surface something genuinely comprehensive at the end.
WISE operates as a tree. A query is submitted. An initial set of sources is retrieved. Then four operations run recursively, layer by layer, until a stopping condition is met.
| System | Diseases Found | Recall | Rare Conditions |
|---|---|---|---|
| WISE | 16 | 0.84 | ✓ |
| ChatGPT (GPT-4o) | 9 | 0.47 | ✗ |
| ChatGPT with Search | 7 | 0.36 | ✗ |
| Google Search | 3 | 0.15 | ✗ |
| Gemini | 2 | 0.10 | ✗ |
WISE's output also scored lower on ROUGE and BLEU overlap metrics than all other systems — meaning it wasn't just finding more, it was finding genuinely different and unique information that no other system surfaced.
We work on a different problem in a different context. But reading this paper carefully, several things came into focus that we want to share honestly.
The challenge of traversing an interconnected knowledge graph without drowning in redundancy is not specific to scientific literature. Any system that needs to reason across a large, heterogeneous corpus of structured entities faces the same fundamental tension: breadth versus depth, volume versus signal, traversal cost versus completeness.
What WISE demonstrates clearly is that the relationship between pieces of knowledge matters as much as the pieces themselves. A system that scores sources by unique contribution — that actively measures what each new source adds relative to everything already known — produces dramatically better results than a system that ranks by popularity or authority alone. The graph structure is not just a storage mechanism. It is a reasoning surface. The edges between entities carry meaning that the entities themselves cannot carry alone.
The researchers also identify knowledge graph integration as a direction they want to explore further — moving from a text-based knowledge container toward a node-and-edge representation where relationships are preserved explicitly rather than merged into accumulating text. Their preliminary experiments — representing the HBB gene entry as a graph of 56 nodes and 55 edges and filtering it to an 11-node, 16-edge subgraph aligned with a specific query — showed that structured representations can preserve relational meaning that text containers lose.
We find this direction genuinely exciting and we look forward to seeing where their future work goes.
RankWithMe.ai is a learning resource. We are building the most machine-readable map of the American business economy that has ever existed, and we are doing it openly — publishing our research, our methodology, our specification, and our thinking as we go.
Part of that commitment is honest engagement with the research community. Papers like this one represent years of careful work by researchers who deserve to be read, cited, and built upon. We read them, we share what we learn, and we point anyone who finds this useful back to the primary source.
If you work in knowledge graph architecture, LLM-based retrieval, linked data systems, or any adjacent field — this paper is worth your time. The full text is available on arXiv. The researchers are at the University of Idaho, Department of Computer Science. Their work is real, their results are independently verifiable, and their methodology is described with enough clarity to build on.
We are grateful they published it.
