Could the bare record of who co-wrote with whom reveal a scientist's expertise, without asking how they think?

This explores how the bare record of who wrote papers with whom, and on what, can stand in for a scientist's expertise without anyone being asked how they think. The clearest example in the collection is the human-aware hypergraph work Can predicting scientists improve discovery forecasts?. It links papers, materials and authors into one network, then runs random walks across it. A walk that goes from a material to an author who studied it, then to a co-author, then to another material they worked on, roughly traces how an idea could actually reach someone. The model never learns what anyone believes. It learns who is in a position to notice a connection. That alone predicted discoveries 43% more precisely than models that read only paper content, and the biggest gains came where the literature was thin.

The surprising part is why this works. Expertise is not only knowledge in someone's head. It is also a position in a social network: whom you talk to, what you have handled, which lab's methods you picked up. Another note makes the reverse argument about language models Can language models distinguish expert arguments from common assumptions?. LLMs read text cut off from reputation and track record, so they cannot tell an expert's claim from a common assumption. Put the two notes side by side and you see that the collaboration graph holds exactly the social context that text-only models lose. Who-worked-with-whom is a compressed record of standing and access.

Community behavior can stand in for judgment in other ways too. One project trained models on 700,000 citation-matched paper pairs to learn 'scientific taste' Can models learn what makes research worth doing?. It never asked researchers why a paper matters. It treated the field's collective citations as the signal. The same logic in reverse explains why fine-tuned LLMs beat neuroscientists at predicting which experimental results really happened Can LLMs predict novel scientific results better than experts?. Patterns across many papers can capture regularities that no single expert puts into words.

This approach has a cost that is easy to miss. If AI tools change who works with whom, the graph these models rely on changes too. AI-augmented researchers publish three times as many papers, yet collaboration across the field fell 22% and the range of topics shrank Does AI help individual scientists while narrowing scientific focus?. A model that reads expertise from co-authorship would see a narrower, more concentrated map of expertise. That map might be accurate, but it would describe a field that is exploring less. Compare this with writing expertise down directly: building expert rules into an agent's scaffolding Can codified expertise let non-experts match specialist output? captures the 'how' that co-authorship graphs skip. That is costly to do, but it is explicit.

A plain caveat: only the hypergraph study in this collection models expertise directly from collaboration structure. The other notes come at the question from the side, through citations, social authority and shifts in collaboration. So the collection supports 'it works, and here is why it might' more than a detailed account of what co-authorship graphs miss.