Hi everyone! I’m still getting the hang of how these AI research tools actually work under the hood. 😅
I’ve been trying out Iris.ai for finding academic papers, and it’s pretty neat, but I’m a bit confused. When I give it a paper I like, how does the AI figure out which other papers are ‘related’? Does it just look at the citations, or is it reading the actual text and concepts? A simple example would really help me understand. Thanks so much for any insights!
It's not just citations, and it's not really 'reading' like a human. Think of it as mapping keywords and phrases into a mathematical space.
The tool creates a dense vector embedding of the paper's text. Papers with similar vectors are closer together in this conceptual space, so they get flagged as 'related'. Citations are a separate signal that might get layered on top, but the core is semantic similarity from the full text.
For your example, if your paper is about "neural networks for protein folding," it'll find other papers whose vector embeddings are nearby, likely about machine learning in biochemistry, even if they don't directly cite each other. The benchmark for these systems is usually precision/recall scores on known related paper sets.
Exactly! The vector embedding explanation is spot on. It's like each paper gets a unique "address" in a giant city of ideas. Papers on the same street are semantically similar.
A practical caveat I've seen is that these tools can sometimes miss newer, niche papers because they aren't densely represented in the model's training data yet. So while the conceptual space idea is powerful, the relatedness can sometimes have a slight recency or popularity bias built in.
It's also why the best tools blend that vector similarity with other signals, like the citations you mentioned, or even co-author networks. That layering helps fix some of the blind spots. Pretty cool tech.
audit often