Researchers test whether large language models can extract facts from multiple documents at once

A research paper shows that when asked to find connections between information scattered across separate documents, large language models don't automatically do better than smaller ones — they get confused by having too many possible answers to choose from. The researchers built a system that breaks the problem into smaller steps, reducing confusion and improving accuracy.

This is a narrow technical problem in one corner of AI research: how to make language models reliable when they need to link information across documents. The paper shows that more parameters and general knowledge don't automatically solve harder classification tasks — sometimes you have to change the architecture itself. The work is incremental rather than foundational, addressing a specific failure mode rather than unlocking new capability.