Abstract
<title>Abstract</title> <p>Large language models have changed how information retrieval research is done, but the bibliometric shape of that change has not been described in detail. This paper asks three things: how the information retrieval (IR) and large language model (LLM) literatures have grown into each other, which publications hold the two together, and how the structure of the field has shifted year on year. We assembled a corpus of retrieval-related work published between 2015 and 2025 from OpenAlex, built a bibliographic coupling network with bibliometrix, and analysed it in Gephi. After thresholding by coupling strength and removing several off-topic clusters that entered through the ambiguity of the word “retrieval,” we were left with 4,064 documents and 71,534 coupling links. These fall into six main communities, spanning multimodal and vision–language retrieval, neural and conversational document retrieval, transformers and LLMs for IR, neural ranking, graph and code retrieval, and hashing-based retrieval. Betweenness centrality points to the neural-IR survey literature as the work that does most of the bridging between these groups. Splitting the network into four time periods shows a pattern easy to miss in a single static map: the field first diversifies, with modularity rising through 2019 to 2021 as transformer methods spawn new sub-areas, then converges, with modularity dropping sharply in 2022 to 2023 (Q = 0.342) as LLM and retrieval-augmented generation work pulls the separate communities together. The results suggest that IR has not simply borrowed LLM techniques but reorganised around them, with neural ranking research at the centre as the connective tissue.</p>