"PRISM" Is Not All You Need: Cross-Topic Name Collisions and Scientific Paper Retrieval

"PRISM" Is Not All You Need. Cross-Topic Name Collisions and Scientific Paper Retrieval. By Jingjie Ning and two others, from Carnegie Mellon University. Published on the first of November.

Scientific names increasingly connect papers from different research topics. When two unrelated papers share a name, a reader has to recover which paper is meant from context. The claim is that this convergence is growing quickly across topics, much faster than within them, and that it leaves a real retrieval problem behind even after a search is given research clues.

The study covers about five hundred and sixty-five thousand AI-related arXiv records. Taking the years two thousand twenty through two thousand twenty-two as a pooled reference, cross-topic name collision grows four point three three-fold by two thousand twenty-six. Within topics, the growth is only one point one two-fold. A second sentence encoder reproduces the pattern, and resampling over name groups quantifies its uncertainty. Fixed bands of word frequency locate the main increase in vocabulary of intermediate frequency, words shared by many research areas.

The second half measures the ambiguity that survives once research clues are supplied. Using the same three-term query construction, targets that share a name keep a matched identification deficit of twenty-five point seven percentage points under title-based BM25, and six point nine points under E5 with titles plus abstracts. Among fixed candidates that all carry the same name, listwise ranking by a language model reaches eighty-nine point two percent using only the query and the candidate titles.

The evidence is drawn from AI-related arXiv records, and the trend is measured against that pooled reference period. Topic-specific comparisons show that context resolves more of the ambiguity across topics, so the deficit is not uniform. The evaluation is built to separate three things: competition among candidates, the document fields used, and contextual ranking, tested through lexical, dense, and language-model baselines.

The title compresses the point. A name alone increasingly fails to identify a paper, and the reader, or the retrieval system, has to lean on context to tell which one is meant. The authors provide source-linked cases and a portable evaluation collection, so the phenomenon can be inspected and reused by people building scientific retrieval systems and research assistants.