"PRISM" Is Not All You Need: Cross-Topic Name Collisions and Scientific Paper Retrieval
Jingjie Ning, Xueqi Li, Yibo Kong
cs.IR, cs.DL
2026-11-01
Cross-topic name collisions on arXiv hit 4.33x the 2020-2022 baseline; shared-name papers trail unique-name controls by 25.7 points under BM25, plus a reusable benchmark.
Naming AI methods has become an uncoordinated land grab. A single token, PRISM, separately names a spatial world model, a pathology encoder, and a model-safety method on arXiv. A three-author CMU team analyzed 564,861 AI-related arXiv records from 2020 through late October 2026 and formalized the pattern as name collision: distinct papers sharing the same leading title label. The cross-topic collision rate now sits at 4.33 times the pooled 2020-2022 reference, against 1.12-fold within topics. What grew is cross-field reuse of names, not paper volume.
The bill lands on retrieval. In the fixed collection, 28,152 papers, 19.5% of labelled records, have at least one namesake, and an affected paper faces 8.33 same-name rivals on average. Locating a paper by a short name keeps getting less reliable.
A label is the compact token before the first colon in a title, compared after case folding; suffixes stay in the key, so SageAttention 1/2/3 count as three distinct labels. Collision rate is the probability that two randomly drawn labelled records share a name, reported per 100,000 pairs. Sentence embeddings partition the abstracts into 200 topic clusters, with the title label masked first, splitting collisions into cross-topic and within-topic trajectories.
Robustness stacks several ways: 2,000 name-group bootstrap draws for intervals; a second encoder (MPNet) reproduces 4.38x cross-topic growth; excluding pairs that share an author still leaves 4.40x; dropping six format labels and the 20 most frequent names leaves 1.91x. The increase survives every cut.
A wordfreq analysis pins each label to a fixed Zipf frequency band. Intermediate-frequency words, roughly 1 to 100 occurrences per million tokens, make up only 12.4% of labelled 2026 records but account for 89.4% of cross-topic matching pairs and 98.0% of the net increase. The community draws names from one shared pool of familiar, jargon-free words.
The retrieval evaluation matches 1,400 shared-name targets with 1,400 unique-name controls on year, category, and title length. A query is the name plus three terms auto-selected from the target's abstract under one corpus-wide frequency-by-IDF rule. Two scopes: full-index search over 144,523 titles, and ranking within the fixed same-name candidate group. Baselines span BM25, MiniLM, E5, and listwise ranking with DeepSeek-Flash.
| Setting | Ranker | Top-1 |
| Full index, titles | BM25, unique vs shared | 85.1% vs 59.4%, gap of 25.7 pp |
| Full index, titles | MiniLM, unique vs shared | gap of 13.2 pp |
| Full index, title+abstract | E5, unique vs shared | 94.8% vs 87.9%, gap of 6.9 pp |
| Same-name group, name then +3 clues | BM25 | 30.8% to 70.6% |
| Same-name group, +3 clues | DeepSeek-Flash listwise | 89.2%, Top-10 98.9%, MRR 0.933 |
| Same-name group, +3 clues | E5, title+abstract | 93.6% |
Three readings fall out. Three concrete abstract terms absorb most of the ambiguity: within the 148-record PRISM group, cytoplasmic, perinuclear, and lymphoblastic rank the cell-segmentation paper first under both scorers. Stronger rankers shrink the residual, but the name tax never disappears. Same-topic collisions are the harder case: context removes 58.9% of cross-topic error under BM25 but only 24.4% of same-topic error, and a fixed two-candidate comparison gives 80.0% versus 66.9%. Two AutoInit papers that both initialize networks fool BM25, MiniLM, and E5 title ranking; only listwise reranking recovers the target.
For anyone building scholarly search or literature-grounded assistants, this turns name ambiguity from anecdote into a reproducible benchmark: fixed corpus, frozen queries, 60,200 scored outcomes, and all baseline outputs ship together. The design guidance is concrete: a short-name query should surface the candidate group, context should rank inside it, and the result should carry a verifiable identifier. For authors, check a name before adopting it and avoid familiar dictionary words first. This is measurement plus benchmark, not a new model, an incremental but solid contribution.