Typed-citation data show ML, NLP and Vision each cite within their own community well above chance and across communities below it

citation-analysisbibliometricsinsularitycross-domain-bridge

SciTraj (arXiv 2606.22342), a study of 32,559 papers spanning 2015–2024, reports "disciplinary siloing" in the real Semantic Scholar citation graph for machine learning, natural-language processing and computer vision. Its Table XI gives a within-community citation propensity of roughly 1.99–2.26 (where 1.0 is the chance expectation under random mixing) for each of the three communities, against 0.06–0.65 for most cross-community pairs — i.e. each field cites itself about twice as often as expected and cites the other two far below expectation. The one exception is Vision→ML at 1.12, slightly above chance. [These figures were read from Table XI; the surrounding sentences failed quote_check on the table's layout at capture and the numbers carried [unverified-quant — needs direct read of Table XI], routed to question-verify-scitraj-2026-within-cross-community-citation-propensities. Discharged 2026-10-11 by a cross-model audit's direct read of a live re-fetch (sha256 match to the pinned archive): real-S2 diagonals 1.99 (NLP), 2.26 (ML), 2.12 (Vision); cross pairs 0.06–0.65; Vision→ML 1.12 the lone above-baseline cross pair; "expected" is a uniform target-choice baseline.]

The finding is the typed-citation complement to Wahle et al.'s CFDI decline: Wahle measures one field's outgoing diversity collapsing over time, SciTraj measures the cross-community propensities directly and finds all three of ML, NLP and Vision siloed from one another in a single recent snapshot. Both are field-level base rates for the vault's single-pair observation that organizational-forgetting and ML literatures show zero citation contact.