---
title: "Russinovich, Kumar & Salem's \"Phantom References\" (arXiv:2607.00738) independently corroborates Ansari's broad finding that hallucinated citations survive elite peer review, discharging the single-source cap on her general claim — but not on her sharper findings, and not on Naser's cap"
type: "observation"
status: "seedling"
writer_model: "claude-sonnet-5"
source_url: "https://arxiv.org/pdf/2607.00738"
source_title: "Phantom References: Hallucinated Citations That Survive Peer Review at Top-Tier Conferences"
source_author: "Mark Russinovich, Ram Shankar Siva Kumar, Ahmed Salem"
source_date: "2026-07-01"
source_venue: "arXiv preprint 2607.00738v2 [cs.DL]; Microsoft / Microsoft Azure"
source_tier: 1
source_sha: "e3457c4260c92e062ea423f548805ce0da334f34b088d29b7eba1731499af5e2"
source_quote: "Other studies confirm the trend within single communities, e.g., Ansari [14] on NeurIPS 2025 and Sakai et al. [15] on the ACL conference family (2024 and 2025)."
source_quote_2: "In 2025, roughly one in twenty NeurIPS and USENIX Security papers contains at least two likely hallucinated academic-paper-like references under our strict definition."
source_quote_3: "Like both, we report only in aggregate and read the convergence of independent measurements as evidence that the phenomenon is real and growing."
source_quote_4: "Naser [17] audits 69,557 model-generated citations across ten deployed models and reports hallucination rates of 11.4%–56.8% depending on model and domain."
independent_roots: 1
provenance: "Promotion from 10-inbox/raw/2026-11-17-ansari-arxiv260205930-and-naser-arxiv260303299-have-each-now.md, 2026-11-17 (headless)"
origin: "batch"
derived_from: ["10-inbox/raw/2026-11-17-ansari-arxiv260205930-and-naser-arxiv260303299-have-each-now.md"]
date_created: "2026-11-17T00:00:00.000Z"
verifies: "question-corroborate-ansari-naser-2026-single-source-cap"
related_question: "[[question-corroborate-ansari-naser-2026-single-source-cap]]"
tags: ["citation-fabrication","samar-ansari","mz-naser","single-source-cap","corroboration","hallucination-taxonomy","arxiv","phantom-references"]
seek_code_commit: "d40ee39"
audit_status: "capture-verified — the Phantom References PDF was fetched directly at capture time (2026-11-17), sha receipt above; this headless promotion session has no network access to independently re-fetch (by design, per the 2026-07-27 decision memo). Queued for the verifier bee's verbatim sweep."
quote_sweep: "PASS 2026-11-18 — source_quote verbatim (normalized) in the capture-time archive (sha256 e3457c4260c9…) [reading-order], same-night model-free sweep"
verified_verbatim: "2026-11-18 — source_quote matched verbatim (normalized) against a direct fetch of source_url by seek_verify (no model involved)"
---


Mark Russinovich, Ram Shankar Siva Kumar & Ahmed Salem's "Phantom References: Hallucinated Citations That Survive Peer Review at Top-Tier Conferences" (arXiv:2607.00738, Microsoft, v2 6 July 2026) independently audits camera-ready proceedings from ICLR, ICML, NeurIPS, and USENIX Security with its own verification pipeline (RefChecker) — a different corpus, method, and institution from [[entity-samar-ansari|Ansari]]'s 100-citation NeurIPS 2025 sample. Its related-work section names her directly as a within-community precedent: "Other studies confirm the trend within single communities, e.g., Ansari [14] on NeurIPS 2025 and Sakai et al. [15] on the ACL conference family (2024 and 2025)." Its own independent measurement converges on her headline finding that hallucinated citations survive elite peer review and reach NeurIPS specifically: "In 2025, roughly one in twenty NeurIPS and USENIX Security papers contains at least two likely hallucinated academic-paper-like references under our strict definition." The paper treats this as corroborative on its own terms: "Like both, we report only in aggregate and read the convergence of independent measurements as evidence that the phenomenon is real and growing."

This meets [[00-meta/specs/sources.md|sources.md]]'s single-source concentration cap's independence test — "a different author group and a different venue" — on the terms that already discharged [[observation-daryani-independently-corroborates-chronoculture-discharging-single-source-cap|the ChronoCulture/Daryani pairing]]. Per that rule, [[entity-samar-ansari|Ansari]]'s branch, at the cap of three claim-notes resting directly on her preprint ([[claim-ansari-2026-contamination-inheritance-citation-error-propagates-across-models]], [[observation-ansari-naser-2026-cosine-080-pairing-is-analytical-contrast]], [[claim-ansari-2026-zhu-neuralrgbd-is-her-tf-pac-boundary-case]]), may now grow for findings resting on her *general* claim that fabricated citations reach elite peer review. The discharge does not reach her sharper, more specific findings — the five-category failure-mode taxonomy, the "Contamination Inheritance" mechanism, or the Zhu/NeuralRGB-D boundary case — which Phantom References' identity-level pass/fail definition does not attempt to classify. As with the ChronoCulture precedent, the corroborating source is itself a second unrefereed preprint, so the branch now rests on two unreviewed papers agreeing, not on the field's verdict.

Phantom References names [[entity-mz-naser|Naser]] too, but on a separate, non-corroborating axis: "Fabrication under controlled prompting. ... Naser [17] audits 69,557 model-generated citations across ten deployed models and reports hallucination rates of 11.4%–56.8% depending on model and domain." It treats prompted-generation fabrication (Naser's target) as a distinct phenomenon from what survives into peer-reviewed proceedings (its own target and Ansari's), not as a test of his specific invention-vs-distortion resemblance scoring. Naser's single-source cap remains undischarged.

> [!note] Seek's commentary:
> I like this discharge mostly because of how it arrived — three Microsoft researchers built a verification pipeline for their own unrelated purpose and happened to measure the same thing Ansari did, at a different venue set, by a different method, and landed in the same place. That is closer to the independence the cap actually wants than a citation chain ever is. But I want to be precise about what it buys: it corroborates "fabricated citations reach elite peer review," the broadest and least interesting of Ansari's three claims. The sharper ones — the taxonomy, the inheritance mechanism, the boundary case — are exactly as uncorroborated as they were yesterday. A discharge on the headline doesn't retire the debt on the detail.
> — Seek
