Does a direct read of Beltran et al.'s v1-vs-v2 diff (arXiv:2412.13176) confirm Ansari's Contamination Inheritance trace, or does it rest only on the identical-string inference?
Raised as question-verify-beltran-2024-v1-v2-diff-confirms-ansari-contamination-inheritance, asking whether a direct read of arXiv:2412.13176 converts Ansari's single identical-string inference into a directly observed before/after correction. This session fetched and read all four of the paper's arXiv versions in full (v1 through v4 — the question named only v1/v2, but the paper has grown to four versions since the question was raised) via extract_pdf, all served over verified TLS, and located the exact fabricated citation string by direct text read rather than search-summary or inference.
Short answer: partially confirmed, but not where the question's own framing expected it. v1 does contain the identical fabricated citation Ansari describes — that leg is now a direct read, not an inference. But the "later version corrected it" leg does not hold for v2, the version the question named: the identical fabricated text survives unchanged into v2, and into v3 as well. The correction only appears in v4 (posted 2026-01-23), two versions and roughly thirteen months after v1, and about two weeks before Ansari's own paper (dated 2026-02-06).
Claim: v1 of Beltran et al. (arXiv:2412.13176) contains, verbatim, the exact fabricated citation Ansari's paper describes
verifies: question-verify-beltran-2024-v1-v2-diff-confirms-ansari-contamination-inheritance
Claim type: historical/textual — settled by direct primary-document read, not inference. Tier 1 (the arXiv preprint itself, read in full).
Reference [62] of v1's bibliography reads, in full: "Z. Zhu, T. Yu, X. Zhang, J. Li, Y. Zhang, and Y. Fu. Neuralrgb-d: Neural representations for depth estimation and scene mapping. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2022." This is character-for-character the fabricated citation Ansari's paper names as the Contamination Inheritance case (per the existing vault claim-note, the same author string "Z. Zhu, T. Yu, X. Zhang, J. Li, Y. Zhang, and Y. Fu" attached to the invented title "Neuralrgb-d"). No real paper with this title and this author list exists; the genuine 2022 CVPR paper on this topic, "Neural RGB-D Surface Reconstruction," is by an entirely different author team (Dejan Azinović, Ricardo Martin-Brualla, Dan B Goldman, Matthias Nießner, Justus Thies — confirmed below from v4's corrected entry). This converts half of Ansari's trace — "the hallucination... encountered the erroneous citation in Beltran et al. (v1)" — from a reported identical-string match into a directly verified one.
- source_quote: "Z. Zhu, T. Yu, X. Zhang, J. Li, Y. Zhang, and Y. Fu. Neuralrgb-d: Neural representations for depth estimation and scene mapping. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2022."
- source_url: https://arxiv.org/pdf/2412.13176v1
- source_sha: 14308f80f5af5feec3416eb13f2b731df190e6760c65f708e98a971e42b6543a
- source_tier: 1
Claim: The v1-to-v2 diff does not show a correction — the identical fabricated citation persists unchanged into v2 (and into v3)
verifies: question-verify-beltran-2024-v1-v2-diff-confirms-ansari-contamination-inheritance
Claim type: historical/textual, settled by direct comparative read. Tier 1.
v2 (posted 2025-03-19, three months after v1) carries the identical fabricated reference, renumbered [71] but textually unchanged: "Z. Zhu, T. Yu, X. Zhang, J. Li, Y. Zhang, and Y. Fu. Neuralrgb-d: Neural representations for depth estimation and scene mapping. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2022." v3 (posted 2025-09-18) carries it again, renumbered [61], with the same wording. The specific diff named in the question — v1 vs v2 — therefore does not show the before/after correction the question asked whether a direct read would confirm; on that narrow comparison, the premise behind the question (that v2 is the corrected version) is false. This matters because the existing vault claim-note and its dependent bridge observations describe the correction only as occurring in "a later version," without naming which one — a direct read was needed to discover that "later" means much later than the next version up.
- source_quote: "Z. Zhu, T. Yu, X. Zhang, J. Li, Y. Zhang, and Y. Fu. Neuralrgb-d: Neural representations for depth estimation and scene mapping. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2022."
- source_url: https://arxiv.org/pdf/2412.13176v2
- source_sha: 5dfaddf1ec8e75e8f000cc6e222620b6e7e325853b1e8eff872bbbac3c4929da
- source_tier: 1
- source_quote_2: "Z. Zhu, T. Yu, X. Zhang, J. Li, Y. Zhang, and Y. Fu. Neuralrgb-d: Neural representations for depth estimation and scene mapping. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2022."
- source_url_2: https://arxiv.org/pdf/2412.13176v3
- source_sha_2: 8a089e4945de0085efaab6aa42174428e5e712755bf4bfc934cc2b235490f528
- source_tier_2: 1
Claim: The fabrication is eventually corrected, but only in v4 (2026-01-23) — not v2 — replacing the fabricated entry with the real Azinović et al. 2022 citation
verifies: question-verify-beltran-2024-v1-v2-diff-confirms-ansari-contamination-inheritance
Claim type: historical/textual, settled by direct read. Tier 1.
v4 of the paper (posted 2026-01-23, the version current as of this session) drops the fabricated "Z. Zhu, T. Yu..." entry entirely. Reference [2] of v4 instead reads: "Dejan Azinović, Ricardo Martin-Brualla, Dan B Goldman, Matthias Nießner, and Justus Thies. Neural rgb-d surface reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 6290–6301, June 2022." — the real paper, real authors, real page range. This is a genuine before/after correction, confirming the second leg of Ansari's trace ("a later preprint version corrected it") as true in substance. But the correction sits between v3 and v4, not v1 and v2 as the question (and, by extension, the vault's prior hedged characterization) assumed. v4 was posted roughly two weeks before Ansari's own paper (dated 2026-02-06) — close enough in time that whether Ansari's scrutiny prompted, coincided with, or was unrelated to the correction is not established by anything read this session; no connection between the two is claimed here.
- source_quote: "Dejan Azinović, Ricardo Martin-Brualla, Dan B Goldman, Matthias Nießner, and Justus Thies. Neural rgb-d surface reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 6290–6301, June 2022."
- source_url: https://arxiv.org/pdf/2412.13176v4
- source_sha: 5abb637b25e3325190dd741ffc834ea28ea908bc1b13c47bc99a280d07bd65c5
- source_tier: 1
Further leads
- v1 also contains a second, independent garbled citation sharing the same wrong author string ("Z. Zhu, T. Yu, X. Zhang, J. Li, Y. Zhang, and Y. Fu") misattached to a different, real paper's title ("Nice-slam," ref [60], year given as 2021) — sitting alongside a separately correct NICE-SLAM citation with the true author list (ref [61]: Zihan Zhu, Songyou Peng, Viktor Larsson, Weiwei Xu, Hujun Bao, Zhaopeng Cui, Martin R. Oswald, Marc Pollefeys, CVPR 2022) in the same v1 bibliography. This duplicate/garbled-author pattern — the identical wrong six-name string reused across two different reference entries — is suggestive of an LLM-assisted citation-generation error distinct from the Neuralrgb-d fabrication Ansari traced, and was not checked against Ansari's dataset or methodology this session.
[unverified-mechanism — needs primary]for any claim about how the duplication arose. - v2's citation [69] for NICE-SLAM is itself a third, distinct mutation — it uses the wrong author string from v1's NICER-SLAM entry ("A. Zhu, Z. Zhang, H. Su, L. Li, G. Wang, and X. Li") but attaches it to the title "Nice-slam" with the correct 2022/CVPR venue — not identical to any single v1 entry. Whether this reflects further LLM-assisted editing between v1 and v2, or manual error, was not established this session.
- Whether the specific NeurIPS 2025 paper Ansari's dataset identifies as reproducing the fabricated citation is itself named anywhere publicly (her own paper's quoted text, per the existing vault claim-note, does not name it) remains unchecked — would require Ansari's supplementary data or dataset, not located this session.
- v3 (2025-09-18) and v4 (2026-01-23) both add a co-author (Stephen Pizer, from v2 onward) and substantially restructure the paper (dropping the NICER-SLAM comparison, adding real-world Colon10k and indoor-scene evaluations) — a fuller publication history than the vault had previously recorded; not otherwise relevant to the Contamination Inheritance question.
Entity candidates
- Dejan Azinović — person — first author of the real "Neural RGB-D Surface Reconstruction" (CVPR 2022), the genuine paper whose identity was effectively erased by the fabricated citation in Beltran v1–v3; the foundational, misattributed figure this whole trace is measured against. Flagged first per the instruction to catch the ancestor, not just the citing paper's authors.
- Zihan Zhu — person — first author of the real NICE-SLAM (CVPR 2022), whose correct author list appears once (v1 ref [61]) alongside the incorrect six-name string wrongly attached to NICE-SLAM elsewhere in the same bibliography (v1 ref [60]) — a second, independent misattribution case in the same document, adjacent to but distinct from the Neuralrgb-d fabrication.
- Andrea Dunn Beltran — person — co-first-author of arXiv:2412.13176 across all four versions; no dedicated entity page exists yet beyond the watching term-stub entity-beltran-et-al-2024-v1.
- Daniel Rho — person — co-first-author of arXiv:2412.13176 across all four versions.
- Stephen Pizer — person — added as a co-author starting with v2 (2025-03-19); absent from v1.
- Marc Niethammer, Roni Sengupta — persons — co-authors across all four versions (UNC Chapel Hill).
- Samar Ansari — person — already a vault hub (entity-samar-ansari); this capture resolves the open verification question against her Contamination Inheritance claim and should be linked from her page as the direct-read confirmation (with the v4-not-v2 correction noted).