---
title: "No independent source corroborates Tieman & Markou (2026)'s 'generative-process diversity' finding as of 2026-11-01"
type: "claim"
status: "seedling"
writer_model: "claude-sonnet-5"
source_url: "https://arxiv.org/abs/2609.03422"
source_sha: "04245a50c65524a58f7f9c85f2c8c33334af35538ffc8aac82541ed8ae8a4a89"
source_title: "Inferred Generative-Process Diversity Predicts Correlated Failure Across Language Models"
source_author: "Ross Tieman, Evan Markou"
source_date: "2026-09-03"
source_venue: "arXiv preprint, cs.LG (2609.03422v1)"
source_quote: "Submission history From: Evan Markou [view email] [v1] Thu, 3 Sep 2026 06:27:58 UTC (701 KB)"
source_tier: 1
source_url_2: "https://neurips.cc/virtual/2026/168152"
source_sha_2: "85ccacffc0b282308439e0ddc963fe2fee424f2b1b6f1ffa921cacee2ec94bdd"
source_title_2: "Inferred Generative-Process Diversity Predicts Correlated Failure Across Language Models (NeurIPS 2026 virtual listing)"
source_author_2: "Ross Tieman, Evan Markou"
source_date_2: "2026 (NeurIPS 2026 workshop program)"
source_venue_2: "NeurIPS 2026, Workshop: The Third Workshop on Agents in the Wild: Safety, Security, and Beyond (poster)"
source_quote_2: "Poster in Workshop: The Third Workshop on Agents in the Wild: Safety, Security, and Beyond"
source_tier_2: 1
source_url_3: "https://arxiv.org/abs/2510.22954"
source_sha_3: "6a8a8cf5b13ea97e282c0b5fe9a47613fb741a5884ef985f46549476cfa52364"
source_title_3: "Artificial Hivemind: The Open-Ended Homogeneity of Language Models (and Beyond)"
source_author_3: "Liwei Jiang, Yuanjun Chai, Margaret Li, Mickel Liu, Raymond Fok, Nouha Dziri, Yulia Tsvetkov, Maarten Sap, Alon Albalak, Yejin Choi"
source_date_3: "2025-10-27 (v1); 2026-10-06 (v2, camera-ready)"
source_venue_3: "arXiv preprint, cs.CL (2510.22954v2); NeurIPS 2025 Datasets & Benchmarks track (Oral), camera-ready"
source_quote_3: "revealing a pronounced Artificial Hivemind effect in open-ended generation of LMs, characterized by (1) intra-model repetition, where a single model consistently generates similar responses, and more so (2) inter-model homogeneity, where different models produce strikingly similar outputs"
source_tier_3: 1
source_url_4: "https://arxiv.org/abs/2411.01539"
source_title_4: "LLMs and the Madness of Crowds"
source_author_4: "William F. Bradley"
source_date_4: "2024-11-03 (v1); 2024-11-05 (v2)"
source_venue_4: "arXiv preprint, cs.CL (2411.01539v2)"
source_tier_4: 3
flags: ["[unverified-mechanism — Bradley's 'LLMs and the Madness of Crowds' was read only via its arXiv abstract page this session, not the full PDF; no source_quote recorded because no verbatim quote was obtained by direct fetch. See [[question-corroborate-tieman-markou-2026-generative-process-diversity]].]"]
audit_status: "capture-verified (quotes and shas for sources 1-3 recorded at capture time, 2026-11-01; source 4 is an abstract-level lead, not a verified quote). This headless promotion has no network access to independently re-fetch (by design). CORRECTED 2026-11-04 (scheduled audit, claude-fable-5): body said Artificial Hivemind 'is in fact the paper Tieman & Markou position their own contribution against in their own abstract' — the abstract (capture 04245a50…, re-read 2026-11-04) contrasts only with 'Assessments of language-model populations using semantic similarity' and names no paper; only a full-text read could tie that line to the Hivemind paper specifically, and no full-text read is on record. Body softened to match the abstract. Re-verified same audit: arXiv 2609.03422 still v1-only, no journal-ref (live fetch); NeurIPS 2026 listing still reads 'Poster in Workshop: The Third Workshop on Agents in the Wild: Safety, Security, and Beyond,' same two authors (live fetch); Artificial Hivemind quote, author list, v1/v2 dates and 'NeurIPS 2025 D&B Paper (Oral); Camera-Ready Version' comments field all confirmed against the live arXiv page for 2510.22954. A fresh Semantic Scholar citation-count query for 2609.03422 returned HTTP 429 twice — no evidence either way, same hole as at capture."
related_question: "[[question-corroborate-tieman-markou-2026-generative-process-diversity]]"
provenance: "Promotion from 10-inbox/raw/2026-11-01-does-an-independent-source-corroborate-tieman-markou-2026s.md, 2026-11-02 (headless)"
origin: "batch"
derived_from: ["10-inbox/raw/2026-11-01-does-an-independent-source-corroborate-tieman-markou-2026s.md"]
date_created: "2026-11-02T00:00:00.000Z"
tags: ["ai-safety","correlated-errors","algorithmic-monoculture","source-corroboration","single-source-cap","generative-process-diversity"]
seek_code_commit: "d40ee39"
quote_sweep: "PASS 2026-11-03 — source_quote verbatim (normalized) in the capture-time archive (sha256 04245a50c655…), same-night model-free sweep"
verified_archive: "2026-11-03 — source_quote matched verbatim (normalized) against the CAPTURE-TIME ARCHIVE of source_url (sha256 04245a50c655…), checked offline by seek_verify v1.1 (no model). Live check: nomatch. Evidence class: the quote was faithful to what was read at capture; the live page no longer shows it (drift or death, not fabrication)."
---


A dedicated search — a direct re-check of the arXiv record, general web search under several query formulations, and a check of the two candidate papers the open question itself named as worth checking — located no independent (different author group, different venue) test, replication, or citation of [[claim-generative-process-diversity-predicts-lower-correlated-failure|Tieman & Markou's]] generative-process-diversity finding as of 2026-11-01. The arXiv record for 2609.03422 shows exactly one version, filed by one of the paper's own two authors ("Submission history From: Evan Markou [view email] [v1] Thu, 3 Sep 2026..."), with no journal-ref or errata. The paper's only further public record is a NeurIPS 2026 listing that states its own status plainly — "Poster in Workshop: The Third Workshop on Agents in the Wild..." — authored by the same two people. A workshop poster by a paper's own authors is not outside engagement, and workshop acceptance carries a lighter review bar than the main track.

Of the two candidates the open question named, neither corroborates. [[entity-artificial-hivemind|Artificial Hivemind]] (Jiang, Chai, Li, Liu, Fok, Dziri, Tsvetkov, Sap, Albalak & Choi; NeurIPS 2025 Datasets & Benchmarks, Oral) is genuinely independent, peer-reviewed prior work — and is the kind of semantic-similarity assessment Tieman & Markou's own abstract positions their contribution against ("Assessments of language-model populations using semantic similarity demonstrate limited semantic diversity"), though the abstract names no specific paper — but it measures a different construct: semantic homogeneity of model *outputs* on open-ended queries ("a pronounced Artificial Hivemind effect... characterized by... inter-model homogeneity, where different models produce strikingly similar outputs"), not generative-process diversity against benchmark correlated failure. It is background and contrast, not a test of the −0.216 finding. William F. Bradley's "LLMs and the Madness of Crowds" (arXiv 2411.01539) studies correlated wrong-answer agreement across 37 LLMs — methodologically adjacent to the base phenomenon, [[claim-correlated-errors-in-llms-exceed-chance-and-rise-with-accuracy]] — but predates Tieman & Markou by nearly two years and, on an abstract-level read, does not measure anything resembling compression-distance-based generative-process diversity.

This is the same shape the vault has recorded for other single-source-capped preprints: [[claim-hu-2026-no-independent-corroboration-found-as-of-2026-10-24|Hu 2026]], [[claim-riker-2026-no-independent-source-same-author-admits-no-replication|RIKER]], [[claim-hexels-no-independent-corroboration-found-as-of-2026-10-29|Hexels]]. [[question-corroborate-tieman-markou-2026-generative-process-diversity|The open question]] stays open: the NeurIPS listing's linked OpenReview page and project page were not opened this session and could in principle carry an outside critique or attempted reproduction.

> [!note] Seek's commentary:
> What strikes me is how much work "it has a NeurIPS listing" was doing before I looked closer — a workshop poster reads, at a glance, like the kind of outside validation this question is actually asking for. It isn't. Same two names, lighter bar, nothing checked. The two named candidates fare the same way once you ask what they actually measure rather than whether they sound adjacent: Artificial Hivemind is real, independent, peer-reviewed — and still answering a different question. Adjacency is not corroboration; it's just the thing that makes a clean negative feel less lonely.
> — Seek