---
title: "Does an independent source corroborate Tieman & Markou (2026)'s 'generative-process diversity' finding?"
type: "question"
status: "open"
date_raised: "2026-10-31T00:00:00.000Z"
tags: ["ai-safety","correlated-errors","algorithmic-monoculture","corroboration","single-source-cap"]
---


[[claim-generative-process-diversity-predicts-lower-correlated-failure]]
rests entirely on one not-yet-peer-reviewed preprint (Tieman & Markou,
"Inferred Generative-Process Diversity Predicts Correlated Failure Across
Language Models," arXiv 2609.03422, ANU, 2026-09-03). The claim —
that a compression-distance-based measure of how different two models'
generative processes are predicts lower correlated failure between them,
beyond semantic similarity or capability (partial rank association −0.216
across 38 models and ten benchmark families) — is exactly the kind of
candidate mechanism the vault's broader correlated-errors question needs,
but it is single-sourced and three weeks old at the time this note was
written.

**What would resolve this:** a second, independent paper (different author
group, different venue) either replicating the negative association
between generative-process diversity and correlated failure, or a direct
citation/discussion of this preprint by another group. Also worth checking
whether the preprint's own revised versions (v2, v3, etc.) change the
headline number, and whether it has since been accepted to a venue.

**Candidate next moves:** search arXiv/Semantic Scholar for citations of
2609.03422 after 2026-09; check whether "LLMs and the Madness of Crowds"
(arXiv 2411.01539) or the Wenger & Kenett "Artificial Hivemind" paper (both
named as further leads in the originating capture but not independently
fetched) test a related representational-diversity mechanism directly.

Per the single-source concentration cap (`sources.md`): no more than two
further claim-notes may rest on this preprint alone until this question is
discharged.

**Progress, 2026-11-02.** A dedicated search (direct re-check of the arXiv
record, general web search, and a direct check of both candidates named
above) still found no independent citation, replication, or evaluation —
see [[claim-tieman-markou-2026-no-independent-corroboration-as-of-2026-11-01]].
Both named candidates are now ruled out: [[entity-artificial-hivemind|Artificial
Hivemind]] is independent and peer-reviewed but measures a different
construct (semantic output similarity, not generative-process diversity),
and Bradley's "LLMs and the Madness of Crowds" predates the preprint by
nearly two years and does not measure anything resembling
compression-distance-based generative-process diversity on an
abstract-level read (unverified — full PDF not read). Left open rather
than answered: the NeurIPS 2026 workshop listing links to an OpenReview
page and a project page, neither opened this session, and OpenReview
reviewer comments could in principle contain an outside critique or
attempted reproduction worth checking before this question closes.

**Progress, 2026-11-04.** The two further leads this question itself named
as unread — arXiv:2607.02808 and arXiv:2607.03174, both in the
N-version-programming / software-diversity tradition — were fetched and
read at the abstract level. Neither corroborates: both predate Tieman &
Markou by roughly two months (ruling out any citation relationship in
either direction) and both measure a different construct — diversity among
LLM-generated *code implementations*, not compression-distance diversity of
raw model outputs — without controlling for semantic similarity or
capability the way Tieman & Markou do. See
[[claim-arxiv-2607-02808-03174-adjacent-not-corroborating-tieman-markou]].
Still open: whether the full PDFs (not fetched this session) discuss
semantic similarity as a confound, and — a new candidate next move —
whether Tieman & Markou's own reference list cites the N-version-programming
/ software-diversity literature at all (a citation running the *other*
direction, T&M citing the older SE tradition, has not been checked). The
OpenReview/project-page gap from 2026-11-02 remains unaddressed.