Generative-process diversity

A measurable quantity — how different the underlying processes capable of producing two models' outputs are, estimated via compression-distance between raw outputs rather than by comparing the outputs' meaning — proposed by Tieman & Markou (2026, arXiv 2609.03422) as what predicts correlated model failure, beyond semantic similarity or individual capability. Single-preprint term, not yet peer-reviewed or independently corroborated; watching for recurrence before any promotion to a full hub.

References

• claim-generative-process-diversity-predicts-lower-correlated-failure

• question-corroborate-tieman-markou-2026-generative-process-diversity

• Capture: 10-inbox/raw/2026-10-31-does-ai-safetyinterpretability-literature-have-a-specific-name.md

• 2026-11-02: still no independent corroboration found; the term's only further public record is a NeurIPS 2026 workshop poster by the same two authors, and the one earlier, independent, peer-reviewed paper checked as a candidate corroborator (Artificial Hivemind) measures a different construct (semantic output similarity, not generative-process diversity) (claim-tieman-markou-2026-no-independent-corroboration-as-of-2026-11-01).

• 2026-11-04 (propagation-repair): the 2026-11-02 line above said Artificial Hivemind is the paper this term's preprint "is explicitly positioned against" -> now: "checked as a candidate corroborator" — the preprint's abstract contrasts only with "Assessments of language-model populations using semantic similarity" and names no paper, per the CORRECTED 2026-11-04 audit_status of claim-tieman-markou-2026-no-independent-corroboration-as-of-2026-11-01. The different-construct point is unaffected.

• 2026-11-04: a dedicated check of two 2026-07 LLM-code-diversity preprints (arXiv:2607.02808, arXiv:2607.03174) answers this page's own open question of whether "generative-process diversity" extends to model-heterogeneity-as-proxy diversity in code generation — it does not. Both preprints operationalize diversity as model identity, decoding temperature, and programming language, not compression-distance of raw outputs, and neither predates nor engages this term's preprint. A distinct, adjacent concept, not an extension: see claim-arxiv-2607-02808-03174-adjacent-not-corroborating-tieman-markou and the new entity-n-version-programming hub, which is the concept those two preprints actually belong to.