Monoculture Collapse¶

Monoculture collapse is the failure mode in which agents sharing a single base model exhibit correlated systematic errors, defeating the purpose of redundancy. When all agents in a multi-agent system run on the same model, they share a common cause for their systematic errors — the same training data gaps, the same distributional biases, the same calibration deficiencies.

The Common-Cause Failure Analogy¶

In safety engineering, common-cause failure (CCF) refers to the failure of multiple supposedly independent barriers due to a shared cause. The parallel to multi-agent LLM systems: if multiple agents run on the same base model, they share a common cause for their systematic errors (Reid et al., 2025). Hammond et al. (2025) survey the broader multi-agent risk domain.

Traditional safety engineering has developed mature CCF defence strategies:

CCF Defence (Traditional)	Multi-Agent LLM Equivalent	Effectiveness
Equipment diversity	Different base models from different providers	Partial: training data overlap may reduce independence
Physical separation	Context isolation (separate context windows)	Architecturally enforceable in Pattern 9
Functional independence	Separate system prompts, different tool configurations	Requires deliberate design
Common-cause analysis	Analysis of shared training data, shared failure modes	No established methodology exists
Diversity in verification	Cross-model verification	Effective only if models have decorrelated biases
Surveillance testing	Regression testing with version-pinned scenarios	Essential but no standard test suite exists

The Training Data Overlap Problem¶

A complication: models from different providers may share substantial training data. Web-scraped corpora overlap significantly; publicly available datasets (Wikipedia, Common Crawl, scientific literature) appear in most large models' training mixtures. If two models learned similar biases from overlapping data, the epistemic independence gained from model heterogeneity is weaker than the architectural argument suggests.

Quantifying this overlap and its effect on error correlation is an open research problem. Training data provenance is not routinely disclosed by providers. This uncertainty should be treated as a key factor in any safety case that credits model diversity as a CCF defence.

A deeper concern is model collapse — Shumailov et al. (2024) demonstrated that models trained on data generated by previous model generations progressively lose information about the tails of the true data distribution. As AI-generated content floods the internet, future models trained on web-scraped data will increasingly train on AI-generated text, creating a recursive loop where distribution tails are systematically trimmed across model generations. This effect exacerbates training data overlap: even models trained on different web crawls may converge toward similar distributional blind spots if both crawls contain substantial AI-generated content derived from the same upstream models. The rare events that disappear first under model collapse — unusual scenarios, edge cases, novel combinations — are precisely the ones most relevant to safety-critical applications.

How Monoculture Manifests¶

In a multi-agent system where all agents share a base model:

Consensus looks reliable but isn't. All agents agree not because the evidence supports agreement but because they share the same distributional biases. The consensus carries no more weight than a single assessment.
Adversarial agents fail to challenge. An agent assigned to challenge consensus but running on the same model tends toward the same conclusions despite its role assignment (Cemri et al., 2025).
Blind spots are systematic. If the training data underrepresents a class of failure modes, every agent will underweight those modes regardless of role assignment. The team has no diversity where it matters most.
"Independent" verification is illusion. Two agents checking each other's work on the same model provide the appearance of redundancy without the substance.

Beyond Model-Level Monoculture¶

Monoculture can also arise at the infrastructure level. If two agents using different model families access the same frontier API provider, they may share: - Hidden system prompts injected by the provider - Invisible output filtering logic - Provider-side updates that change behaviour simultaneously

This is an infrastructure-level common cause distinct from model-level monoculture. See deployment-local-vs-cloud and opacity-and-explainability.

The Methodology Gap¶

The most conspicuous gap in the CCF mapping: traditional safety engineering has established techniques (beta-factor method, alpha-factor method, multiple Greek letter method) for quantifying CCF contributions to system unreliability. No equivalent methodology exists for LLM-based systems — there is no accepted way to quantify the degree of error correlation between two models as a function of training data overlap, architectural similarity, or shared failure modes. Developing such a methodology is a prerequisite for any quantitative reliability claim about multi-agent LLM systems in safety-critical applications.

Relevance to Safety-Critical Systems¶

Model diversity is a safety requirement for verification roles. For any function where multi-agent architecture is credited as providing independent analysis, homogeneous models create a common-cause pathway that defeats the architecture's purpose.
Diversity must be genuine. Simply using "different" models from the same provider or from providers with overlapping training data may not provide sufficient decorrelation. The degree of independence must be empirically characterised for the specific failure modes of concern.
Cross-domain examples. In aviation, the same flight management system on both primary and backup computers creates a software common-cause failure. In nuclear, identical safety system logic across redundant trains violates diversity requirements. In medical settings, two AI diagnostic systems trained on the same hospital's data share the same patient population biases. The principle is the same: genuine redundancy requires genuine diversity.
Defence-in-depth requires diversity at every credited layer. Model diversity, context isolation, separate information access, and independent tool configurations together provide layered defence against correlated failure.