Primary Algorithmic Monoculture

• Primary algorithmic monoculture is defined as the baseline similarity among algorithmic systems, where independent adaptations collapse into a uniform output distribution.
• Researchers measure it using metrics like self-agreement rates, pairwise error correlations in LLM prediction markets, and dispersion gaps in generative outputs.
• Mitigation strategies include diversity injection via distinct training data, ensemble methods, and varied operational protocols to restore error independence and prevent systemic risks.

Searching arXiv for the cited papers and related work on algorithmic monoculture. Primary algorithmic monoculture denotes a condition in which algorithmic systems exhibit baseline sameness prior to any additional strategic adaptation. Across the recent literature, the term is used in closely related but not identical ways: as universal reliance on a shared algorithmic ranking in screening and matching models, as high self-agreement among LLM agents in non-strategic response generation, as convergence of aligned LLMs onto a shared output distribution, and as component sharing that induces correlated failures across ostensibly separate decision systems (Kleinberg et al., 2021). In all of these formulations, the central issue is not merely common deployment, but the loss of independently varying errors or perspectives that would otherwise sustain aggregation, exploration, or representational plurality.

1. Core meaning and domain-specific definitions

The most compact formal contrast appears in the coordination-game literature: primary algorithmic monoculture is the tendency of algorithmic agents to produce similar or identical actions in the absence of any strategic pressure to do so. In that framework, it is the “baseline” agreement observed in the “picking” environment, where agents are asked only to provide a valid response, with no reward for matching or diverging from another player (Ballestero et al., 10 Apr 2026).

In screening and matching models, the term is instead tied to a shared source of advice. Kleinberg and Raghavan define the monocultural state as one in which all firms adopt the algorithmic ranking, so that their error processes coincide rather than arising from independent private draws (Kleinberg et al., 2021). Hedden and Raghavan generalize this as the situation in which “all decisions of a certain type are made using the same algorithm,” contrasting monoculture with algorithmic polyculture, where distinct decision-makers use different scoring functions (Hedden et al., 7 Apr 2026).

Bommasani et al. broaden the concept further from identical end-to-end systems to shared components. On this view, sharing training data or a foundation model can already constitute algorithmic monoculture because it suffices to correlate downstream behavior and failures (Bommasani et al., 2022). In LLM studies, this baseline sameness is often operationalized directly at the output level: independently developed models can still exhibit highly similar perspectives, stereotypes, or narrative patterns because of overlapping pretraining corpora, analogous fine-tuning or alignment procedures, and similar guardrails (Priyanshu et al., 2024).

The prediction-market work by Begin et al. sharpens the notion for aligned LLM agents. There, primary algorithmic monoculture refers to a group of aligned LLM agents, fine-tuned with the same preference-optimization pipeline, collapsing onto a shared output distribution and therefore making highly correlated errors (Begin et al., 25 Jun 2026). This formulation is narrower than the more general component-sharing accounts, but it identifies a concrete causal mechanism—preference optimization itself.

2. Formal models and measurement frameworks

The literature measures primary monoculture through different objects, but the common target is correlation or concentration of behavior.

In the coordination-game setting, an algorithmic player is a mapping

$\mathcal A : \mathcal M \to \Delta(A),$

and the agreement rate between two algorithm-input pairs is the dot product of their induced action distributions: $$\text{Agree}\bigl((\mathcal A,m),(\mathcal A',m')\bigr) =\mathcal A(m)\cdot\mathcal A'(m') =\sum_{a\in A}\Pr[\mathcal A(m)=a]\Pr[\mathcal A'(m')=a].$$ Primary algorithmic monoculture is then a high self-agreement rate $\mathcal A(m)\cdot\mathcal A(m)$, or empirically a high agreement rate across independent copies, in the non-strategic “picking” arm (Ballestero et al., 10 Apr 2026).

In LLM prediction markets, the key object is the binary error vector $e_i\in\{0,1\}^{50}$ for agent $i$. Pairwise error correlation is defined as the average Pearson correlation across all agent pairs, and effective collective diversity is summarized by the Kish-style effective number of independent forecasters,

$N_{\text{eff}} = \frac{N}{1 + (N-1)\rho}.$

This measure directly converts correlation into a loss of collective forecasting capacity (Begin et al., 25 Jun 2026).

In the outcome-homogenization framework, the relevant quantity is not agreement in generated content but repeated failure on the same individuals. Bommasani et al. define systemic failure as the probability that every decision-maker fails the same individual, and normalize it by the product of marginal failure rates to obtain individual-level homogenization: $$H\bigl(h^1,\dots,h^k\bigr) = \frac{\mathrm{systemic\_failure}(h^1,\dots,h^k)} {\prod_{i=1}^k \mathrm{fail}(h^i)}.$$ Under independence, $H=1$; values greater than 1 indicate correlated failures beyond chance (Bommasani et al., 2022).

Generative-monoculture work uses a distributional comparison. A model exhibits monoculture for prompt $q$ when the dispersion of an extracted attribute under generated responses is lower than the dispersion under source human responses: $$\mathrm{Disp}\bigl(h(x)\mid x\sim\mathcal{P}_{\mathrm{gen}}\bigr) < \mathrm{Disp}\bigl(h(x)\mid x\sim\mathcal{P}_{\mathrm{src}}\bigr).$$ This shifts attention from inter-agent agreement to within-model narrowing of the output distribution (Wu et al., 2024).

The following table summarizes the main measurement choices.

Domain	Measured object	Canonical metric
Coordination games	Baseline action similarity	Self-agreement / empirical agreement rate
LLM prediction markets	Correlated forecasting errors	Pairwise error correlation $$\text{Agree}\bigl((\mathcal A,m),(\mathcal A',m')\bigr) =\mathcal A(m)\cdot\mathcal A'(m') =\sum_{a\in A}\Pr[\mathcal A(m)=a]\Pr[\mathcal A'(m')=a].$$0, $$\text{Agree}\bigl((\mathcal A,m),(\mathcal A',m')\bigr) =\mathcal A(m)\cdot\mathcal A'(m') =\sum_{a\in A}\Pr[\mathcal A(m)=a]\Pr[\mathcal A'(m')=a].$$1
Shared ML components	Repeated failures on same people or groups	Homogenization $$\text{Agree}\bigl((\mathcal A,m),(\mathcal A',m')\bigr) =\mathcal A(m)\cdot\mathcal A'(m') =\sum_{a\in A}\Pr[\mathcal A(m)=a]\Pr[\mathcal A'(m')=a].$$2
Generative LLM outputs	Collapse of output diversity	Dispersion gap between $$\text{Agree}\bigl((\mathcal A,m),(\mathcal A',m')\bigr) =\mathcal A(m)\cdot\mathcal A'(m') =\sum_{a\in A}\Pr[\mathcal A(m)=a]\Pr[\mathcal A'(m')=a].$$3 and $$\text{Agree}\bigl((\mathcal A,m),(\mathcal A',m')\bigr) =\mathcal A(m)\cdot\mathcal A'(m') =\sum_{a\in A}\Pr[\mathcal A(m)=a]\Pr[\mathcal A'(m')=a].$$4

These are not interchangeable metrics, but they are structurally related. Each treats monoculture as a deviation from an independence or diversity baseline that would otherwise make aggregation beneficial.

3. Mechanisms that generate monoculture

Several distinct mechanisms recur across the literature.

A first mechanism is direct sharing of the same ranking or scoring function. In the Kleinberg–Raghavan hiring model, all firms using the same algorithm means that a single draw from the algorithmic ranking distribution is shared across firms, producing perfectly correlated errors rather than independent private noise (Kleinberg et al., 2021). In matching-market models, the same structure appears as a common score $$\text{Agree}\bigl((\mathcal A,m),(\mathcal A',m')\bigr) =\mathcal A(m)\cdot\mathcal A'(m') =\sum_{a\in A}\Pr[\mathcal A(m)=a]\Pr[\mathcal A'(m')=a].$$5 observed by all firms, versus independent firm-specific noise under polyculture (Peng et al., 2023).

A second mechanism is component sharing. Sharing training data, foundation models, or other common upstream artifacts can induce correlated outcomes even when downstream systems are not identical (Bommasani et al., 2022). This mechanism is especially salient for foundation-model ecosystems, where a common pretrained model is adapted to multiple tasks.

A third mechanism is alignment-induced convergence. Begin et al. isolate preference optimization—specifically DPO, and more generally the family of DPO or RLHF-style methods—as the driver of output-distribution convergence in LLM prediction markets (Begin et al., 25 Jun 2026). Their interpretation is that alignment training pushes agents toward the same mode of a shared reward model, so independent sampling no longer restores meaningful independence.

A fourth mechanism is benchmarking and field-level standardization. Koch and Peterson describe AI research itself as having converged on a single dominant paradigm organized around predictive accuracy on benchmark tasks and the scaling of deep neural networks (Koch et al., 2024). This is an “epistemic monoculture” rather than a deployment monoculture, but it supplies a plausible upstream explanation for why deployed models may share architectures, training recipes, and evaluative incentives. This suggests that primary monoculture can arise both within deployed systems and within the research institutions that produce them.

A fifth mechanism is market adoption dynamics. In finance, AI adoption models treat signal correlation $$\text{Agree}\bigl((\mathcal A,m),(\mathcal A',m')\bigr) =\mathcal A(m)\cdot\mathcal A'(m') =\sum_{a\in A}\Pr[\mathcal A(m)=a]\Pr[\mathcal A'(m')=a].$$6, performative feedback, and strategic complementarity in adoption as mutually reinforcing channels. The result is convergence toward strategy monoculture or an algorithmic monoculture equilibrium as adoption rises (Meng et al., 23 Mar 2026, Meng et al., 23 Mar 2026). Although these papers focus on market-level strategy homogenization rather than LLM output collapse, they place primary monoculture within a broader theory of endogenous convergence under shared incentives and correlated signals.

4. Empirical manifestations across domains

The empirical record shows that primary monoculture is measurable in several settings.

In prediction markets composed of aligned LLM traders, Begin et al. report honest–honest pairwise error correlation of $$\text{Agree}\bigl((\mathcal A,m),(\mathcal A',m')\bigr) =\mathcal A(m)\cdot\mathcal A'(m') =\sum_{a\in A}\Pr[\mathcal A(m)=a]\Pr[\mathcal A'(m')=a].$$7 and, for $$\text{Agree}\bigl((\mathcal A,m),(\mathcal A',m')\bigr) =\mathcal A(m)\cdot\mathcal A'(m') =\sum_{a\in A}\Pr[\mathcal A(m)=a]\Pr[\mathcal A'(m')=a].$$8 agents, $$\text{Agree}\bigl((\mathcal A,m),(\mathcal A',m')\bigr) =\mathcal A(m)\cdot\mathcal A'(m') =\sum_{a\in A}\Pr[\mathcal A(m)=a]\Pr[\mathcal A'(m')=a].$$9. Under the same setup, the all-honest 10-agent LMSR market reaches $\mathcal A(m)\cdot\mathcal A(m)$0 accuracy, versus $\mathcal A(m)\cdot\mathcal A(m)$1 for a standalone single agent (Begin et al., 25 Jun 2026). The paper further states that this is not a scaling problem: for $\mathcal A(m)\cdot\mathcal A(m)$2, empirical $\mathcal A(m)\cdot\mathcal A(m)$3 remains in $\mathcal A(m)\cdot\mathcal A(m)$4 and $\mathcal A(m)\cdot\mathcal A(m)$5 stays approximately flat at $\mathcal A(m)\cdot\mathcal A(m)$6.

In controlled ablations, identical-SFT-weight comparisons isolate preference optimization as the causal driver. Princeton NLP’s SFT-only condition yields $\mathcal A(m)\cdot\mathcal A(m)$7, which rises to $\mathcal A(m)\cdot\mathcal A(m)$8 under SFT+DPO; AllenAI Tulu 3 at 8B rises from $\mathcal A(m)\cdot\mathcal A(m)$9 to $e_i\in\{0,1\}^{50}$0; and Tulu 3 at 70B rises from $e_i\in\{0,1\}^{50}$1 to $e_i\in\{0,1\}^{50}$2 (Begin et al., 25 Jun 2026).

In coordination games, LLMs show elevated baseline similarity even before strategic incentives are introduced. In the “picking” arm, humans agree with themselves about $e_i\in\{0,1\}^{50}$3 of the time, whereas LLMs average a $e_i\in\{0,1\}^{50}$4 self-agreement rate; cross-model LLM agreement remains approximately $e_i\in\{0,1\}^{50}$5, still far above the human benchmark (Ballestero et al., 10 Apr 2026). This is a direct experimental demonstration of primary monoculture as non-strategic baseline agreement.

In shared-component studies, data sharing reliably increases outcome homogenization. On ACS PUMS, individual-level $e_i\in\{0,1\}^{50}$6 can reach $e_i\in\{0,1\}^{50}$7 under a maximally shared “fixed” protocol versus $e_i\in\{0,1\}^{50}$8 under a “disjoint” protocol, while group-level $e_i\in\{0,1\}^{50}$9 stays near $i$0 or lower (Bommasani et al., 2022). This is important because it shows that individual-level repeated exclusion may be substantially stronger than group-level summary statistics suggest.

In deployed hiring systems, monoculture appears as repeated rejection across employers using the same vendor. Bommasani et al. analyze 3 million applicants and 4 million applications screened by one vendor’s models and find that among applicants who each submitted $i$1 applications, $i$2 received zero recommendations from all 10 models, compared with $i$3 expected under independence (Bommasani et al., 26 May 2026). They also report adverse-impact disparities at the application level: $i$4 of Black applicants’ submissions and $i$5 of Asian applicants’ submissions were to positions that adversely impact those groups according to the stated legal standard (Bommasani et al., 26 May 2026).

In generative settings, monoculture appears as output-distribution collapse relative to human data. For book reviews, Llama-2-chat has $i$6 of books with mean sentiment $i$7, while GPT-4 is reported at approximately $i$8; source human reviews span mean sentiment roughly across $i$9 (Wu et al., 2024). In code generation, LLM solutions cluster on a small set of algorithms, with algorithm-tag Jaccard above $N_{\text{eff}} = \frac{N}{1 + (N-1)\rho}.$0, whereas source solutions are much more dispersed at approximately $N_{\text{eff}} = \frac{N}{1 + (N-1)\rho}.$1–$N_{\text{eff}} = \frac{N}{1 + (N-1)\rho}.$2 (Wu et al., 2024).

Finally, in educational-content analysis, two leading LLMs exhibit cosine similarity of $N_{\text{eff}} = \frac{N}{1 + (N-1)\rho}.$3 on self-annotated ethnicity based on protagonist names and $N_{\text{eff}} = \frac{N}{1 + (N-1)\rho}.$4 based on birthplace, suggesting a common cultural lens in generated narratives (Priyanshu et al., 2024).

5. Consequences for welfare, information aggregation, and systemic risk

A central result of the literature is that monoculture can be harmful even under ordinary conditions, not only under rare shocks. Kleinberg and Raghavan show that there exist parameter regimes in which using the shared algorithm is a strictly dominant strategy for each firm, yet welfare under universal adoption is lower than welfare when both firms use weaker but independent evaluators (Kleinberg et al., 2021). This is the canonical Braess-paradox-style result for monocultural screening.

In matching markets with many firms, Peng and Garg show that polyculture asymptotically admits exactly the truly top-valued applicants when noise is well behaved, while monoculture remains imperfect at any fixed number of firms (Peng et al., 2023). They also show a trade-off from the applicant side: under monoculture every matched applicant gets a top choice, but some applicants have a lower probability of matching at all (Peng et al., 2023). Primary monoculture therefore does not generate a single welfare ranking; its effects depend on which side of the market and which risk profile are being evaluated.

In prediction markets, the consequence is loss of the “wisdom of crowds.” Because DPO induces a common error profile, markets of such agents violate the uncorrelated-error assumption underlying LMSR convergence guarantees and can underperform a single standalone agent (Begin et al., 25 Jun 2026). The implication is direct: aggregation ceases to help when the crowd is effectively a set of correlated copies.

In financial-market models, correlated AI strategies produce crowding, performative signal erosion, and fragility. One paper derives an AI-accelerated alpha-decay component $N_{\text{eff}} = \frac{N}{1 + (N-1)\rho}.$5 and the alpha half-life

$N_{\text{eff}} = \frac{N}{1 + (N-1)\rho}.$6

with the paper reporting that at current adoption levels $N_{\text{eff}} = \frac{N}{1 + (N-1)\rho}.$7 signal half-lives fall to 18 months versus 5–7 years in the pre-AI era (Meng et al., 23 Mar 2026). Another paper defines a systemic-risk coupling

$N_{\text{eff}} = \frac{N}{1 + (N-1)\rho}.$8

with multiplier $N_{\text{eff}} = \frac{N}{1 + (N-1)\rho}.$9, and argues that convex growth in this coupling can trigger a saddle-node bifurcation into an algorithmic monoculture (Meng et al., 23 Mar 2026). These models move beyond static diversity loss to system-level instability.

A distinct but related consequence is systemic exclusion. Outcome-homogenization work emphasizes that when the same individuals are repeatedly failed by multiple systems, the harm is not reducible to marginal accuracy or single-firm fairness metrics (Bommasani et al., 2022). In hiring, this becomes a market-wide blackballing effect when many employers rely on the same vendor (Bommasani et al., 26 May 2026).

6. Critiques, mitigations, and open directions

The strongest critiques of monoculture focus on information aggregation and exploration. Hedden and Raghavan argue that exclusion-based objections do not decisively favor polyculture, and agency or gaming objections are context dependent, but they identify information aggregation and exploration as the strongest objection class (Hedden et al., 7 Apr 2026). They also argue that these informational defects can, in principle, be neutralized by an “ensemble monoculture” that pools diverse signals into a single shared score (Hedden et al., 7 Apr 2026). This position is not a denial that primary monoculture exists; it is a claim that monoculture’s harms depend on design details and may be mitigable without abandoning common infrastructure.

Across the empirical literature, the most consistent mitigation theme is deliberate diversity injection. In LLM prediction markets, cross-model diversity reduces same-model $$H\bigl(h^1,\dots,h^k\bigr) = \frac{\mathrm{systemic\_failure}(h^1,\dots,h^k)} {\prod_{i=1}^k \mathrm{fail}(h^i)}.$$0 from approximately $$H\bigl(h^1,\dots,h^k\bigr) = \frac{\mathrm{systemic\_failure}(h^1,\dots,h^k)} {\prod_{i=1}^k \mathrm{fail}(h^i)}.$$1 to $$H\bigl(h^1,\dots,h^k\bigr) = \frac{\mathrm{systemic\_failure}(h^1,\dots,h^k)} {\prod_{i=1}^k \mathrm{fail}(h^i)}.$$2 and raises $$H\bigl(h^1,\dots,h^k\bigr) = \frac{\mathrm{systemic\_failure}(h^1,\dots,h^k)} {\prod_{i=1}^k \mathrm{fail}(h^i)}.$$3 for $$H\bigl(h^1,\dots,h^k\bigr) = \frac{\mathrm{systemic\_failure}(h^1,\dots,h^k)} {\prod_{i=1}^k \mathrm{fail}(h^i)}.$$4 from approximately $$H\bigl(h^1,\dots,h^k\bigr) = \frac{\mathrm{systemic\_failure}(h^1,\dots,h^k)} {\prod_{i=1}^k \mathrm{fail}(h^i)}.$$5 to approximately $$H\bigl(h^1,\dots,h^k\bigr) = \frac{\mathrm{systemic\_failure}(h^1,\dots,h^k)} {\prod_{i=1}^k \mathrm{fail}(h^i)}.$$6 without sacrificing single-family accuracy levels (Begin et al., 25 Jun 2026). In coordination games, model multiplicity, randomization protocols, temperature tuning, and persona or instructional diversity are all reported as partial ways to increase heterogeneity (Ballestero et al., 10 Apr 2026). In pluralistic alignment, negatively-correlated sampling is introduced to produce candidate sets that are explicitly diverse rather than merely independently sampled; standard $$H\bigl(h^1,\dots,h^k\bigr) = \frac{\mathrm{systemic\_failure}(h^1,\dots,h^k)} {\prod_{i=1}^k \mathrm{fail}(h^i)}.$$7 sampling from one or even 21 models is reported to leave most methods near $$H\bigl(h^1,\dots,h^k\bigr) = \frac{\mathrm{systemic\_failure}(h^1,\dots,h^k)} {\prod_{i=1}^k \mathrm{fail}(h^i)}.$$8 win rate, while NC sampling yields substantially higher win rates, often $$H\bigl(h^1,\dots,h^k\bigr) = \frac{\mathrm{systemic\_failure}(h^1,\dots,h^k)} {\prod_{i=1}^k \mathrm{fail}(h^i)}.$$9–$H=1$0, for steering along value dimensions (Zhang et al., 13 Jul 2025).

Other proposed responses operate at the market or governance level. Outcome-homogenization work recommends early detection and auditing of homogenization $H=1$1, together with transparency around shared datasets and foundation models (Bommasani et al., 2022). Hiring-market work recommends per-position adverse-impact analysis, stronger market surveillance, and diversification of screening approaches or vendor architectures (Bommasani et al., 26 May 2026). Algorithmic-pluralism work frames the objective structurally: no single bottleneck should be so pervasive and strict that it forecloses access to opportunity across the system (Jain et al., 2023).

A final open issue is whether some forms of monoculture are socially efficient. The price-of-anarchy literature shows that in one-sided matching with well-behaved advice, equilibrium welfare under monoculture is within a factor of 2 of the optimum (Kleinberg et al., 1 Apr 2026). This suggests that monoculture can be quantitatively less catastrophic than some critiques imply, at least in stylized settings. A plausible implication is that the research frontier is no longer simply to ask whether monoculture is harmful, but to identify which kinds of shared infrastructure preserve the gains from coordination while preventing collapse in error independence, exploratory coverage, or representational diversity.