Minority-Tendency in Collective Dynamics

• Minority-tendency is a framework describing asymmetry when underrepresented groups interact with majority forces across social, decision-making, and computational domains.
• It reveals phenomena such as phase transitions, bias amplification, suboptimal detection, and structural mismatches in settings from collective choice to imbalanced classification.
• The paradigm informs strategies in network analysis, machine learning oversampling, and group dynamics, highlighting trade-offs between minority visibility and optimization.

Minority-tendency is an umbrella designation for several asymmetry phenomena that arise when a minority position—numerical, strategic, structural, or representational—interacts with majority-driven dynamics. In the literature considered here, it does not denote a single formal theory. Rather, it names a family of problems in which minority-seeking agents, minority classes, minority communities, or socially subordinate groups exhibit distinctive behavior under majority rule, class imbalance, homophily, punishment, preference aggregation, or machine-learning optimization. Across these settings, recurring outcomes include phase transitions near critical minority fractions, frustration under incompatible incentives, systematic under-detection of small communities, bias amplification toward majority structure, and flattening of minority representation in generated language (Holme et al., 2016, Bachoc et al., 19 May 2025, Lee et al., 2024).

1. Conceptual scope and formalizations

The term is used heterogeneously across domains. In collective-decision models, the minority may be an agent type that seeks to end in the minority; in public-goods models, it may be the locally less represented strategy within a group; in imbalanced learning, it is the underrepresented class; in social-psychological analyses of LLMs, it denotes socially subordinate groups; and in network science it refers to small-scale communities whose total population is substantially smaller than that of larger communities.

Domain	Minority unit	Formalization in the literature
Binary collective choice	Minority seekers	Fraction $\phi$ of agents wants to end in the minority
Public goods game	Local strategy minority	In a group of 5, the less represented strategy is the minority
Imbalanced classification	Minority class	Classification categories are not approximately equally represented
LLM social representation	Socially subordinate groups	African, Asian, Hispanic Americans and women are treated as subordinate relative to White Americans and men
Network community detection	Minority communities	Total node fraction $\rho$ satisfies $\rho < q_s/q$

This heterogeneity matters because the operative mechanism changes with the formalization. In some models the minority is defined by preference over outcomes; in others by class frequency, network size, or social status. A plausible implication is that “minority-tendency” is best understood not as a single measurable trait but as a recurrent asymmetry pattern that emerges whenever local update rules, information constraints, or optimization objectives are coupled to unequal group sizes or unequal social positions (Yang et al., 2018, Chawla et al., 2011, Lee et al., 2024, Li et al., 20 Apr 2026).

2. Strategic minorities in collective decision dynamics

A central formalization appears in the binary-decision model of a population of $N$ agents, each making an irreversible choice $+1$ or $-1$, where a fraction $\phi$ are minority seekers and the remaining $1-\phi$ are majority seekers. Agents accumulate a state variable $s_i$ through pairwise interactions,

$s_i(t+1)=s_i(t)+\mathrm{sgn}\,s_j(t),$

and finalize when $\rho$0 exceeds a threshold $\rho$1; minority seekers then flip the sign of $\rho$2, while majority seekers do not. When $\rho$3, the system self-organizes efficiently: majority seekers end in the majority, minority seekers in the minority, $\rho$4, and for initial bias $\rho$5 the average opinion satisfies $\rho$6. When $\rho$7, frustration appears, the population tends toward approximately equal-sized opinion groups, $\rho$8 becomes negative, and just above $\rho$9 the satisfaction $\rho < q_s/q$0 drops abruptly from near $\rho < q_s/q$1 to close to $\rho < q_s/q$2. The optimal satisfaction bound is

$\rho < q_s/q$3

For $\rho < q_s/q$4, the paper further states that no Nash equilibrium exists (Holme et al., 2016).

A related but distinct formulation is the Galam Majority Model, which studies repeated local discussions in groups of size $\rho < q_s/q$5 with parameters $\rho < q_s/q$6: initial support for opinion A, tie prejudice in favor of A, and contrarian fraction. For odd groups, $\rho < q_s/q$7 is irrelevant; for even groups, prejudice can lower the tipping point far below $\rho < q_s/q$8. For example, with $\rho < q_s/q$9 and $N$0, opinion A wins if $N$1, so a minority with only $N$2 support can become majority. Increasing $N$3 shrinks the two-attractor regime and replaces it with a single-attractor regime in which the final outcome is predetermined by $N$4 rather than by initial support; for very high $N$5 an alternating regime appears (Galam, 25 Feb 2025).

These models converge on a common point: local majority rules do not guarantee globally “democratic” outcomes. This suggests that minority-tendency in collective choice is not exhausted by simple contrarianism. It also includes threshold effects, hidden asymmetries in tie-breaking, and incompatibility between private incentives and collective feasibility.

3. Network topology, media fields, and minority visibility

Network structure strongly conditions whether minorities can persist, expand, or even be detected. Under majority dynamics on a graph, a profile is stable if every agent agrees with at least half of its neighbors. The robust topologies—those in which an initial majority is never subverted in any reachable stable profile—are extremely limited: a graph must have no edge, be a clique, or be very close to a clique. Outside this small class, every graph admits at least one initial profile in which the majority dynamics can subvert the initial majority. The same work shows that deciding whether a graph admits a minority that becomes majority is NP-hard when the minority size is at most $N$6 of the social network size (Auletta et al., 2014).

In an Axelrod-style globalization model with endogenous mass media, minority growth becomes possible even under a global field equal to the statistical mode of the population. The model identifies four phases using the order parameters $N$7 and $N$8: ordered, semi-ordered, disordered, and chimera-like. The distinctive “minority-growth” phase satisfies $N$9 with both $+1$0 and $+1$1 near $+1$2, meaning that the second-largest domain, in a state different from the mass media trend, grows to almost half of the population. This phase requires sufficient long-range connectivity: on Erdös–Rényi networks it appears for $+1$3, and on rings with long-range connections for $+1$4 (Cosenza et al., 2019).

A different network asymmetry is the “chasm effect” in homophilous growing networks. There, minority representation is non-monotonic across the hierarchy: it first increases as one ascends from lower levels, then decreases in the tail. The sufficient and necessary condition reported for this effect is the combination of selective homophily on the rich-get-richer mechanism with the presence of an equal-chance mechanism. General homophily without that interaction reproduces the glass ceiling but not the chasm (Zhang et al., 2021).

Minority visibility also appears as an inference problem. In a stochastic block model with minority communities of total fraction $+1$5, three phases are identified: detectable, distinguishable, and resolvable. The transitions occur at the Kesten-Stigum threshold and two additional thresholds,

$+1$6

corresponding respectively to global detectability, collective distinguishability of minority communities, and full resolution of each minority community. The same analysis finds that spectral clustering with the Bethe Hessian is significantly weaker than belief propagation for minority-community recovery, even when macroscopic structure is recoverable (Li et al., 20 Apr 2026).

4. Minority classes in imbalanced machine learning

In statistical learning, minority-tendency often denotes the degradation of classifier performance caused by class imbalance. The classic response is oversampling or under-sampling, but the decisive issue is how minority support is expanded. SMOTE constructs synthetic minority examples rather than replicating existing ones. For each minority sample, it finds $+1$7 nearest minority neighbors, chooses a neighbor, and interpolates feature-wise along the segment between them:

$+1$8

The reported result is that combining synthetic minority over-sampling with majority under-sampling achieves better classifier performance in ROC space than under-sampling alone, and also outperforms varying loss ratios in Ripper or class priors in Naive Bayes on almost all datasets considered (Chawla et al., 2011).

Subsequent work refines the same idea by using local structure. A three-step method first heavily oversamples with SMOTE, then clusters the original and synthetic minority data with a Gaussian Mixture Model, and finally selects synthetic samples according to cluster weights determined by the distribution of majority samples. The weight

$+1$9

suppresses clusters contaminated by majority points. On 27 binary-class imbalanced datasets, the method is reported to outperform baseline, ROS, SMOTE, BSMOTE1, BSMOTE2, SVMSMOTE, and ADASYN in 22/27 tasks by $-1$0 and 16/27 by $-1$1 (Tripathi et al., 2021).

For long-tailed visual recognition, Context-rich Minority Oversampling (CMO) shifts the design from interpolation to context recombination. Minority foreground objects are pasted onto majority-class background images using a CutMix-style construction,

$-1$2

with background images sampled from the original distribution $-1$3 and foreground images from a minor-class-weighted distribution

$-1$4

The reported finding is that majority context can improve minority generalization; if both foreground and background are drawn from the minority-oversampled distribution, effectiveness collapses (Park et al., 2021).

A more generative response is Annealing Genetic GAN, which reframes GAN training as an evolutionary process with simulated annealing. Multiple offspring generators are produced under different objectives, the best offspring is evaluated, and the Metropolis criterion

$-1$5

permits occasional acceptance of worse solutions to escape local optima. The paper states that, under its Markovian and monotonicity conditions, the search process reaches an optimum with probability $-1$6 and yields better minority-class distribution coverage than standard GAN variants under scarcity (Hao et al., 2020).

Optimization itself can amplify majority structure. In a majority-minority learning framework with loss decomposition $-1$7, the “stereotypical predictor” minimizes only the majority loss $-1$8. The paper shows that a critical point of $-1$9 lies close to a critical point of the full loss,

$\phi$0

and that the majority-adverse region is confined to neighborhoods of radius $\phi$1 around critical points of $\phi$2. It also gives a lower bound on the additional training required to move from stereotypical to representative solutions,

$\phi$3

which formalizes why global accuracy can improve long before minority accuracy does (Bachoc et al., 19 May 2025).

Not all minority inclusion effects are minority-only. The MIME result shows that including some minority samples in the training set can improve test performance on the majority group itself. Theoretical existence proofs are given for perceptron settings with and without domain gap, and experiments on six datasets show a non-monotonic majority-accuracy curve whose peak occurs at a nonzero minority ratio on the realistic datasets tested, while a high-domain-gap negative control does not exhibit the effect (Chari et al., 2022).

5. Representation, communication, and AI systems

Minority-tendency also appears as a representational bias rather than a prediction error. In one study of ChatGPT, 500 short texts were generated for each of eight intersectional U.S. groups, yielding 52,000 texts in total. Homogeneity was operationalized as average pairwise cosine similarity between sentence embeddings derived from the second-to-last layer of BERT-base-uncased, with 124,750 pairwise similarities per group. Texts about African Americans, Asian Americans, and Hispanic Americans were more homogeneous than texts about White Americans, with standardized differences of $\phi$4, $\phi$5, and $\phi$6 SD, respectively; women were also portrayed as slightly more homogeneous than men by $\phi$7 SD. The effect persisted across embedding models and after topical controls (Lee et al., 2024).

A closely related issue arises in mental-health NLP. Traditional psychological stress models trained on general-population data underperform on minority stress detection for sexual and gender minorities. The reported remedy is emotion infusion: among the tested architectures, the classical multi-task model using stress and emotion objectives performs best on minority stress detection, and the Multi + MentalRoBERTa configuration reaches $\phi$8 and Accuracy $\phi$9 on MStress without training on minority stress data. The paper attributes the gain to distinct emotion distributions in minority communities and to the robustness of multi-task learning in low-data environments (Ivey et al., 2023).

Minority status also affects collaborative discourse. In collaborative problem-solving, Group Communication Analysis measures Participation, Internal Cohesion, Social Impact, Overall Responsivity, and Newness. Mixed-effects models show that group composition is more consequential than individual status for most sociocognitive features: Asian Majority groups had lower Internal Cohesion, Social Impact, and Overall Responsivity than several other group types, while more demographically diverse groups displayed more productive communication behaviors (Cavazos et al., 2024).

Intervention studies with LLMs extend this line from diagnosis to design. In power-imbalanced four-person groups with three Seniors and one Junior, AI-generated counterarguments and AI-mediated anonymous messages have different consequences. AI-generated counterarguments improved satisfaction and fostered a more flexible atmosphere, whereas AI-mediated messages increased minority participation but reduced psychological safety for Juniors; the reported coefficient for the psychological-safety drop is $1-\phi$0, $1-\phi$1, $1-\phi$2. The study therefore identifies a support paradox between participation and psychological safety (Lee et al., 30 Jun 2026).

6. Institutional selection, aggregation, and normative enforcement

In some settings minority-tendency is produced not by information scarcity alone but by rules that reward conformity. In a spatial public goods game on groups of size 5, the majority strategy in each group punishes the minority strategy with fine $1-\phi$3, and the cost is shared by the majority members. For no punishment or small $1-\phi$4, the cooperator fraction $1-\phi$5 increases continuously with the multiplication factor $1-\phi$6. For large $1-\phi$7, the transition becomes discontinuous; with $1-\phi$8, the paper reports a critical value $1-\phi$9 at which $s_i$0 jumps from $s_i$1 to $s_i$2, and the coexistence phase disappears for $s_i$3. Here minority status itself, regardless of whether it corresponds to cooperation or defection, becomes locally costly (Yang et al., 2018).

Preference aggregation can underrepresent minorities even when all reports are truthful. In the two-group analysis of the Moral Machine averaging mechanism, group A is the majority of size $s_i$4 and group D is the minority of size $s_i$5, with homogeneous within-group preference vectors. The aggregate vector is

$s_i$6

The paper shows that, in contested cases, the fraction of disputes in which the minority prevails is less than proportionate in $s_i$7, and the sub-proportionality worsens as disagreement between groups increases. Under strategic reporting, pure strategy equilibria do not always exist; whenever one exists, the majority prevails $s_i$8 of the time. Alternatives discussed include randomized dictatorship and median-based mechanisms (Feffer et al., 2023).

Selection procedures can similarly backfire. In a sequential search-and-evaluation model of hiring, “soft” affirmative action policies raise the relative proportion of minority candidates in the pool, but minority candidates are evaluated with noisier signals. Under both myopic and forward-looking rules, increasing the minority inflow can reduce the ex-ante probability of hiring a minority candidate, even when minorities are at least as qualified and as valuable as non-minorities. The mechanism is premature abandonment: noisier evaluation makes continued search more attractive than continuing to assess harder-to-evaluate minority candidates (Fershtman et al., 2020).

Not every minority role is suppressive or disadvantaged. In repeated public-goods games with group selection, indirect reciprocity, and imperfect observation, a minority of norm internalizers can remain numerically small while amplifying cooperation. The reported mechanisms are polarization, catalysis, and stabilization: norm internalizers increase between-group variance, trigger transitions out of low-cooperation troughs, and prolong high-cooperation periods, even though they often constitute only a minority of the population themselves (Odouard et al., 2021).

Taken together, these results show that minority-tendency is best treated as a structural property of update rules, information channels, and aggregation procedures rather than as a property of minority actors alone. A minority can be suppressed, merged, flattened, misdetected, over-penalized, or strategically disadvantaged; but under other rules it can catalyze cooperation, improve majority-group generalization, or grow to parity with the majority. The unifying technical question is therefore not whether minorities matter, but which local mechanisms determine when minority positions are erased, when they are merely visible, and when they become system-shaping.