Generative Stratification: Dynamic Hierarchy

• Generative stratification is the process where dynamic rules, model-based synthesis, or learned partitions generate stratified outcomes rather than relying on static groupings.
• It underpins diverse applications, from reproducing social inequality via occupational diffusion and intergenerational feedback to enhancing precision in experimental designs and adaptive randomization.
• It extends to deep generative modeling by capturing data on stratified spaces—unions of manifolds with varying dimensions—to better represent complex real-world structures.

Across recent literature, generative stratification designates a family of ideas in which stratified outcomes are produced, reproduced, or operationalized by dynamic rules, model-based synthesis, or learned partitions. In sociology and labor studies, the term is used for inequality reproduced through the structured diffusion of occupational requirements, intergenerational feedback loops linking wealth and institutional selection, and unequal access to AI-exposed work (Cantillan et al., 24 Feb 2026, Acharya et al., 2022, Mishra, 11 Jun 2026). In experimental design and simulation, it denotes procedures that use LLMs, pilot data, or informative Gaussian directions to construct strata before randomization or sampling [(Gui et al., 30 Sep 2025); (Tabord-Meehan, 2018); (Jourdain et al., 2010)]. A related machine-learning usage concerns deep generative modeling on stratified spaces, where the object of learning is a union of manifolds of varying dimensions rather than a single manifold (Martinez et al., 12 Apr 2026). Taken together, these usages suggest a common analytic motif: hierarchy or partition is not treated as a static backdrop, but as something generated, updated, or exploited by repeated local mechanisms.

1. Conceptual range

In the strongest sociological usage, generative stratification shifts attention away from the static distribution of persons across positions and toward the rules by which positions themselves change. One paper states that occupational hierarchy may persist not only because workers are sorted into unequal positions, but because the very process through which jobs are updated reproduces the hierarchy; another frames the core question as whether wealth differences are endogenously reproduced, amplified, or eliminated across generations through a repeated feedback loop linking parental wealth, observed signals, admission or hiring, and next-generation wealth (Cantillan et al., 24 Feb 2026, Acharya et al., 2022). In this sense, stratification is generative because inequality is continuously produced by directional diffusion, institutional thresholding, or recursive wealth updates rather than merely observed at equilibrium.

A second usage concerns the privatization and automation of social capacities by generative AI. Here the central object is not occupational skill diffusion or intergenerational selection, but what one paper calls social doing: “what we exert to build a social connection with another.” The paper argues that generative models automate and privatize the production of this social fabric, producing Synthetic Sociality, “a social reality in part fabricated by Silicon Valley’s privately owned and undemocratically governed generative models” (Dodik et al., 13 May 2026). This usage places generative stratification in a political-economic register of ownership, dependence, and unequal governance.

A third usage is procedural rather than substantive. In experimental design, Generative Stratification is an LLM-based method that predicts unit-level potential outcomes from high-dimensional pretreatment covariates and uses their sum as a prognostic score for blocking before randomization (Gui et al., 30 Sep 2025). In adaptive randomization, strata are learned from pilot data through stratification trees that jointly choose covariates, split points, and stratum-specific treatment probabilities (Tabord-Meehan, 2018). In Monte Carlo, multiple directions of stratification generate informative partitions of Gaussian latent space and allocate samples across them to reduce estimator variance (Jourdain et al., 2010). A plausible implication is that the phrase now spans both theories of inequality reproduction and technical methods that generate stratified designs.

2. Occupational change as a generative mechanism of hierarchy

"Structurally Conditioned Diffusion Reproduces Skills-Based Stratification" defines skills-based stratification as inequality organized through the distribution and structure of occupational skill portfolios, and introduces Asymmetric Trajectory Channeling (ATC) as the directional diffusion rule governing its reproduction (Cantillan et al., 24 Feb 2026). Using O*NET 2015–2024, the study analyzes 17.3 million directed diffusion opportunities linking 873 occupations and 161 skills, knowledge, and ability items. A directed diffusion opportunity is a source-target-skill triple $(i,j,s)$ such that source occupation $i$ is specialized in skill $s$ at baseline while target occupation $j$ is not. Specialization is defined by Revealed Comparative Advantage, with occupation $j$ specialized in skill $s$ if $\mathrm{RCA}(j,s)\ge 1$.

The theoretical baseline is a gravity-style conception of propagation,

$$P_{ijk} \propto \frac{M_i \cdot M_j \cdot S_k(\mathbf{x}_k)}{D_{ij}^{\gamma}},$$

where $M_i$ is source emission capacity, $M_j$ is target absorptive capacity, $i$0 is intrinsic diffusibility, and $i$1 is structural distance. The extension is directional. Hierarchy is represented by predetermined 2015 occupational wages, with signed wage gap

$i$2

or equivalently

$i$3

Upward and downward components are

$i$4

The main empirical specification is a complementary log-log gravity-hazard model with source, target, and skill fixed effects, directional friction parameters $i$5 and $i$6, an upward-boundary effect $i$7, and a distance-decay coefficient $i$8. Identification comes from within-occupation variation in propagation direction, and the results are robust to destination-side and origin-side specifications, placebo tests, domain-label permutation tests, within-stratum permutation tests, alternative RCA thresholds, alternative structural-distance measures, and sub-period analyses.

The core empirical pattern is domain-specific directional asymmetry. Socio-cognitive requirements propagate upward more often than downward, 20.7% vs. 14.9%, while sensory/physical requirements propagate downward more often than upward, 19.5% vs. 10.3%. A one-standard-deviation increase in composite structural distance reduces physical adoption by approximately 70% more than cognitive adoption. In model-based results, sensory/physical requirements show pronounced downward facilitation, with $i$9 in the source fixed-effects specification and about $s$0 in the destination fixed-effects specification; because $s$1, more negative $s$2 implies stronger downward facilitation. The upward component for physical requirements is statistically indistinguishable from zero.

ATC is generated by two mechanisms. Directional incorporation asymmetry means that wage gradients create distinct receiving environments: upward-moving socio-cognitive requirements encounter complementary routines, tools, credentialing systems, and evaluative standards, whereas upward-moving sensory or physical requirements encounter structural indifference. Structural portability constraint means that a requirement’s dependency position governs portability because skills anchoring long prerequisite chains impose larger co-adoption burdens. Nestedness intensifies these asymmetries in opposite directions. Skills are grouped into SC_Scaffolding, SC_Specialized, and Physical_Terminal archetypes according to domain and contribution to system-wide nestedness $s$3, computed as the standardized change in NODF when a skill is removed from the baseline occupation-skill matrix. For socio-cognitive requirements, the upward coefficient rises from 0.19 to 0.58 as one moves from low-nestedness socio-cognitive skills to high-nestedness scaffolding skills; for Physical_Terminal skills, Panel A reports $s$4 and $s$5, implying strong upward penalty and mild downward facilitation. The paper’s broader claim is that hierarchy can be reproduced without assortative preferences, discriminatory intent, or coordinated action: high-wage occupations accumulate socio-cognitive complexity, while lower-wage occupations remain more tied to sensory or physical requirements because the diffusion channel itself is directional.

3. Intergenerational feedback, fixed points, and persistent inequality

"Wealth Dynamics Over Generations: Analysis and Interventions" gives a stylized intergenerational model of generative stratification in which two populations have the same talent distribution but may converge to different long-run wealth levels because wealth contaminates the observed signal used by a university or employer (Acharya et al., 2022). Talent and wealth are Gaussian,

$s$6

with talent fixed across groups and over time, while mean wealth $s$7 evolves. The simplest signal is

$s$8

and the more general specification is

$s$9

The institution’s objective is

$j$0

and it admits or hires iff

$j$1

Because of Gaussian linearity, the decision rule reduces to a signal cutoff. The dynamic state is the group’s mean wealth $j$2, and the main update rule is

$j$3

Under the baseline model and with $j$4, the induced map is

$j$5

This is the formal mechanism: initial wealth advantage raises current mean wealth, higher $j$6 shifts the signal distribution upward, the institution uses the signal to estimate the objective, higher-wealth groups are admitted or hired at a higher rate, and admission or hiring determines next-generation wealth.

The model distinguishes two regimes. If

$j$7

then the process has a single attracting fixed point, and wealth inequality is transitory. A sufficient condition is

$j$8

If there exists $j$9 such that $j$0, the map is not globally contractive. The paper shows that there can be at most three fixed points, corresponding to a low attracting fixed point, a middle unstable fixed point, and a high attracting fixed point. Then history matters: groups starting below the unstable threshold converge to the low-wealth equilibrium, and groups starting above it converge to the high-wealth equilibrium. Persistent stratification therefore does not require between-group talent differences; it arises from signal contamination by wealth, Bayesian updating under partial observability, institutional thresholding, and intergenerational feedback.

The intervention analysis reshapes the update map rather than replacing the model. Lowering the effective threshold $j$1 can eliminate low-wealth traps. Aligning the signal with institutional preferences yields the summary claim

$j$2

because misalignment deepens disadvantage when the observable overweights wealth relative to what the institution values. The paper also studies direct transfers. If there are three fixed points $j$3, then a one-shot transfer

$j$4

pushes a disadvantaged group past the unstable threshold. For repeated subsidies, defining

$j$5

any constant map-shift $j$6 is sufficient to escape the low basin. The model is intentionally stylized, but its contribution is precise: equal talent is not enough to prevent long-run inequality when wealth enters the signal and institutional decisions recursively reproduce it.

4. Generative AI, privatized sociality, and unequal exposure

"Synthetic Sociality: How Generative Models Privatize the Social Fabric" argues that generative models automate not only intellectual labor or intelligence but a broader set of human social capacities called social doing, defined as “what we exert to build a social connection with another” (Dodik et al., 13 May 2026). The paper distinguishes use social doing from exchange social doing, and also distinguishes substitutive automation from mediative automation. Substitutive automation includes companion chatbots, AI therapy bots, bot farms, and synthetic sexual imagery; mediative automation includes having the model write messages, produce visual artifacts for communication, summarize someone else’s writing, or generate social media content. The paper’s culminating concept, Synthetic Sociality, names “a social reality in part fabricated by Silicon Valley’s privately owned and undemocratically governed generative models.” Its central stratification claim is that many contribute sociality as data and content, while a few own the infrastructures, training pipelines, and interfaces that turn that “dead social doing” into commodified synthetic mediation.

"The Privilege of Exposure: Caste and Generative AI in India's Graduate Labour Market" makes this distributional problem empirical in a labor-market setting (Mishra, 11 Jun 2026). Using India’s redesigned Periodic Labour Force Survey for calendar year 2025, the paper studies 82,830 employed graduates and maps three occupational AI-exposure indices to NCO-2015 occupations. Its headline result is a steep caste gradient among 83,000 employed graduates: graduates from the Scheduled Castes and the Scheduled Tribes are 0.24--0.37 standard deviations less exposed than upper-caste graduates within the same district. Two channels drive the gap. First, one in four SC and one in three ST graduates work in farm or elementary occupations untouched by AI. Second, among those in white-collar work, disadvantaged-caste graduates are underrepresented in managerial, software, and finance occupations. Because exposure carries a wage premium of up to 20 per cent, the paper concludes that generative AI stands to widen, not narrow, India’s caste earnings gap.

The exposure measures are occupation-level estimates of potential interaction between tasks and current AI capabilities, not adoption. This distinction is explicit: exposure is not evidence that a worker is using AI, has access to AI, knows how to use it, or is benefitting from it. Yet the wage results show why exposure matters. For regular salaried graduates with positive monthly earnings, a one-standard-deviation increase in AIOE exposure is associated with 0.200 log points higher monthly earnings, roughly a 20 per cent wage premium; GPT exposure yields 0.184 log points. For the ILO GenAI score, the linear term is not significant until a squared term is added, producing an inverted-U pattern interpreted as evidence that some exposure dimensions are augmentative and rewarded while others are more automatable routine-cognitive exposure. The paper’s core interpretive move is that low exposure is not protection in this setting. SC and ST graduates are “insulated from AI for the worst possible reason”: they are excluded from the occupations where AI is likely to augment work.

Taken together, these papers place generative stratification in the domain of ownership and access. One paper emphasizes privatization of the means of social fabrication; the other shows that proximity to AI complements is already stratified by caste. A plausible implication is that generative AI can deepen hierarchy through both platform control over social capacities and unequal labor-market access to the occupations where AI exposure is economically valuable.

5. Experimental, adaptive, and simulation-based procedures

"Leveraging LLMs to Improve Experimental Design: A Generative Stratification Approach" uses the phrase in a design-stage sense: an LLM-based procedure that synthesizes high-dimensional covariates into a one-dimensional prognostic score for blocking before randomization (Gui et al., 30 Sep 2025). The theoretical backbone comes from Bai (2022). For treatment probability one-half, the optimal score for stratification is

$j$7

The method asks an LLM to predict both potential outcomes for each unit,

$j$8

and then constructs

$j$9

for $s$0, or

$s$1

for unequal assignment probabilities. Units are then sorted and paired, or combined with Mahalanobis distance through the hybrid cost

$s$2

Across four empirical applications, the paper reports that generative stratification reduces MSE by about 10\% to 50\% relative to simple randomization, lowers standard errors by about 5.6\% to 30.4\%, and can further improve on full-covariate Mahalanobis matching.

"Stratification Trees for Adaptive Randomization in Randomized Controlled Trials" addresses a related problem without LLMs: how to learn strata from a first-wave experiment so as to minimize the asymptotic variance of the second-wave ATE estimator (Tabord-Meehan, 2018). A stratification tree is

$s$3

where $s$4 is a decision-tree partition of covariate space and $s$5 gives stratum-specific treatment probabilities. Under a fixed tree, the asymptotic variance is

$s$6

The oracle target is

$s$7

and the main theorem establishes

$s$8

The paper also proposes cross-validation for depth selection and an evolutionary algorithm for searching tree space. Here generative stratification means that the partition is data-adaptive, outcome-aware, and learned from pilot data rather than fixed ex ante.

"Convenient Multiple Directions of Stratification" treats stratification as a variance-reduction problem for Monte Carlo with Gaussian inputs (Jourdain et al., 2010). The key object is the stratified estimator

$s$9

with variance

$\mathrm{RCA}(j,s)\ge 1$0

Optimal allocation is

$\mathrm{RCA}(j,s)\ge 1$1

The paper studies PCA, Linear Transformation, and a new Linear Approximation rule, for which the first direction is

$\mathrm{RCA}(j,s)\ge 1$2

It also gives a novel algorithm for generating normal vectors stratified along non-orthogonal directions by Gram–Schmidt orthogonalization plus sequential truncated-normal sampling. The empirical conclusion is that carefully chosen low-dimensional stratification, especially with optimal allocation, can outperform Latin Hypercube Sampling both statistically and computationally.

These procedures use the same word in a narrower, technical sense than the social-science literature. The shared feature is operational: strata are not assumed; they are generated from predictive models, pilot outcomes, or informative latent directions and then used to improve precision.

6. Stratified spaces and deep generative modeling

"A Deep Generative Approach to Stratified Learning" develops a geometric and statistical framework for data concentrated near a stratified space

$\mathrm{RCA}(j,s)\ge 1$3

where each $\mathrm{RCA}(j,s)\ge 1$4 is a compact embedded manifold of intrinsic dimension $\mathrm{RCA}(j,s)\ge 1$5, with $\mathrm{RCA}(j,s)\ge 1$6 (Martinez et al., 12 Apr 2026). The singular set is the union of pairwise intersections $\mathrm{RCA}(j,s)\ge 1$7, and the paper assumes transversal intersections, positive reach on each stratum viewed individually, and lower-dimensional singularities. The intrinsic distribution is

$\mathrm{RCA}(j,s)\ge 1$8

while observations are noisy,

$\mathrm{RCA}(j,s)\ge 1$9

so the ambient distribution is

$$P_{ijk} \propto \frac{M_i \cdot M_j \cdot S_k(\mathbf{x}_k)}{D_{ij}^{\gamma}},$$0

The paper proposes two generative frameworks. The first is a sieve maximum-likelihood model realized as a dimension-aware mixture of variational autoencoders. Its key geometric construction is a hierarchical generator in which different experts use only the first $$P_{ijk} \propto \frac{M_i \cdot M_j \cdot S_k(\mathbf{x}_k)}{D_{ij}^{\gamma}},$$1 latent coordinates, thereby matching latent dimension to stratum dimension. The second is a diffusion-based framework built on forward and reverse diffusion and the mixture structure of the score field. A central identity is

$$P_{ijk} \propto \frac{M_i \cdot M_j \cdot S_k(\mathbf{x}_k)}{D_{ij}^{\gamma}},$$2

so the global score is a convex combination of component scores with weights equal to posterior probabilities of stratum membership. This is what allows the score field to encode local stratum information.

The theoretical results separate ambient-distribution learning, intrinsic-distribution learning, and structure recovery. For the sieve MLE, Hellinger convergence rates depend on the hardest stratum through intrinsic dimension and smoothness rather than ambient dimension. For intrinsic recovery, if $$P_{ijk} \propto \frac{M_i \cdot M_j \cdot S_k(\mathbf{x}_k)}{D_{ij}^{\gamma}},$$3, then

$$P_{ijk} \propto \frac{M_i \cdot M_j \cdot S_k(\mathbf{x}_k)}{D_{ij}^{\gamma}},$$4

For the diffusion model, the score-learning and $$P_{ijk} \propto \frac{M_i \cdot M_j \cdot S_k(\mathbf{x}_k)}{D_{ij}^{\gamma}},$$5 rates likewise depend on intrinsic dimensions, stratum smoothness, time, and noise level. The paper also proves consistency for a local intrinsic-dimension estimator based on singular values of sampled score vectors,

$$P_{ijk} \propto \frac{M_i \cdot M_j \cdot S_k(\mathbf{x}_k)}{D_{ij}^{\gamma}},$$6

and proposes an algorithm that consistently estimates both the number of strata and their dimensions. Experiments on synthetic mixtures, butane, alanine dipeptide, and mixture-of-VAE reconstructions support the claim that complex data may be better modeled as unions of manifolds of varying dimensions than as a single manifold.

In this mathematical usage, generative stratification does not refer to social hierarchy or experimental blocking. It refers to a deep generative treatment of data whose support is itself stratified. This suggests a broader conceptual continuity across otherwise distant literatures: the term recurrently marks settings in which local geometry, local rules, or local predictions generate a global partitioned structure rather than presupposing one.