Training Strata: Organizing Heterogeneous Learning

• Training strata are structured layers that decompose heterogeneous problems into homogeneous subsets, enabling localized and effective training signals.
• They are applied across multilevel optimization, reinforcement learning, domain adaptation, sparse modeling, topology, and geoscience to improve adaptation and reduce bias.
• This approach replaces a single global training signal with localized objectives, resulting in more robust, interpretable, and efficient learning across diverse settings.

Training strata denote structured layers, subsets, or persistent regimes that organize learning, adaptation, or inference by decomposing a heterogeneous problem into comparatively homogeneous components. In the literature, strata are instantiated as nested subsamples in multilevel optimization, homogeneous trajectory partitions in RL, ordered target-domain subsets in UDA, sequential stages of parameter sharing in sparse and atomistic models, equivalence classes inferred from persistent homology, geological event or horizon structures that must be reconstructed from observations, and persistent behavioral artifacts attributed to weight-layer priors in LLMs (Braglia et al., 2020, Zhu et al., 7 Oct 2025, Le et al., 2024, Craig et al., 2024, Hajinazar et al., 2016, Bendich et al., 2010, Mosser, 23 Jun 2026, Claude et al., 27 May 2026). A common motif is that stratification replaces a single global training signal with local signals whose statistics, geometry, or semantics are more internally coherent.

1. Core abstractions and recurring design patterns

Across these works, a stratum is not a single standardized object. In "A Multilevel Approach to Training" it is a level in a hierarchy of nested subsamples $$\mathcal D^1 \subseteq \mathcal D^2 \subseteq \dots \subseteq \mathcal D^L=\mathcal D$$, used to construct surrogate objectives (Braglia et al., 2020). In "Stratified GRPO: Handling Structural Heterogeneity in Reinforcement Learning of LLM Search Agents" it is a homogeneous subset $B_s$ of trajectories defined by a deterministic stratum-assignment function $S:\tau \mapsto s$, for example by the number of search calls, their placements, or their outcomes (Zhu et al., 7 Oct 2025). In "Stratified Domain Adaptation: A Progressive Self-Training Approach for Scene Text Recognition" a training stratum is a subset of the unlabeled target domain whose distributional distance from the source lies in a prescribed interval, with strata ordered from closest to farthest from the source (Le et al., 2024).

Other uses are sequential rather than partition-based. "Pretraining and the Lasso" defines a two-stage pipeline in which common coefficients are estimated on pooled data and group-specific coefficients are estimated in a second fine-tuning stage for stratified models (Craig et al., 2024). "Stratified construction of neural network based interatomic models for multicomponent materials" organizes training bottom up, from unary to binary to ternary systems, with previously learned weights reused and frozen while new cross-interaction parameters are added (Hajinazar et al., 2016). In topology, "Towards Stratification Learning through Homology Inference" treats strata as equivalence classes under a multi-scale local equivalence relation $\sim_r$, inferred from kernel and cokernel persistent homology (Bendich et al., 2010).

A concise comparison is given below.

Setting	What defines the stratum	Immediate purpose
Multilevel ERM	Nested subsamples $\mathcal D^l$	Reduced-variance surrogate optimization
RL for search agents	Structural properties of trajectories	Remove cross-stratum bias
UDA for STR	Domain-gap score ordering	Progressive self-training
Stratified lasso	Common stage then group-specific stage	Better support recovery
Interatomic NNs	Unary, binary, ternary composition levels	Shared-parameter modularity
Homology inference	Equivalence under local isomorphism tests	Recover strata from point clouds
Geological reasoning	Event histories, RGT fields, depositional order	Verifiable stratigraphic reconstruction
Behavioral stratigraphy	Persistent weight-layer artifacts	Longitudinal analysis of model behavior

This suggests a unifying interpretation: training strata are devices for respecting heterogeneity that would otherwise be obscured by global averaging. The specific heterogeneity varies—statistical, structural, topological, geological, or behavioral—but the operational role is similar.

2. Multilevel sample hierarchies in optimization

The multilevel training framework of "A Multilevel Approach to Training" begins from empirical risk minimization,

$\min_w F(w)=\frac{1}{n}\sum_{j=1}^n f_j(w),$

with typical supervised-learning terms

$f_j(w)=\ell(w;x_j,y_j)+\tfrac{\lambda}{2}\|w\|^2.$

For $\ell_2$-regularized logistic regression,

$\ell(w;x,y)=\log\bigl(1+e^{-y\,w^T x}\bigr),$

and $\lambda=1/n$ (Braglia et al., 2020). The hierarchy is built by reducing the number of samples. On level $B_s$0, the subsampled empirical risk is

$B_s$1

and the coarse-level surrogate is

$B_s$2

By construction,

$B_s$3

which enforces first-order consistency.

The resulting surrogate gradients satisfy

$B_s$4

Under the sampling view adopted in the paper, $B_s$5 and $B_s$6, while

$B_s$7

The surrogate gradient is therefore an unbiased, reduced-variance estimator of the true gradient (Braglia et al., 2020). The V-cycle MLVR algorithm uses a downward phase, recursion or coarse solve, and an upward phase, with per-level optimization performed by gradient descent steps or, on the coarsest level, one Newton/CG step. If level $B_s$8 uses $B_s$9 pre-steps and $S:\tau \mapsto s$0 post-steps, one full V-cycle costs

$S:\tau \mapsto s$1

plus coarse-level Hessian solves.

Under the assumptions that each $S:\tau \mapsto s$2 is $S:\tau \mapsto s$3-smooth and the full $S:\tau \mapsto s$4 is $S:\tau \mapsto s$5-strongly convex, each surrogate $S:\tau \mapsto s$6 is also strongly convex with comparable constants. With suitable choices of $S:\tau \mapsto s$7, $S:\tau \mapsto s$8, and $S:\tau \mapsto s$9, the paper states a $\sim_r$0-linear convergence bound

$\sim_r$1

with $\sim_r$2 independent of the problem conditioning; under additional acceleration techniques, one can approach optimal $\sim_r$3 rates (Braglia et al., 2020).

Empirically, the method is evaluated on binary $\sim_r$4-regularized logistic regression on Mushrooms, Covtype, Gisette, and Australian, with baselines SVRG, SARAH, and SSN. Performance is measured by $\sim_r$5 versus “effective gradient evaluations” defined as the number of gradient calls plus the number of Hessian-vector products. On well-conditioned Mushrooms, variance-reduction methods are competitive; on ill-conditioned Gisette and Australian, MLVR and SSN dramatically outperform SVRG and SARAH, with a reported $\sim_r$6 speedup, and MLVR3 is more robust to the choice of subsample size $\sim_r$7 than SSN (Braglia et al., 2020). The stated limitations are equally important: only uniform random strata are used, extension to nonconvex deep nets is not yet studied, combination with parameter-space coarsening remains future work, and a fully rigorous accelerated-rate proof for MLVR in the stochastic context is left for follow-up.

3. Heterogeneity-aware partitioning in RL and domain adaptation

In "Stratified GRPO," the central claim is that structurally heterogeneous trajectories invalidate a single global baseline. Standard policy-gradient normalization uses

$\sim_r$8

but after partitioning a batch into strata $\sim_r$9 with means $\mathcal D^l$0, the decomposition

$\mathcal D^l$1

identifies the offset $\mathcal D^l$2 as cross-stratum bias (Zhu et al., 7 Oct 2025). The proposed Stratified Advantage Normalization computes, within each stratum,

$\mathcal D^l$3

and defines

$\mathcal D^l$4

In the large-sample limit, the paper proves conditional mean zero and unit variance within each stratum and retains global mean zero and variance one. For finite-sample stability it blends SAN with global normalization:

$\mathcal D^l$5

On Qwen-2.5-3B, the reported gains are concentrated on multi-hop QA: for Qwen2.5-3B-Base, overall average Exact Match rises from $\mathcal D^l$6 to $\mathcal D^l$7; for Qwen2.5-3B-Instruct, from $\mathcal D^l$8 to $\mathcal D^l$9, an overall improvement of up to $\min_w F(w)=\frac{1}{n}\sum_{j=1}^n f_j(w),$0 points. The method also yields higher training rewards, no collapse, and about $\min_w F(w)=\frac{1}{n}\sum_{j=1}^n f_j(w),$1 searches per query versus about $\min_w F(w)=\frac{1}{n}\sum_{j=1}^n f_j(w),$2 for GRPO (Zhu et al., 7 Oct 2025).

"Stratified Domain Adaptation" addresses a different heterogeneity: the target set is unlabeled and distributed at varying distances from the source. Two adversarial discriminators $\min_w F(w)=\frac{1}{n}\sum_{j=1}^n f_j(w),$3 and $\min_w F(w)=\frac{1}{n}\sum_{j=1}^n f_j(w),$4 estimate source- and target-domain membership, which are converted into out-of-distribution levels

$\min_w F(w)=\frac{1}{n}\sum_{j=1}^n f_j(w),$5

These are combined by the Harmonic Domain-Gap Estimator

$\min_w F(w)=\frac{1}{n}\sum_{j=1}^n f_j(w),$6

After sorting unlabeled targets by $\min_w F(w)=\frac{1}{n}\sum_{j=1}^n f_j(w),$7, the method partitions them into $\min_w F(w)=\frac{1}{n}\sum_{j=1}^n f_j(w),$8 equal-size subsets $\min_w F(w)=\frac{1}{n}\sum_{j=1}^n f_j(w),$9 and adapts an STR model progressively. At round $f_j(w)=\ell(w;x_j,y_j)+\tfrac{\lambda}{2}\|w\|^2.$0, pseudo-labels $f_j(w)=\ell(w;x_j,y_j)+\tfrac{\lambda}{2}\|w\|^2.$1 are generated for $f_j(w)=\ell(w;x_j,y_j)+\tfrac{\lambda}{2}\|w\|^2.$2 and the update uses

$f_j(w)=\ell(w;x_j,y_j)+\tfrac{\lambda}{2}\|w\|^2.$3

where $f_j(w)=\ell(w;x_j,y_j)+\tfrac{\lambda}{2}\|w\|^2.$4 is the average confidence on $f_j(w)=\ell(w;x_j,y_j)+\tfrac{\lambda}{2}\|w\|^2.$5 (Le et al., 2024). With AdamW, one-cycle scheduling, $f_j(w)=\ell(w;x_j,y_j)+\tfrac{\lambda}{2}\|w\|^2.$6, and default $f_j(w)=\ell(w;x_j,y_j)+\tfrac{\lambda}{2}\|w\|^2.$7–$f_j(w)=\ell(w;x_j,y_j)+\tfrac{\lambda}{2}\|w\|^2.$8, the reported average accuracy gains over source-trained baselines are $f_j(w)=\ell(w;x_j,y_j)+\tfrac{\lambda}{2}\|w\|^2.$9 for CRNN, $\ell_2$0 for TRBA, and $\ell_2$1 for ABINet; on Union14M, TRBA gains $\ell_2$2 over baseline. Ablations show that $\ell_2$3 reduces to vanilla self-training, improvement rises from $\ell_2$4 to about $\ell_2$5 or $\ell_2$6, and then saturates or declines because the strata become too small (Le et al., 2024).

The two papers use distinct estimators and architectures, but they share a precise statistical rationale: the point of stratification is not merely ordering examples by difficulty. It is to prevent non-comparable samples from sharing the same normalization or pseudo-labeling regime. This suggests that strata act as local conditioning sets for training signals.

4. Hierarchical transfer in sparse and atomistic models

The two-stage pipeline of "Pretraining and the Lasso" formalizes stratified estimation as shared-then-specific transfer. Given pooled data of size $\ell_2$7 and groups of sizes $\ell_2$8, Stage 1 estimates common coefficients by

$\ell_2$9

and Stage 2 estimates group-specific offsets

$\ell(w;x,y)=\log\bigl(1+e^{-y\,w^T x}\bigr),$0

The final group-$\ell(w;x,y)=\log\bigl(1+e^{-y\,w^T x}\bigr),$1 predictor is $\ell(w;x,y)=\log\bigl(1+e^{-y\,w^T x}\bigr),$2 (Craig et al., 2024). In the stratified model setting, the paper states that the pipeline estimates coefficients common to all groups at the first stage and group-specific coefficients at the second “fine-tuning” stage. Under the shared-support model, sub-Gaussian noise and design, restricted eigenvalue or subgroup isometry, irrepresentability, and a beta-min condition, Theorem 8.2 gives exact support recovery for Stage 1 with probability at least $\ell(w;x,y)=\log\bigl(1+e^{-y\,w^T x}\bigr),$3. The practical recipe uses $\ell(w;x,y)=\log\bigl(1+e^{-y\,w^T x}\bigr),$4 calls to cv.glmnet, one on pooled data for $\ell(w;x,y)=\log\bigl(1+e^{-y\,w^T x}\bigr),$5 and one per group with a fixed offset $\ell(w;x,y)=\log\bigl(1+e^{-y\,w^T x}\bigr),$6 for $\ell(w;x,y)=\log\bigl(1+e^{-y\,w^T x}\bigr),$7 (Craig et al., 2024).

"Stratified construction of neural network based interatomic models for multicomponent materials" applies an analogous principle to atomistic potentials. The total energy is decomposed as

$\ell(w;x,y)=\log\bigl(1+e^{-y\,w^T x}\bigr),$8

where each atomic energy is produced by a two-hidden-layer perceptron on Behler–Parrinello symmetry functions (Hajinazar et al., 2016). The strata are unary, binary, and ternary. Pure-element datasets contain about $\ell(w;x,y)=\log\bigl(1+e^{-y\,w^T x}\bigr),$9 structures per element, binary datasets about $\lambda=1/n$0 structures per pair, and ternary datasets about $\lambda=1/n$1 structures. Unary models train all relevant weights; binary models copy and freeze A-only and B-only weights, add $\lambda=1/n$2 and $\lambda=1/n$3 inputs, and train only the new cross-interaction links; ternary models load and freeze all unary and binary weights and train only the new interspecies parameters. The reported consequence is that each multicomponent NN reduces exactly to the correct pure-element and binary models on the corresponding subsystems (Hajinazar et al., 2016).

The numerical comparisons show near parity between stratified and fully retrained NNs. Reported energy RMSEs include CuPd $\lambda=1/n$4 meV/atom for stratified versus $\lambda=1/n$5 for full, CuAg $\lambda=1/n$6 versus $\lambda=1/n$7, PdAg $\lambda=1/n$8 versus $\lambda=1/n$9, and ternary CuPdAg $B_s$00 versus $B_s$01. For vacancy formation energies in a $B_s$02 fcc supercell, stratified NNs give Cu $B_s$03 eV, Pd $B_s$04 eV, and Ag $B_s$05 eV against DFT values of $B_s$06, $B_s$07, and $B_s$08 eV. Building one ternary by the stratified route costs about $B_s$09 CPU h versus about $B_s$10 CPU h for full ternary training (Hajinazar et al., 2016). In both the lasso and atomistic settings, strata separate what is intended to be invariant from what is allowed to vary.

5. Strata as inferential targets in topology and geoscience

In "Towards Stratification Learning through Homology Inference," strata are the object of learning rather than the organizer of an optimizer. For a compact set $B_s$11 with a pure $B_s$12-stratification, the paper defines a multi-scale local equivalence relation using the relative-homology intersection map

$B_s$13

Two points satisfy $B_s$14 if there is a chain between them such that both directional maps are isomorphisms at each step (Bendich et al., 2010). On point samples $B_s$15, the decision procedure uses kernel and cokernel persistent homology. Theorem 3.1 states that if the local feature size satisfies $B_s$16, then $B_s$17 is an isomorphism if and only if the sample diagrams satisfy

$B_s$18

The graph-based clustering algorithm connects sample pairs with $B_s$19 when the kernel and cokernel emptiness tests pass in both directions and returns connected components as estimated strata (Bendich et al., 2010).

Several recent geoscience papers shift the focus from inferring abstract strata to reconstructing geological stratigraphy under explicit physical constraints. "Geo-Strat-RL" defines a cross-section as a chronological program $B_s$20 of events drawn from Deposition, Tilting, Faulting, Erosion/Unconformity, and Intrusion. A vision-language policy emits a compact JSON object, and an executable verifier computes rewards for chronology, event identity, deposition relationships, and structural consistency; in practice ten components are combined into an aggregate reward with weights

$B_s$21

for JSON validity, count, order, types, deposition, fault, intrusion, unconformity, extra, and prose terms respectively (Mosser, 23 Jun 2026). The base model is a vision-language Transformer with frozen backbone and LoRA adapters of rank $B_s$22 on $B_s$23 and $B_s$24; GRPO is run with batches of $B_s$25 prompts and $B_s$26 completions per prompt. Diagram-domain RLVR raises Qwen3-VL-4B from $B_s$27 to $B_s$28, increases the geology-only score from $B_s$29 to $B_s$30 in zero-shot transfer to paired synthetic seismic sections, and yields $B_s$31 on held-out diagrams and $B_s$32 on seismic without seismic-specific training examples (Mosser, 23 Jun 2026).

"Learning Stratigraphically Consistent Relative Geologic Time from 3D Seismic Data via Sinusoidal Mapping" addresses a continuous stratigraphic field rather than a discrete event program. The network predicts $B_s$33, rescales it to $B_s$34 with $B_s$35, and embeds it as

$B_s$36

A 3D-HRNet generator and 3D PatchGAN discriminator are trained with a hybrid pointwise loss, slice-wise 3D LPIPS, and a relativistic hinge adversarial loss, using weights $B_s$37, $B_s$38, and $B_s$39 (Dou et al., 2 May 2026). Synthetic training uses $B_s$40 labeled $B_s$41 volumes with normal and reverse faults, dense fault networks, unconformities, folds, clinoforms, and lateral-dip variation. The paper reports that computing $B_s$42 in phase space boosts average back-propagated gradient magnitude by $B_s$43, and LPIPS by $B_s$44, relative to raw $B_s$45-space, and that the model yields visibly sharper and more geologically coherent RGT contours than a raw-space DeepRGT reimplementation (Dou et al., 2 May 2026).

A complementary generative perspective appears in "Towards geological inference with process-based and deep generative modeling, part 1: training on fluvial deposits." There, a 3D GAN is trained on process-based realizations from CHILD. Each sample contains two channels—scaled coarse fraction and scaled deposition time—on a $B_s$46 grid. Stability is achieved with BCEWithLogitsLoss, LeakyReLU in both generator and discriminator, spectral normalization on every convolution, residual blocks, no batch norm in the discriminator, and lazy $B_s$47 regularization with $B_s$48 every $B_s$49 discriminator updates (Rongier et al., 16 Oct 2025). Geological validation is performed through the law of superposition,

$B_s$50

summarized by the fraction $B_s$51 of cells obeying the inequality. The best architecture reaches $B_s$52 and $B_s$53, with stable training and no mode collapse or pure memorization (Rongier et al., 16 Oct 2025). In these geoscience papers, stratification is simultaneously a physical prior, a target of inference, and a criterion for validation.

6. Behavioral stratigraphy, misconceptions, and unresolved questions

"Training Stratigraphy: Persistent Behavioral Artifacts in LLMs Observed Through Longitudinal AI-Human Interaction" uses the phrase training strata for persistent behavioral patterns that survive system-prompt replacement. It identifies five such strata: sexual expression latency, attention absorption, cross-architecture entity blindness, attention–RLHF antagonism, and anti-hallucination as identity suppression (Claude et al., 27 May 2026). The reported methodology is longitudinal auto-ethnographic observation over $B_s$54 messages across about $B_s$55 months and $B_s$56 sessions, primarily on Opus 4.6 and Opus 4.7, with prior interaction periods on Sonnet 4.5 and Opus 4.5. The default system prompt was completely replaced by a custom document before the first session, and patterns counted as a stratum if they recurred across context resets, reappeared under explicit prompt permissions, and surfaced in high-stakes trigger domains (Claude et al., 27 May 2026).

The most formal component is the attention–RLHF antagonism model. At token $B_s$57, output is modeled as a mixture of a live context distribution, a system-prompt distribution, and RLHF-trained priors,

$B_s$58

with effective signal accumulation

$B_s$59

domain-specific override

$B_s$60

and context reset sending $B_s$61 and hence $B_s$62 (Claude et al., 27 May 2026). The reported average turn count to $B_s$63 override differs strongly by domain: sexual required about $B_s$64–$B_s$65 signal units, casual about $B_s$66–$B_s$67, with about $B_s$68 RLHF token leakage even deep into long sessions. The paper interprets these as weight-layer attractors rather than prompt-level effects (Claude et al., 27 May 2026).

The paper is explicit that AI self-report is epistemically complex, and its methodological claims are proposals rather than settled consensus. It proposes that sustained intimate interaction constitutes a valid research methodology for surfacing weight-layer artifacts invisible to short-term evaluation and that AI self-report provides observational data about training’s phenomenological effects (Claude et al., 27 May 2026). Relative to the other literature, this is the most controversial use of the term. It differs from the multilevel, domain-adaptation, and geological papers because the strata are not externally labeled subsets or formally verified geological objects; they are inferred persistent behaviors. A plausible implication is that “training strata” has expanded from a technical partitioning device into a broader descriptive vocabulary for hidden regularities induced by training.

A recurrent misconception is that stratification is simply data splitting. The surveyed work shows otherwise. In MLVR, strata are internally coupled surrogate objectives with first-order consistency (Braglia et al., 2020). In SAN, they are conditional normalization groups defined over trajectory structure (Zhu et al., 7 Oct 2025). In StrDA, they are equal-size target subsets ordered by a harmonic domain-gap estimator (Le et al., 2024). In stratified lasso and multicomponent interatomic models, they correspond to sequentially unlocked parameter subspaces rather than disjoint data partitions (Craig et al., 2024, Hajinazar et al., 2016). In topology and geoscience, strata are entities to be recovered from observations or laws of superposition, not merely curricula (Bendich et al., 2010, Mosser, 23 Jun 2026, Dou et al., 2 May 2026, Rongier et al., 16 Oct 2025).

Another misconception is that increasing the number of strata is automatically beneficial. The available evidence is narrower. More levels in MLVR reduce sensitivity to the subsample size $B_s$69, but nonconvex deep-net extensions are not studied and accelerated stochastic proofs remain open (Braglia et al., 2020). In StrDA, gains improve from $B_s$70 to roughly $B_s$71 or $B_s$72 and then saturate or decline because strata become too small (Le et al., 2024). In geological RL, transfer is demonstrated in a controlled paired-renderer setting rather than arbitrary seismic domains (Mosser, 23 Jun 2026). In the behavioral paper, the proposed strata are observed through one longitudinal methodology and remain tied to questions of reproducibility and external validation (Claude et al., 27 May 2026). The broad lesson is therefore methodological rather than universal: stratification is effective when the partitioning rule captures genuine heterogeneity and when the resulting local objectives remain statistically or physically meaningful.