Stratified Attention Mechanisms

• Stratified attention is an approach where attentional selection is distributed across nested layers or strata, aligning with abstraction levels, spatial regions, or feature dimensions.
• It appears in diverse applications such as neurosymbolic cognition, dimensional-shift tasks, and high-resolution image analysis, clearly separating low-level perception from high-level reasoning.
• This design pattern enhances performance by optimizing resource allocation, reducing attention overload, and improving outcomes in tasks like image generation, object recognition, and anomaly detection.

Searching arXiv for relevant papers on stratified attention and related uses across cognitive architectures, decision-making, vision, and generative modeling. Stratified attention denotes a class of attention mechanisms in which attentional selection is organized across distinct levels, regimes, or partitions rather than implemented as a single undifferentiated weighting operation. Across the literature, the term refers to several related but non-identical ideas: a multilevel attentional architecture aligned with levels of abstraction in neurosymbolic cognition; a structured selector over feature dimensions in dimensional-shift learning; a combination of spatially stratified sampling and attention-based aggregation for high-resolution image analysis; separate attention flows over heterogeneous key-value sources in image-prompted diffusion; and attention-allocation schemes that distinguish sharply relevant, weakly relevant, and irrelevant tokens in sequence models. In all of these settings, the common principle is that attentional control is not flat: different information sources, spatial regions, dimensions, or abstraction levels are assigned different attentional roles or strata (Latapie et al., 2021, Malloy et al., 19 Jan 2025, Han, 1 Jan 2026, Jo et al., 4 Aug 2025, Fu et al., 1 Jan 2026).

1. Conceptual definition and scope

In cognitive and computational usage, stratified attention is best understood as attention organized in nested layers aligned with levels of abstraction, where low-level attention is fast, automatic, and perception-grounded, while higher-level attention is slower, deliberate, and goal-directed (Latapie et al., 2021). This formulation rejects a hard symbolic-versus-subsymbolic split and instead treats both low and high levels as containing “neurosymbolic information,” with the primary differentiator being the character of the attention mechanism rather than the representational substrate itself (Latapie et al., 2021).

The same idea appears in more task-specific forms. In dimensional-shift bandit tasks, attention is structured because the learner must first determine which dimension is relevant and then use values within that dimension to guide choice (Malloy et al., 19 Jan 2025). In high-resolution image detection, the stratification is spatial and operational: a model processes a globally resized image while also sampling multiple local crops from different image regions, then reweights those crops by informativeness (Han, 1 Jan 2026). In image-prompted diffusion, stratification refers to separate attention computations over generated-image features and prompt-image features, combined only after each stream has formed its own attention map (Jo et al., 4 Aug 2025). In Transformer sequence modeling, a related perspective treats attention as stratified when informative tokens receive sharp, discriminative focus while irrelevant tokens are suppressed, rather than all tokens being forced into a single normalized pool (Fu et al., 1 Jan 2026).

This suggests that “stratified attention” is not a single canonical architecture. A plausible implication is that it is better regarded as a design pattern: attentional selection is decomposed across levels, sources, or functional modes so that relevance can be represented more explicitly than in a monolithic attention distribution.

2. Multilevel attention in neurosymbolic cognition

A central formulation appears in “Neurosymbolic Systems of Perception & Cognition: The Role of Attention” (Latapie et al., 2021). The paper argues against a rigid System-1/System-2 interpretation in which fast processing is subsymbolic and slow processing is symbolic. Instead, it proposes a continuous, multilevel architecture in which both low and high levels contain symbolic and subsymbolic content, and the critical difference lies in the kind of attention they employ (Latapie et al., 2021).

Within this framework, low-level cognition and perception use high-speed, non-conscious attention, whereas higher-level cognition is more single-threaded, slower, and involves conscious attention, working memory, and deliberate control (Latapie et al., 2021). The authors emphasize that attention is the “fundamental currency of intelligence” and use this to organize the relation between perception, cognition, and incremental learning (Latapie et al., 2021). Attention is therefore stratified both across abstraction levels and across functional modes: it can be autonomous and bottom-up, or goal-directed and top-down (Latapie et al., 2021).

The paper makes a consequential distinction between knowledge representation and information processing. The symbolic/subsymbolic distinction pertains to representation, whereas the automatic/controlled distinction pertains to processing; the two are often conflated but need not align (Latapie et al., 2021). The bike-riding example serves as a concrete illustration: learning begins as controlled and conscious, with attention focused on gross goals such as “don’t fall,” and over time becomes distributed to dynamic variables such as tilt angle and steering corrections, until the control loop is increasingly internalized and automatic (Latapie et al., 2021). In this account, failure is not only a deficit of representation but also a deficit in knowing what to attend to.

The paper also anchors multilevel attention in neurobiological discussion of the thalamus and thalamocortical/corticothalamic loops. Thalamocortical pathways create and maintain task-related representations and help update cortical representations, while corticothalamic pathways direct cognitive resources and contribute to learning (Latapie et al., 2021). The authors summarize this relation by stating that the thalamocortical pathway defines the world for the cortex, while the corticothalamic pathway uses attention to tell the thalamus what the cortex needs (Latapie et al., 2021). Stratification here is not a separation into isolated modules; it is a coupled bidirectional process in which lower-level sensory integration grounds higher-level representation and higher-level goals bias lower-level selection.

3. Architectural mechanisms: hierarchy, partitioning, and sub focus

The neurosymbolic metamodel in (Latapie et al., 2021) provides one of the clearest architectural realizations of stratified attention. The hierarchy is specified as L0 for raw sensor data, L1 for links between raw data and abstractions, L2 for higher integrated abstractions learned via statistical learning and reasoning, and L* for high-level goals and motivations (Latapie et al., 2021). The paper further allows nested sublevels within L2, implying unbounded abstraction depth (Latapie et al., 2021).

A key mechanism is that the metamodel can be partitioned into subgraphs “using the hierarchical abstraction substructure associated with the current focus of attention” (Latapie et al., 2021). Attention therefore determines how the system decomposes a problem. The paper introduces the notion of a sub focus of attention (sFoA), where a partition may trigger a subfocus that requests either subsymbolic data from System-1 or symbolic answers from System-2 (Latapie et al., 2021). This operationalizes attention as a nested control structure rather than a single selector.

Attention is also treated as resource allocation. The paper links this role to the Non-Axiomatic Reasoning System (NARS), where attention allocates inference steps and supports revision, composition, and reliability tracking, and to AERA (Auto-catalytic Endogenous Reflective Architecture), where attention functions as system-wide control of computational resources (Latapie et al., 2021). Incremental learning, on this view, depends on a recurrent loop in which new sensory evidence is related to existing abstractions, surprising observations trigger revision, and higher-level discoveries can propagate back into lower-level routines (Latapie et al., 2021).

The paper’s application cases illustrate this operational meaning. In retail object and shelf recognition, operating on bounding boxes with spatial semantics rather than training a Deep Learning model on product or shelf images, the use of focus of attention increased overall accuracy from 46.30% without FoA to 94.73% with FoA (Latapie et al., 2021). In network telemetry and anomaly detection, for more than 200,000 time series and 30 million data points, Matrix Profile algorithms identify regime changes, condensing millions of sensory points into a smaller number of events and explainable knowledge structures (Latapie et al., 2021). In both cases, attentional partitioning acts as a stratified reduction from raw data to interpretable abstractions.

4. Structured attention over dimensions and memory

A distinct but related notion appears in “Modeling Attention during Dimensional Shifts with Counterfactual and Delayed Feedback” (Malloy et al., 19 Jan 2025). The task is a contextual bandit with three abstract attributes, each having six possible values; one feature value yields reward with probability 0.9, while other choices yield reward with probability 0.1, with Gaussian noise $\mathcal{N}(0,0.1)$ (Malloy et al., 19 Jan 2025). Halfway through training, the task undergoes either an intradimensional shift, in which the relevant dimension stays the same but the high-reward feature value changes, or an extradimensional shift, in which the relevant dimension itself changes (Malloy et al., 19 Jan 2025).

The paper contrasts two attentional learning rules. The first is an RL-based mechanism in which dimension weights $\omega$ are updated from reward prediction errors, with option value defined as

$V(C) = \sum_{f \in C} \omega_f V(f)$

and weights updated by

$\omega_j' = \omega_j + \alpha (r - \omega_j)$

(Malloy et al., 19 Jan 2025). The second is an Instance-Based Learning (IBL) model in which individual experiences are stored in memory $\mathcal{M}$ and dimension weights are updated from mutual-information-based relevance:

$w_j = w_j + \alpha (I_i(j|u) - w_j)$

(Malloy et al., 19 Jan 2025).

The paper’s central claim is that the memory-based, information-theoretic approach is better able to account for human-like behavior in tasks that shift dimensions and alter feedback presentation (Malloy et al., 19 Jan 2025). Under immediate feedback, only the IBL + mutual-information model shows the human-like jumpstart performance difference between intradimensional and extradimensional shifts, where jumpstart is the average percent correct of the first 10 post-shift trials (Malloy et al., 19 Jan 2025). Under clustered or delayed feedback, where reward is observed only after 5 selections as a sum, all models show some early performance drop relative to immediate feedback, but only IBL with mutual-information learning shows a persistent reduction across training (Malloy et al., 19 Jan 2025). Under additional or counterfactual feedback, where outcomes of unchosen options are also observed, the memory-based model improves faster in the intradimensional condition and exhibits a clearer post-shift drop in the extradimensional condition (Malloy et al., 19 Jan 2025).

This formulation of stratified attention is not about levels of abstraction in the neurosymbolic sense. Rather, attention is structured over feature dimensions, and its relevance is inferred from accumulated evidence across memory rather than updated solely from local error signals. A plausible implication is that this is a higher-order attentional selector: it chooses not just among features, but among feature dimensions as abstract explanatory axes.

5. Spatial and visual forms of stratification

Stratified attention has also been used to describe spatially layered attentional effects in vision science. In “Attentional Modulation of Visual Spatial Integration: Psychophysical Evidence Supported by Population Coding Modeling” (Grillini et al., 2019), attention alters visual spatial integration at a later pooling stage rather than primarily changing early encoding. The study distinguishes spatial attention (SA) and feature-based attention (FBA) using a gaze-contingent search task followed by a 2-AFC orientation discrimination task (Grillini et al., 2019).

Integration strength is defined as

$$\text{Integration strength} = \text{JND}_{\text{crowded}} - \text{JND}_{\text{isolated}}$$

(Grillini et al., 2019). The main psychophysical result is that spatial attention produces a strong, localized reduction in integration strength, whereas feature-based attention produces a more modest, global reduction (Grillini et al., 2019). Visual deprivation versus baseline yields a significant reduction in integration strength at $7^\circ$ with $P = 0.008$, while information deprivation versus baseline yields a smaller but significant overall reduction with $P = 0.027$ (Grillini et al., 2019). Crucially, isolated targets show no significant attentional effect, so the modulation is specific to integration or crowding rather than general sensitivity (Grillini et al., 2019).

The proposed mechanism is a population-coding model with four functional layers: encoding, noisy population response, spatial integration, and decoding (Grillini et al., 2019). Attention acts at the integration layer by modulating Gaussian-shaped weighting functions

$\omega$0

(Grillini et al., 2019). Spatial attention narrows the weighting function at the attended location, while feature-based attention narrows weighting functions across the whole field more mildly (Grillini et al., 2019). In this use, “stratified” refers to functionally distinct layers of processing and to localized versus field-wide regimes of integration.

A different spatial realization appears in “Attention to Detail: Global-Local Attention for High-Resolution AI-Generated Image Detection” (Han, 1 Jan 2026). The proposed GLASS architecture combines a globally resized image with several local crops from the original-resolution image. The global branch processes the full image resized to $\omega$1, while the local branch processes multiple $\omega$2 crops extracted from the original image (Han, 1 Jan 2026). The local sampler defines a grid

$\omega$3

and, when $\omega$4, selects $\omega$5 distinct grid cells and samples one crop from each, thereby enforcing spatially stratified sampling (Han, 1 Jan 2026).

Each crop yields an embedding $\omega$6, which is scored by additive attention using a two-layer MLP with Tanh and dropout:

$\omega$7

and aggregated as

$\omega$8

(Han, 1 Jan 2026). The local summary is concatenated with the global embedding before final classification (Han, 1 Jan 2026). On the reported test sets, GLASS improves over standard transfer learning across all three backbones: ViT-Base/16 from 93.244% to 97.533%, ConvNeXt-Tiny from 96.133% to 98.611%, and ResNet-50 from 93.656% to 97.656% (Han, 1 Jan 2026). Here the stratification lies in the combination of broad spatial coverage and selective crop weighting.

6. Stratified attention in generative and sequence models

In generative diffusion modeling, “Devil is in the Detail: Towards Injecting Fine Details of Image Prompt in Image Generation via Conflict-free Guidance and Stratified Attention” (Jo et al., 4 Aug 2025) introduces the term explicitly. The paper considers image-prompted generation, where a prompt image $\omega$9 should transfer fine details such as textures, color tone, patterns, and makeup into a generated image $V(C) = \sum_{f \in C} \omega_f V(f)$0 (Jo et al., 4 Aug 2025). Earlier self-attention modifications either replace the generated image’s keys and values with prompt-derived ones or concatenate prompt and generated keys/values (Jo et al., 4 Aug 2025). The authors identify a realism-versus-alignment trade-off: more prompt keys and values improve alignment, while more generated-image keys and values improve realism (Jo et al., 4 Aug 2025).

Their solution is Stratified Attention, which computes prompt-related and generated-image-related attention separately and combines the resulting outputs:

$V(C) = \sum_{f \in C} \omega_f V(f)$1

with

$V(C) = \sum_{f \in C} \omega_f V(f)$2

(Jo et al., 4 Aug 2025). This avoids forcing prompt and generated features to compete within a single joint softmax. The method is paired with conflict-free guidance, in which the image prompt is used only in the positive branch of classifier-free guidance and omitted from the negative branch, thereby removing contradictory conditioning signals (Jo et al., 4 Aug 2025).

The paper evaluates the method on COCO 2017 validation, ImageNet-R-TI2I, and CelebAHQ, across image variation, cross-prompt image generation, and structure-guided image generation (Jo et al., 4 Aug 2025). Reported image variation scores include CelebAHQ CLIP-I of 0.878 versus 0.845 for IP-Adapter, 0.850 for SSR-Encoder, and 0.739 for RIVAL; ImageNet-R-TI2I CLIP-I of 0.883; and COCO CLIP-I of 0.873 (Jo et al., 4 Aug 2025). The paper recommends using Stratified Attention for about 10 to 30 steps, and recommends $V(C) = \sum_{f \in C} \omega_f V(f)$3 based on the balance between prompt alignment and generated-image contribution (Jo et al., 4 Aug 2025).

A related but differently motivated development appears in “Attention Needs to Focus: A Unified Perspective on Attention Allocation” (Fu et al., 1 Jan 2026). This paper does not use the term to denote multilevel abstraction or multi-source conditioning, but it argues that standard attention fails because of improper attention allocation, specifically Attention Overload and Attention Underload (Fu et al., 1 Jan 2026). Its Lazy Attention introduces positional discrimination across heads and dimensions and Elastic-Softmax:

$V(C) = \sum_{f \in C} \omega_f V(f)$4

to allow attention on irrelevant tokens to be suppressed to zero (Fu et al., 1 Jan 2026). The model reaches up to 59.58% attention sparsity and reduces attention sink while remaining competitive on nine benchmarks evaluated on the FineWeb-Edu corpus (Fu et al., 1 Jan 2026). This suggests a broader interpretation of stratification as attention organized into qualitatively different relevance tiers rather than distributed uniformly over all candidates.

7. Comparative interpretation, misconceptions, and significance

Across these works, stratified attention does not denote a single equation or universally accepted module. The term spans at least five recurring organizational motifs.

Motif	Representative formulation	Representative paper
Abstraction-level stratification	Fast non-conscious low-level attention and slower conscious higher-level attention	(Latapie et al., 2021)
Dimensional stratification	Attention learned over feature dimensions from mutual information in memory	(Malloy et al., 19 Jan 2025)
Spatial stratification	Distinct grid-based crop sampling with attention-based reweighting	(Han, 1 Jan 2026)
Source-wise stratification	Separate attention flows for generated and prompt features	(Jo et al., 4 Aug 2025)
Relevance-tier stratification	Sharp focus on relevant tokens and suppression of irrelevant ones	(Fu et al., 1 Jan 2026)

One common misconception is to equate stratified attention with a strict two-system architecture. The neurosymbolic account explicitly rejects that binary, arguing instead for a layered continuum in which all levels are neurosymbolic and differ mainly in attentional dynamics (Latapie et al., 2021). Another misconception is to treat stratification as merely sparse attention. In some papers, sparsity is relevant, as in Lazy Attention (Fu et al., 1 Jan 2026), but in others the key idea is not sparsity but separation of attentional roles, as in separate prompt and generated-image attention flows (Jo et al., 4 Aug 2025) or distinct global and local image streams (Han, 1 Jan 2026). A further misconception is that stratification always refers to anatomical layers; in the psychophysical study it refers instead to functional layers of processing, especially the integration stage (Grillini et al., 2019).

The broader significance of the idea is methodological. Stratified attention provides a way to encode task structure into attentional allocation. In neurosymbolic systems, it links perception, cognition, and cumulative learning through bottom-up refreshing and top-down control (Latapie et al., 2021). In dimensional-shift learning, it supports transfer by representing which dimension is informative across accumulated experience (Malloy et al., 19 Jan 2025). In high-resolution vision, it preserves global semantics while retaining fine local evidence under feasible computational constraints (Han, 1 Jan 2026). In image prompting, it mitigates realism-versus-alignment trade-offs by preventing heterogeneous sources from competing too early (Jo et al., 4 Aug 2025). In sequence modeling, it reframes collapse and sink as failures of attention allocation rather than isolated pathologies (Fu et al., 1 Jan 2026).

Taken together, the literature supports a general definition: stratified attention is an attentional organization in which relevance is distributed across differentiated strata—of abstraction, dimension, space, source, or token status—so that selection, integration, and control can be matched more closely to the structure of the task and the form of the information being processed.