Hierarchical Identity Learning Overview

• Hierarchical Identity Learning denotes methods that model identity across multiple levels—from instance to coarse abstractions—capturing nuanced semantic information.
• It employs architectures like hierarchical attention, dual concept bottlenecks, and multi-center clustering to effectively preserve and infer identity.
• Applications span multi-subject video generation, image self-supervised learning, and acoustic and text identification, yielding improved performance metrics.

Searching arXiv for the cited HIL-related papers to ground the article in current records. arxiv_search query: Hierarchical Identity Learning ID-Composer (Pan et al., 1 Nov 2025) arxiv_search query: (Shi et al., 15 Sep 2025) Hierarchical Identity Learning unsupervised visible infrared person re-identification arxiv_search query: (Xu et al., 2022) HIRL hierarchical image representation learning; (Xie et al., 2 Apr 2026) HIL-CBM hierarchical interpretable label-free concept bottleneck model Hierarchical Identity Learning (HIL) denotes a family of methods that represent, preserve, infer, or control “identity” across multiple levels of abstraction rather than through a single flat embedding or policy. In the cited literature, the notion appears in several distinct but structurally related forms: multi-subject video generation with hierarchical identity-preserving attention, self-supervised image representation learning with fine-to-coarse semantic codes, visible–infrared person re-identification with coarse and fine identity memories, acoustic identification with individual–species–taxon structure, concept bottleneck models with basic- and subordinate-level concepts, LLM role-playing with compositional personality and profession identities, hierarchical text classification with entropy-optimized structure encoders, and hierarchical imitation learning where the latent “identity” is an option or sub-skill rather than a person or object (Pan et al., 1 Nov 2025, Xu et al., 2022, Shi et al., 15 Sep 2025, Nolasco et al., 2024, Xie et al., 2 Apr 2026, Sun et al., 2024, Zhu et al., 2024, Gao et al., 2022, Chen et al., 2022, Buamanee et al., 4 Mar 2026).

1. Terminology and scope

The literature does not present a single standardized use of the term. Some works explicitly name a method “Hierarchical Identity Learning,” as in unsupervised visible–infrared person re-identification, where identity is organized as instance $\rightarrow$ fine centers $\rightarrow$ coarse centers (Shi et al., 15 Sep 2025). Other works implement the same idea implicitly without naming it as HIL, as in multi-subject video generation, where identity is preserved through intra-subject, inter-subject, and multi-modal attention but “the paper does not introduce a separate acronym ‘HIL’ or an explicit ‘Hierarchical Identity Learning’ module” (Pan et al., 1 Nov 2025). Still other papers use the acronym HIL for “Hierarchical Imitation Learning,” where the operative identity is a latent skill, option, or subtask rather than a visual entity (Gao et al., 2022, Chen et al., 2022, Buamanee et al., 4 Mar 2026).

A common pattern nevertheless recurs: identity is not treated as a single atomic label. Instead, it is represented through multiple levels such as local versus global, subordinate versus basic, individual versus species, or subtask versus task. This suggests that HIL is best understood as an architectural and learning principle: preserve or infer identity through structured multi-level representations rather than flat conditioning.

Domain	Hierarchical identity structure	Representative work
Multi-subject video generation	Intra-subject $\rightarrow$ inter-subject $\rightarrow$ multi-modal	(Pan et al., 1 Nov 2025)
Image SSL	Fine-grained $z^0$ to coarse $z^L$	(Xu et al., 2022)
USVI-ReID	Instance $\rightarrow$ fine centers $\rightarrow$ coarse centers	(Shi et al., 15 Sep 2025)
Acoustic identification	Individual $\rightarrow$ species $\rightarrow$ taxon	(Nolasco et al., 2024)
LLM role-playing	Personality identities + profession identities	(Sun et al., 2024)
Concept bottlenecks	Lower/subordinate concepts + higher/basic concepts	(Xie et al., 2 Apr 2026)
HTC and imitation learning	Label-path or sub-skill hierarchy	(Zhu et al., 2024, Gao et al., 2022)

2. Representation hierarchies

A central formulation appears in Hierarchical Image Representation Learning, where each image is assigned a chain

$\rightarrow$0

with $\rightarrow$1 encoding the most fine-grained semantics and higher $\rightarrow$2 encoding increasingly coarse semantics. The framework factorizes

$\rightarrow$3

so that a baseline SSL method learns the most detailed representation and a hierarchical mechanism learns coarse semantics conditioned on it (Xu et al., 2022). In this formulation, identity is explicitly multi-granular: a single image may simultaneously carry instance-level, subtype-level, and superclass-level identity.

An analogous structure appears in label-free concept bottlenecks. HIL-CBM learns two concept vectors from a shared backbone, a higher-level concept prediction $\rightarrow$4 and a lower-level concept prediction $\rightarrow$5, then applies dual classification heads

$\rightarrow$6

Here, coarse identity corresponds to basic-level category prediction and fine identity corresponds to subordinate prediction, with explanations aligned to the same abstraction level as the label (Xie et al., 2 Apr 2026).

In acoustic identification of individual animals, the hierarchy is fully explicit in the labels: $\rightarrow$7 with each individual belonging to one species and each species belonging to one taxon (Nolasco et al., 2024). In visible–infrared re-identification, the same principle is instantiated not by taxonomy but by prototype structure: coarse-grained DBSCAN clusters define identity-level memories, and K-means sub-clusters within each coarse cluster define finer identity modes (Shi et al., 15 Sep 2025).

ID-Composer expresses a related hierarchy in token space. Each reference image is encoded into spatial subject tokens

$\rightarrow$8

and the model then refines identity first within subject, then across subjects, and finally across modalities with text and latent video tokens (Pan et al., 1 Nov 2025). The paper characterizes this as a “hierarchical identity-preserving attention mechanism,” which can be interpreted as an implicit hierarchical identity learning scheme.

A different but compatible notion of hierarchical identity appears in LLM role-playing. HIRPF organizes identities into personality categories and profession categories, so a role is a composition such as “high agreeableness + low extraversion + doctor.” Identity is therefore not a single persona token but a structured combination of coexisting and category-specific identities (Sun et al., 2024).

3. Architectural realizations

The most direct architectural realization in generative modeling is ID-Composer’s three-stage conditioning hierarchy inside each MMDiT transformer block. The conditioning side applies: (1) intra-subject attention to tokens from a single reference image, (2) inter-subject attention over the concatenation of subject tokens, and (3) multi-modal attention over subject features, text tokens, and latent video tokens. The paper describes this as a token-level, spatial hierarchy over conditioning tokens, while temporal consistency is handled by the base video DiT rather than an additional temporal hierarchy (Pan et al., 1 Nov 2025).

In representation learning, HIRL realizes hierarchy through prototypes and semantic paths. Hierarchical K-means produces prototype sets $\rightarrow$9 and edges $\rightarrow$0, yielding a tree of prototypes. For each image, the positive semantic path $\rightarrow$1 and sampled negative paths are contrasted using

$\rightarrow$2

This path-based design forces all levels of the representation to be jointly correct rather than independently plausible (Xu et al., 2022).

In HIL-CBM, the hierarchy is realized by dual concept bottleneck layers operating on a shared backbone plus two consistency mechanisms. The visual consistency term

$\rightarrow$3

encourages higher- and lower-level concepts to focus on similar spatial regions, while a Tree-path KL divergence enforces consistency between higher- and lower-level labels (Xie et al., 2 Apr 2026).

Visible–infrared re-identification uses a memory-bank architecture. Coarse cluster centers are

$\rightarrow$4

and each coarse cluster is split into $\rightarrow$5 fine centers $\rightarrow$6. Multi-Center Contrastive Learning then contrasts each instance against all fine centers in its assigned identity and one hard fine center from every other identity (Shi et al., 15 Sep 2025).

In LLM agents, HIRPF uses a masked mixture of LoRA experts. For block $\rightarrow$7,

$\rightarrow$8

where $\rightarrow$9 is a diagonal mask that activates only the selected identities in that block. Personality-related and profession-related LoRA modules are inserted alternately across transformer layers, yielding both conceptual and architectural hierarchy (Sun et al., 2024).

The Bayesian contextual face model implements hierarchy probabilistically. A global identity distribution $\rightarrow$0 is shared across contexts, while each context has its own context-specific weights $\rightarrow$1 over the same identities. In the unbounded-context extension, the model nests a DP over context-specific identity distributions,

$\rightarrow$2

so that frames first select a context and then sample identities from the corresponding context-specific measure (Castro et al., 2018).

For hierarchical text classification, HILL constructs a coding tree by structural entropy minimization and uses it as a structure encoder. Leaf nodes are initialized from a document-conditioned label feature matrix, and internal nodes are built bottom-up by

$\rightarrow$3

The pooled multi-level representation $\rightarrow$4 then acts as a hierarchy-aware positive view of the document (Zhu et al., 2024).

4. Learning objectives and inference

A major divide in the literature concerns how hierarchical identity is optimized. In generative video, ID-Composer combines a Rectified Flow objective

$\rightarrow$5

with an online RL phase that maximizes

$\rightarrow$6

The paper explicitly motivates this by stating that standard diffusion loss often fails in aligning critical concepts like subject ID (Pan et al., 1 Nov 2025).

HIRL adds hierarchy to arbitrary SSL backbones by optimizing

$\rightarrow$7

where $\rightarrow$8 is the baseline SSL loss and $\rightarrow$9 is a semantic path discrimination loss over positive and negative prototype paths (Xu et al., 2022). HIL-CBM uses cross-entropy plus elastic net for both classification heads, augments concept learning with the cubic-cosine alignment loss $z^0$0, and adds both visual and semantic consistency terms: $z^0$1 The explicit aim is to align spatial and label-space hierarchy (Xie et al., 2 Apr 2026).

Hierarchy-aware contrastive learning is prominent in acoustic identity. HiMulCon averages supervised contrastive losses across levels,

$z^0$2

and HiMulConE further constrains finer-level loss values using the maximum pairwise loss from the previous level. The effect is to enforce a geometry in which taxon, species, and individual identity remain nested rather than independent (Nolasco et al., 2024).

In USVI-ReID, the total objective is

$z^0$3

combining coarse identity contrastive loss, instance-level neighbor contrastive loss, and multi-center contrastive loss over fine identity memories. Cross-modal label quality is improved by Bidirectional Reverse Selection Transmission, which accepts a visible-to-IR or IR-to-visible association only if a reverse-selection condition such as $z^0$4 is satisfied (Shi et al., 15 Sep 2025).

HILL combines supervised multi-label prediction with contrastive alignment of text and structure views: $z^0$5 where $z^0$6 is an NT-Xent loss between the text encoder representation and the structure encoder representation. The paper frames this as “information lossless contrastive learning,” arguing that a coding-tree-based positive view preserves at least as much label-relevant information as an augmentation-based view (Zhu et al., 2024).

In the imitation-learning lineage, hierarchy is optimized through latent-variable inference over skill identities. DMIL meta-learns both a high-level selector $z^0$7 and low-level sub-skills $z^0$8 with sequential inner adaptation and simultaneous outer update (Gao et al., 2022). HierAIRL extends AIRL to option-aware trajectories by defining a discriminator on $z^0$9 and combining an imitation term with directed information,

$z^L$0

so that options are causally linked to behavior (Chen et al., 2022). Bi-HIL uses a high-level loss

$z^L$1

and a low-level CVAE objective

$z^L$2

where the “identity” variable is the active subtask and its subtask-level progress rate (Buamanee et al., 4 Mar 2026).

5. Empirical evidence across domains

In multi-subject video generation, the role of hierarchy is quantified directly. In ID-Composer’s ablation, removing hierarchical attention drops FaceSim from 58.12% to 51.34% and Total Score from 54.33% to 50.11%. On the subject-to-video custom benchmark, the 1.3B base model reaches FaceSim 58.12%, while the 14B version reaches 60.50%, which the paper identifies as the highest among all models in that table (Pan et al., 1 Nov 2025).

In image SSL, HIRL reports broad improvements over flat baselines. Representative examples include SwAV improving from KNN 63.45 to 63.99, linear 72.68 to 73.43, and fine-tune 76.82 to 77.18; DINO improving from KNN 76.01 to 76.84, linear 78.07 to 78.32, and fine-tune 82.09 to 83.24; and iBOT improving from fine-tune 82.47 to 83.37. Clustering quality also improves substantially, for example MoCo v2 Acc 29.3 to 32.9 (Xu et al., 2022).

HIL-CBM reports gains in both lower- and higher-level accuracy relative to sparse CBM baselines. With a ResNet backbone, it attains CIFAR-100 69.50% lower / 73.55% higher, CUB-200 75.35% lower / 83.08% higher, Places365 47.59% lower / 57.91% higher, and ImageNet 75.54% lower / 81.50% higher. The ablation on consistency terms is especially diagnostic: on ImageNet lower-level accuracy, the sequence “Neither” 72.03%, “Visual” 72.74%, “Semantic” 74.90%, and “Both” 75.54% indicates that both cross-level constraints contribute to hierarchical prediction quality (Xie et al., 2 Apr 2026).

Hierarchy-aware acoustic identification yields the clearest fine-grained gains at the identity level. On closed-set evaluation, the flat supervised-contrastive baseline obtains ID balanced accuracy 61.0%, while HC reaches 73.2% and HCE$z^L$3 reaches 72.3%. On unseen individuals, flat SC gives ID balanced accuracy 80.9%, while HC$z^L$4 reaches 88.1%. In 1-shot identification of unseen individuals, SC yields 6.0% at ID level, whereas HC$z^L$5 reaches 15.3%. The paper also reports no hierarchical inconsistency errors: predicted ID always belongs to predicted species, and predicted species always belongs to predicted taxon (Nolasco et al., 2024).

In unsupervised visible–infrared re-identification, the full HIL framework achieves Rank-1 66.30%, mAP 64.95%, and mINP 52.62% on SYSU-MM01 All Search, and Rank-1 92.82%, mAP 86.61%, and mINP 73.69% on RegDB V$z^L$6I. Ablation shows that BRST contributes the largest jump over the baseline, while MCCL adds further gains: on SYSU All Search, the baseline has R1 54.88 and mAP 52.62, adding BRST yields +10.65 R1 and +10.98 mAP, and the full model reaches 66.30 / 64.95 (Shi et al., 15 Sep 2025).

In LLM role-playing, the hierarchical controller primarily improves identity disentanglement rather than generic text quality. In personality-scale ablations, HIRPF achieves, for agreeableness, H = 4.85 and L = 1.35, compared with HIRP-Dense H = 4.05 and L = 1.90. In profession-scale ablations for the artist identity, HIRPF attains A = 4.85, D = 1.55, and P = 1.50, indicating high own-profession fidelity and low cross-profession leakage. On the open situation test, HIRPF reports AGR 60.59, CON 48.89, EXT 41.18, EMS 43.29, OPE 27.6, and Profession 14.49 (Sun et al., 2024).

For hierarchical text classification, HILL outperforms both flat and hierarchy-aware baselines. On WOS it reports 87.28 / 81.77 Micro-/Macro-F1, on RCV1-v2 87.31 / 70.12, and on NYTimes 80.47 / 69.96. Removing the contrastive term or replacing the entropy-optimized structure encoder with a trivial tree degrades performance, especially in Macro-F1, which suggests that the hierarchy-aware structural view improves rare or fine-grained label discrimination (Zhu et al., 2024).

In robotic imitation learning, explicit hierarchical structure improves long-horizon execution success. DMIL reaches 0.376 / 0.640 on ML45 meta-testing in 1-shot / 3-shot, above MIL 0.205 / 0.510 and OptionGAIL 0.220 / 0.481 (Gao et al., 2022). H-AIRL obtains 113.94 ± 2.54 episodic return on AntPush against 81.58 ± 39.75 for Option-AIRL and 80.23 ± 30.29 for H-GAIL (Chen et al., 2022). Bi-HIL reaches 100% success on Put-Three-Balls-in-Drawer, whereas Bi-ACT reaches 80%, and attains 80% success on both 6-Cup Downstack and 4-Peg-in-Hole, exceeding flat bilateral baselines (Buamanee et al., 4 Mar 2026).

6. Limitations, misconceptions, and research directions

A common misconception is that HIL names a single, agreed-upon module. The record is more heterogeneous. ID-Composer explicitly states that it does not introduce a separate acronym “HIL” or an explicit HIL module, even though its hierarchical attention mechanism implements a hierarchical identity model (Pan et al., 1 Nov 2025). In robotics, several influential papers use HIL to mean Hierarchical Imitation Learning rather than Hierarchical Identity Learning, and the corresponding latent variable is a skill or option index rather than a person, object, or semantic concept (Gao et al., 2022, Chen et al., 2022, Buamanee et al., 4 Mar 2026). The term therefore functions more as a family resemblance than a standardized formalism.

Another limitation is the prevalence of fixed or hand-specified hierarchy. HIRL uses a fixed number of levels $z^L$7 and hierarchical K-means prototypes (Xu et al., 2022). HIL-CBM restricts itself to a two-level hierarchy and uses GPT-4 to construct higher-level groupings and concept sets (Xie et al., 2 Apr 2026). The acoustic identification framework uses exactly three levels—taxon, species, and individual—and notes that deeper and more complex hierarchies remain open (Nolasco et al., 2024). USVI-ReID fixes the number of fine sub-clusters per coarse identity, with $z^L$8 on SYSU-MM01 and $z^L$9 on RegDB (Shi et al., 15 Sep 2025). Bi-HIL depends on manual subtask decomposition and annotated subtask boundaries (Buamanee et al., 4 Mar 2026).

Many approaches also rely on auxiliary structure that may itself be noisy or costly. HIRPF is trained on dialogues generated by ChatGPT and explicitly notes data realism as a limitation (Sun et al., 2024). HILL depends on an offline coding tree built from a known label hierarchy and structural entropy minimization (Zhu et al., 2024). The contextual Bayesian face model offers a principled nonparametric hierarchy over contexts and identities, but inference in nested DPs is computationally demanding and the paper characterizes the extended context model as preliminary (Castro et al., 2018).

The strongest forward directions suggested by the corpus are deeper hierarchies, adaptive hierarchy discovery, stronger open-set transfer, and multimodal generalization. HIRL proposes extensions to videos, text, and multimodal identity hierarchies (Xu et al., 2022). HIL-CBM suggests deeper taxonomic structures, tree-structured concept hierarchies, and multimodal extensions (Xie et al., 2 Apr 2026). The acoustic identification work identifies backbone fine-tuning, deeper hierarchies, and stronger few-shot open-set methods as natural next steps (Nolasco et al., 2024). USVI-ReID points toward adaptive numbers of sub-centers, differentiable clustering, and broader cross-modal retrieval tasks (Shi et al., 15 Sep 2025). Taken together, these works suggest that the durable contribution of HIL is not a single algorithmic recipe, but the claim that identity is often intrinsically multi-level and that models benefit when this structure is made explicit in representation, conditioning, and optimization.