Prototype-Based Knowledge Modulation

• Prototype-based knowledge modulation is an approach that uses compact prototypes as reusable knowledge primitives to steer representation learning across diverse domains.
• It employs modulation operators like generation, evolution, alignment, and retrieval to adjust representations and enhance model performance.
• Applications span federated learning, few-shot tasks, and medical imaging, demonstrating practical gains in accuracy, robustness, and effective knowledge transfer.

Searching arXiv for the cited papers and closely related prototype-based methods to ground the article in current literature. Prototype-based knowledge modulation denotes a class of methods in which a model stores, exchanges, retrieves, evolves, or composes knowledge through prototypes and then uses those prototypes to alter representation learning, decision boundaries, or symbolic inheritance. In recent arXiv literature, a prototype may be a low-dimensional class vector summarizing an embedding space, a task slot in a learned prototype matrix, a modality-level class centroid, a trainable server-side class representative, an answer-option visual exemplar bank, or a symbolic object defined by inheritance and override rules. Despite this heterogeneity, the common pattern is stable: prototypes act as compact carriers of prior structure, and downstream computation is modulated by distances, affinities, residual corrections, or inheritance operations rather than by raw observations alone (Wang et al., 13 Feb 2025, Chennoufi et al., 7 Jul 2025, Oh et al., 12 Jan 2026, Chen et al., 2024, Hossen et al., 26 Aug 2025, Pellegrini et al., 12 Mar 2026, Cochez et al., 2016).

1. Conceptual forms of prototypes

Across current work, the notion of a prototype is instantiated in several technically distinct ways. In vertical federated continual learning, each class is represented by a low-dimensional prototype vector $\mu$ summarizing the joint embedding of all parties’ features, with

$$\mu_c^t = \frac{1}{|D_c|}\sum_{i:y_i=c} E_i^t, \qquad E_i^t=\sum_{k=2}^K E_{ik}^t.$$

In federated intrusion detection, a client-level class prototype is the mean embedding

$$C_{i,j}^t=\frac{1}{N_{i,j}}\sum_{x\in D_i:\,\mathrm{label}=j}\phi(x),$$

and the server forms global prototypes by simple averaging. In partially annotated multi-task learning, prototypes are organized as a task prototype matrix

$V=[v_1;v_2;\dots;v_T]\in\mathbb{R}^{T\times d},$

with each slot intended to capture the “essence” of one task. In multi-modality semi-supervised segmentation, each modality $k$ maintains class-wise prototypes $p_c^k\in\mathbb{R}^D$. In structured radiology reporting, each answer option $\ell$ is represented by a visual prototype

$$p_\ell=\max_{1\le k\le K} f_{\mathrm{img}}(x_{\ell,k}),$$

constructed from up to $K=5$ exemplar images (Wang et al., 13 Feb 2025, Chennoufi et al., 7 Jul 2025, Oh et al., 12 Jan 2026, Chen et al., 2024, Pellegrini et al., 12 Mar 2026).

Setting	Prototype form	Operational role
V-LETO	$\mu_c^t$, $$\mu_c^t = \frac{1}{|D_c|}\sum_{i:y_i=c} E_i^t, \qquad E_i^t=\sum_{k=2}^K E_{ik}^t.$$0	preserve class and feature knowledge
PROTEAN	$$\mu_c^t = \frac{1}{|D_c|}\sum_{i:y_i=c} E_i^t, \qquad E_i^t=\sum_{k=2}^K E_{ik}^t.$$1, $$\mu_c^t = \frac{1}{|D_c|}\sum_{i:y_i=c} E_i^t, \qquad E_i^t=\sum_{k=2}^K E_{ik}^t.$$2	client-server knowledge sharing
Task prototype retrieval	$$\mu_c^t = \frac{1}{|D_c|}\sum_{i:y_i=c} E_i^t, \qquad E_i^t=\sum_{k=2}^K E_{ik}^t.$$3	quantify task associations
DBDC	$$\mu_c^t = \frac{1}{|D_c|}\sum_{i:y_i=c} E_i^t, \qquad E_i^t=\sum_{k=2}^K E_{ik}^t.$$4	modality-specific knowledge bank
ProtoSR	$$\mu_c^t = \frac{1}{|D_c|}\sum_{i:y_i=c} E_i^t, \qquad E_i^t=\sum_{k=2}^K E_{ik}^t.$$5	answer-option visual memory
Web prototype ontologies	$$\mu_c^t = \frac{1}{|D_c|}\sum_{i:y_i=c} E_i^t, \qquad E_i^t=\sum_{k=2}^K E_{ik}^t.$$6	symbolic inheritance and override

The same term extends beyond vector centroids. In web knowledge representation, a prototype is a 4-tuple

$$\mu_c^t = \frac{1}{|D_c|}\sum_{i:y_i=c} E_i^t, \qquad E_i^t=\sum_{k=2}^K E_{ik}^t.$$7

whose denotation is defined recursively by

$$\mu_c^t = \frac{1}{|D_c|}\sum_{i:y_i=c} E_i^t, \qquad E_i^t=\sum_{k=2}^K E_{ik}^t.$$8

In structured dynamical systems, “schemata” function as prototypes of procedural knowledge: shared parameter sets $$\mu_c^t = \frac{1}{|D_c|}\sum_{i:y_i=c} E_i^t, \qquad E_i^t=\sum_{k=2}^K E_{ik}^t.$$9 prescribe state transitions for multiple object files, thereby separating declarative object state from reusable dynamics (Cochez et al., 2016, Goyal et al., 2020).

A common misconception is that prototype-based methods are limited to nearest-centroid classification. The current literature shows otherwise. Prototypes may be fixed summaries, trainable parameters, anchor libraries selected by minimal intra-class variance, prompt-conditioned semantic carriers, or symbolic inheritance objects. This suggests that “prototype” is better understood as a reusable knowledge primitive than as a single geometric operation (Zhu et al., 27 Nov 2025, Hossen et al., 26 Aug 2025).

2. Modulation operators and knowledge flow

Prototype-based modulation is implemented through a recurring set of operators: generation, evolution, alignment, retrieval, enhancement, and override. In V-LETO, Prototype Generation computes class means from aggregated party embeddings, and Prototype Evolving then updates the global prototype repository through Class Evolving and Feature Evolving. For missing past classes $$C_{i,j}^t=\frac{1}{N_{i,j}}\sum_{x\in D_i:\,\mathrm{label}=j}\phi(x),$$0, Class Evolving adjusts the stored global prototype by

$$C_{i,j}^t=\frac{1}{N_{i,j}}\sum_{x\in D_i:\,\mathrm{label}=j}\phi(x),$$1

with cosine similarity

$$C_{i,j}^t=\frac{1}{N_{i,j}}\sum_{x\in D_i:\,\mathrm{label}=j}\phi(x),$$2

For classes already present, Feature Evolving blends old and new information through

$$C_{i,j}^t=\frac{1}{N_{i,j}}\sum_{x\in D_i:\,\mathrm{label}=j}\phi(x),$$3

The result is a prototype repository spanning past and present classes and features (Wang et al., 13 Feb 2025).

In PROTEAN, modulation is primarily alignment-based. Clients send $$C_{i,j}^t=\frac{1}{N_{i,j}}\sum_{x\in D_i:\,\mathrm{label}=j}\phi(x),$$4 to the server; the server averages both model parameters and prototypes, then broadcasts $$C_{i,j}^t=\frac{1}{N_{i,j}}\sum_{x\in D_i:\,\mathrm{label}=j}\phi(x),$$5 back. Each client is then pulled simultaneously toward the global model and toward the global prototypes of attack classes. Because the prototype alignment term acts in embedding space, clients can be drawn toward representations of classes missing from their local data, enabling few-/zero-shot knowledge transfer under highly non-IID distributions (Chennoufi et al., 7 Jul 2025).

Task Prototype-Based Knowledge Retrieval implements modulation through affinities and cross-attention. Given flattened task-specific features $$C_{i,j}^t=\frac{1}{N_{i,j}}\sum_{x\in D_i:\,\mathrm{label}=j}\phi(x),$$6, cosine similarities $$C_{i,j}^t=\frac{1}{N_{i,j}}\sum_{x\in D_i:\,\mathrm{label}=j}\phi(x),$$7 are converted to row-wise affinities $$C_{i,j}^t=\frac{1}{N_{i,j}}\sum_{x\in D_i:\,\mathrm{label}=j}\phi(x),$$8, and the task-affinity feature is formed as

$$C_{i,j}^t=\frac{1}{N_{i,j}}\sum_{x\in D_i:\,\mathrm{label}=j}\phi(x),$$9

A Knowledge Retrieval Transformer then applies self-attention to $V=[v_1;v_2;\dots;v_T]\in\mathbb{R}^{T\times d},$0 and cross-attention from the refined feature to the prototype-derived keys and values. Modulation therefore occurs per spatial location and is conditioned by estimated task associations rather than by pseudo-labels from unlabeled tasks (Oh et al., 12 Jan 2026).

Other systems realize the same principle with different operators. HAPM begins with a variance spectrum-driven anchor prototype library and then applies Pathological Semantic Injector and Discriminative Prototype Enhancer modules, both implemented through cross-attention from visual prototypes to gated semantic prompts. ProtoSR retrieves only those prototypes whose labels are valid answer options for the current question, computes retrieval weights

$V=[v_1;v_2;\dots;v_T]\in\mathbb{R}^{T\times d},$1

and adds a prototype-conditioned residual to the base logits:

$V=[v_1;v_2;\dots;v_T]\in\mathbb{R}^{T\times d},$2

In web prototype ontologies, modulation is symbolic rather than metric: new knowledge is inserted by constructing a prototype that extends a base with fresh [ADD](https://www.emergentmind.com/topics/adversarial-diffusion-distillation-add) and [REM](https://www.emergentmind.com/topics/random-energy-model-rem) sets, and lookup resolves inheritance by fixed-point computation and memoization (Zhu et al., 27 Nov 2025, Pellegrini et al., 12 Mar 2026, Cochez et al., 2016).

3. Optimization regimes

Prototype-based modulation is not defined solely by representation; it is also defined by the losses that enforce prototype utility. In V-LETO, the server combines fresh-data supervision with replay-like prototype supervision:

$V=[v_1;v_2;\dots;v_T]\in\mathbb{R}^{T\times d},$3

where $V=[v_1;v_2;\dots;v_T]\in\mathbb{R}^{T\times d},$4 re-predicts from adjusted missing-class prototypes and $V=[v_1;v_2;\dots;v_T]\in\mathbb{R}^{T\times d},$5 re-predicts from blended current-class prototypes. Forgetting on passive parties is further mitigated by a Fisher-based freezing rule. For each local parameter $V=[v_1;v_2;\dots;v_T]\in\mathbb{R}^{T\times d},$6,

$V=[v_1;v_2;\dots;v_T]\in\mathbb{R}^{T\times d},$7

with threshold

$V=[v_1;v_2;\dots;v_T]\in\mathbb{R}^{T\times d},$8

Parameters with $V=[v_1;v_2;\dots;v_T]\in\mathbb{R}^{T\times d},$9 are frozen; the remainder are updated by gradient descent (Wang et al., 13 Feb 2025).

PROTEAN uses a joint client objective

$k$0

where $k$1. The first term is standard cross-entropy, the second aligns local and global class prototypes, and the third is a FedProx-style proximal regularizer. Here prototype modulation and model-drift control are explicitly coupled (Chennoufi et al., 7 Jul 2025).

In partially annotated multi-task learning, the total objective is

$k$2

with

$k$3

Task Knowledge Embedding forces affinity mass onto the correct task slot when only that task is labeled, while Task Consistency imposes a margin between intra-task and inter-task similarity. The stated goal is to ensure that each prototype slot specializes in its own task while remaining distinguishable from the others (Oh et al., 12 Jan 2026).

Several adjacent lines of work use prototype losses to regularize different structures. DBDC defines

$k$4

combining prototype-segmentation, inter-modality contrast, inter-class contrast within a modality, and intra-class compactness. FedProtoKD defines a server-side objective

$k$5

where $k$6 trains server prototypes with adaptive class-wise margins. ProC-KD jointly optimizes embedding-layer distillation, prototype learning, and student supervision through

$k$7

Prototype Completion with Primitive Knowledge adds a Gaussian-based prototype fusion step that treats one prototype estimate as a prior and another as a likelihood, producing a closed-form posterior mean $k$8 and variance $k$9 (Chen et al., 2024, Hossen et al., 26 Aug 2025, Li et al., 2022, Zhang et al., 2020).

4. Federated and distributed learning

Federated learning has become a primary site for prototype-based knowledge modulation because prototypes can be exchanged without sharing raw data. V-LETO addresses the conjunction of Vertical Federated Learning with class-incremental and feature-incremental learning. Its Evolving Prototype Knowledge module carries forward past-task information through Prototype Generation and Prototype Evolving, while constrained local optimization freezes Fisher-important passive-party parameters. On MNIST, FMNIST, CIFAR-10, and CINIC-10, the method is reported to outperform Pyvertical, Pass+VFL, FedSpace+VFL, and FedProK+VFL. For CIFAR-10, a representative result gives 52.26% CIL average accuracy and 88.34% FIL task-4 accuracy, with stated gains of 10.39% for CIL and 35.15% for FIL over the state of the art (Wang et al., 13 Feb 2025).

PROTEAN addresses horizontal federated intrusion detection under severe heterogeneity by exchanging class prototypes of attack types in addition to model parameters. The evaluation uses X-IIoTID and 5G-NIDD with Dirichlet $p_c^k\in\mathbb{R}^D$0 over $p_c^k\in\mathbb{R}^D$1 clients, Accuracy, Macro-Accuracy, F1, Precision, per-round convergence, and rare-class accuracy. The reported effects are specific: PROTEAN yields 5–23 pp higher macro-F1 than SOTA FedAvg-based methods when $p_c^k\in\mathbb{R}^D$2; rare classes absent locally gain 30–80 pp in accuracy over Cerberus; clients detect unseen classes with $p_c^k\in\mathbb{R}^D$3 accuracy versus $p_c^k\in\mathbb{R}^D$4 locally; and stable macro-accuracy is reached in 2 rounds versus 5+ for FedAvg. Privacy analysis further reports that a semi-honest server can formulate a reconstruction attack, but reconstruction MSE is large, and adding differential-privacy noise further increases MSE with $p_c^k\in\mathbb{R}^D$5 F1 drop (Chennoufi et al., 7 Jul 2025).

FedProtoKD targets heterogeneous federated learning, where fixed weighted averaging of client prototypes is said to shrink prototype margins. Instead of accepting

$p_c^k\in\mathbb{R}^D$6

the method introduces trainable server prototypes

$p_c^k\in\mathbb{R}^D$7

learned with class-wise adaptive margin

$p_c^k\in\mathbb{R}^D$8

The server also computes public-sample importance from prototype closeness and combines logits-based and prototype-based distillation. Reported results include average gains from 1.13% up to 34.13%, server accuracy of 63.14% on CIFAR-10 with Dirichlet $p_c^k\in\mathbb{R}^D$9 versus 54.84% for Fed2PKD and 62.58% for FedPKD, minimum inter-class prototype margins rising from about 12 to over 30, and client-side gains up to 15.1% under extreme non-IID splits (Hossen et al., 26 Aug 2025).

Distributed prototype mechanisms are not limited to neural federated learning. In web-scale prototype ontologies, a RemoteKB fetches prototype definitions over HTTP, ChainedKB queries multiple sources in turn, and memoized fixed-point computation supports reuse of inherited states. Benchmarks on synthetic families with up to 4 million prototypes show linear scaling in consistency checks and fixpoint computation, while a simulated WAN with 100 parallel requests yields approximately 50 fixpoints/sec (Cochez et al., 2016).

5. Partial annotation, few-shot learning, and transfer

Prototype-based modulation is also prominent when supervision is sparse, incomplete, or transferred across tasks. In Task Prototype-Based Knowledge Retrieval for partially annotated multi-task learning, the prototype matrix $\ell$0 is optimized so that each slot captures task-specific characteristics, and a Knowledge Retrieval Transformer injects prototype-derived affinities into task features. On PASCAL-Context in the one-label setting, the reported method reaches 59.78 mIoU for semantic segmentation, 59.08 for parsing, 78.62 for saliency, 15.63 mean angular error for normals, and 65.10 odsF for boundary detection. On NYUD-v2 in the one-label setting, it reports 45.95 mIoU, 0.4865 depth absolute error, and 25.64 normal error. The ablation on PASCAL-Context is especially direct: the semantic-segmentation score moves from 44.73 without VQ or AKG, to 44.83 with TAE only, to 58.21 with TAE+TKE, and to 59.78 with full AKG (Oh et al., 12 Jan 2026).

In few-shot learning, Prototype Completion with Primitive Knowledge argues that novel-class samples in a pre-trained feature space have large intra-class variance and that improving prototype estimates is more effective than fine-tuning the feature extractor. Primitive knowledge is encoded as attribute-feature priors

$\ell$1

processed by ProtoComNet, and then fused with the raw mean prototype by a Gaussian product rule. The reported effects are 2–9% absolute gains over strong baselines on miniImageNet, tieredImageNet, and CUB, higher prototype-to-ground-truth cosine similarity, and robustness to noise in the attribute matrix $\ell$2 (Zhang et al., 2020).

Few-shot action recognition extends the same logic to multimodal prompt conditioning. CLIP-SPM uses Hierarchical Synergistic Motion Refinement, Semantic Prototype Modulation, and Prototype-Anchor Dual Modulation. In the CLIP-ViT-B variant, reported 1/3/5-shot accuracies are 78.2/86.3/88.6 on HMDB51, 96.2/98.2/98.7 on UCF101, 92.8/94.1/94.3 on Kinetics, 66.7/74.8/77.3 on SSv2-Full, and 57.8/62.4/66.2 on SSv2-Small. The ablation states that adding SPM yields a large jump, with 1-shot gains of 9–10 points, and PADM provides further improvement in inter-class separation (Li et al., 22 Dec 2025).

Prototype-guided cross-task knowledge distillation broadens the idea to teacher-student transfer across disparate label spaces. ProC-KD learns teacher-side local pattern prototypes $\ell$3 and uses student-side prototype attention for feature augmentation. The reported effects include Office→ClipArt accuracy increasing from approximately 40% to approximately 41.3%, long-tailed CIFAR-100 improving from 72.7% to 78.3%, an ablation gain of +1.6 mAP on FoggyCityscapes when both modules are used, and a +1.3% boost over REFILLED in same-task KD on CIFAR-100 (Li et al., 2022).

Transfer prototype-based fuzzy clustering shows that prototype modulation also has an optimization-theoretic form outside deep learning. T-FCM augments the FCM objective with a source-prototype supervised term and a centroid regularization term,

$\ell$4

and analogous formulations are given for T-FSC and T-FKPC. Reported gains include TFCM improving NMI on T1 from 0.656 to 0.785 and RI from 0.860 to 0.919, while TFSC on S4→T4 reaches approximately 1.00 for both NMI and RI versus 0.708 and 0.822 for FSC (Deng et al., 2014).

6. Medical imaging and structured reporting

Medical imaging has adopted prototype-based modulation both for cross-center robustness and for fine-grained decision support. DBDC addresses multi-modality semi-supervised medical image segmentation through two plug-in banks: the Modality-Level Modulation Bank, which stores per-modality affine normalization parameters and EMA statistics, and the Modality-Level Prototype Bank, which stores class-wise modality prototypes updated by Sinkhorn-based assignments and EMA. The total learning objective combines supervised segmentation, Modality Adaptive Weighting, Dual Consistency, and MPCL. On a 2-modality heart task with 20% labeled data, the reported mean performance is DSC 84.56% versus 75.17% for SUMML and 74.76% for MASS, with HD$\ell$5 4.21 mm versus 6.28 and 6.18 mm. On a 4-modality abdomen task with 20% labeled data, the reported mean DSC is 87.61% versus 79.87% for CPS and 81.50% for EVIL, with HD$\ell$6 4.62 mm versus 11.91 and 14.06 mm. The framework is also described as lightweight, adding 0.05M parameters per modality with FLOPs comparable to single U-Net (Chen et al., 2024).

HAPM for multicenter diabetic retinopathy diagnosis uses prototypes as clinically grounded anchors rather than as simple class means. For each grade $\ell$7, it selects $\ell$8 anchor images by minimizing intra-class feature variance and initializes

$\ell$9

A hierarchical differential prompt gating mechanism then selects discriminative semantic prompts from LVLM and LLM sources, after which PSI and DPE modulate prototypes by cross-attention. Across EyePACS, Messidor, IDRiD, APTOS, DeepDR, FGADR, RLDR, and DDR, the reported ESDG average is 50.1 ACC and 37.7 F1, compared with 46.7 and 32.2 for GDRNet; Leave-One-Out DG average F1 is 58.9% versus 52.4% for GDRNet and 52.1% for CLIP-DR. On the APTOS ablation, performance progresses from 24.6% for the base system to 55.3% for full HAPM (Zhu et al., 27 Nov 2025).

ProtoSR applies prototype-based guidance to structured radiology reporting by constructing a multimodal knowledge base from 80k+ MIMIC-CXR studies using an instruction-tuned LLM and then retrieving answer-option prototypes relevant to an image-question pair. The retrieved evidence yields a residual “second opinion” rather than replacing the base prediction. On Rad-ReStruct, ProtoSR reports best overall F$$p_\ell=\max_{1\le k\le K} f_{\mathrm{img}}(x_{\ell,k}),$$0=34.4% versus 32.9% for Context-VQA, with the largest gain at the detailed attribute level: L3-F$$p_\ell=\max_{1\le k\le K} f_{\mathrm{img}}(x_{\ell,k}),$$1 improves from 3.2% to 7.4%. L2-F$$p_\ell=\max_{1\le k\le K} f_{\mathrm{img}}(x_{\ell,k}),$$2 moves from 71.8% to 72.8%, while L1-F$$p_\ell=\max_{1\le k\le K} f_{\mathrm{img}}(x_{\ell,k}),$$3 remains competitive at 66.2% versus 67.2%. The ablation is also informative: early fusion of prototypes into the prompt gives 32.5% overall F$$p_\ell=\max_{1\le k\le K} f_{\mathrm{img}}(x_{\ell,k}),$$4, randomized prototypes give 32.7%, and retrieval with late residual fusion gives 34.4% (Pellegrini et al., 12 Mar 2026).

These medical systems clarify an important point. Prototype-based modulation is not confined to generic class discrimination; it can integrate pathology descriptions, modality-specific priors, or free-text-derived exemplar knowledge into highly structured clinical decision pipelines. This suggests that prototypes are increasingly being used as interfaces between heterogeneous knowledge sources rather than merely as summary statistics.

7. Limitations, misconceptions, and open directions

Several limitations recur in the literature. In task-prototype retrieval for partially annotated MTL, prototypes are fixed to the $$p_\ell=\max_{1\le k\le K} f_{\mathrm{img}}(x_{\ell,k}),$$5 tasks seen at training time, so zero-shot addition of new tasks is not directly supported; the method also introduces extra parameters in the form of prototype slots and transformer heads, which may be non-trivial for extremely lightweight settings (Oh et al., 12 Jan 2026). In web prototype ontologies, there is no native support for more complex OWL constructs such as cardinality and restrictions, conflict-resolution semantics are left to the application layer, remove-all semantics are implemented as a special case, and current benchmarks are synthetic rather than real-world prototype datasets (Cochez et al., 2016).

Privacy and robustness remain active issues. PROTEAN explicitly notes that a semi-honest server might attempt a reconstruction attack by solving

$$p_\ell=\max_{1\le k\le K} f_{\mathrm{img}}(x_{\ell,k}),$$6

although the reported reconstruction MSE is large and differential-privacy noise further degrades reconstruction while preserving most F1 (Chennoufi et al., 7 Jul 2025). FedProtoKD identifies a different failure mode, prototype-margin shrinking under weighted server averaging, and responds with adaptive class-wise margins and contrastive server-prototype training (Hossen et al., 26 Aug 2025).

A second misconception is that prototype methods eliminate negative transfer by construction. The surveyed work does not support that claim. Several systems instead add explicit safeguards: Fisher-based freezing in V-LETO, proximal regularization in PROTEAN, Task Consistency in AKG, adaptive class-wise margins in FedProtoKD, and Gaussian fusion in Prototype Completion with Primitive Knowledge. This suggests that prototypes are useful but not self-sufficient; they typically require auxiliary constraints to remain discriminative, stable, or privacy-preserving (Wang et al., 13 Feb 2025, Zhang et al., 2020).

The future directions stated in the literature are correspondingly concrete. Proposed extensions include zero-shot task extension, hierarchical prototypes, cross-modal prototypes, and self-supervised pre-training for task-prototype retrieval; adaptive weighting of source versus target knowledge for transfer clustering; and SPARQL-like query interfaces together with closer alignment to mainstream Linked Data vocabularies for prototype-based web knowledge bases (Oh et al., 12 Jan 2026, Deng et al., 2014, Cochez et al., 2016). Collectively, these directions indicate that prototype-based knowledge modulation is evolving from a local representation trick into a broader design pattern for controllable knowledge transfer across statistical, semantic, and symbolic regimes.