Dynamic Prototypes in Adaptive Machine Learning

• Dynamic Prototypes are methods that replace fixed representatives with mutable models that adapt to data geometry, query context, and temporal evolution.
• They employ mechanisms like adaptive cardinality, query-conditioned adaptation, and memory-based updates to enhance performance in tasks such as few-shot learning and density estimation.
• This approach improves model robustness in diverse applications including deepfake detection, point cloud segmentation, and iterative UI prototyping by addressing domain shifts and multimodal challenges.

Dynamic Prototypes (DP) is an umbrella concept for methods in which prototypes are not treated as fixed, globally static representatives. Across recent literature, this dynamicity appears in several distinct forms: the number of prototypes can be inferred from data on a per-episode basis; support-derived prototypes can be adapted to a particular query; spatial prototypes can deform during matching; temporal or graph prototypes can evolve through memory and cross-domain updates; and, in a broader human-computer interaction usage, prototypes can be treated as revisable development artifacts rather than frozen deliverables (Allen et al., 2019, Liu et al., 2024, Donnelly et al., 2021, Ma et al., 2024). The term is therefore heterogeneous. It denotes a family resemblance centered on prototype mutability, but the mechanism of change differs substantially across few-shot learning, dense prediction, interpretable vision, temporal modeling, federated learning, graph anomaly detection, activation steering, and iterative UI prototyping.

1. Conceptual scope and recurring meanings

A useful synthesis is that the literature contains several recurrent modes of dynamicity rather than a single canonical DP formalism. Standard prototype methods typically assume one fixed representative per class or one fixed prototype bank learned once and then reused unchanged. Dynamic prototype methods relax that assumption by letting prototype cardinality, placement, structure, or use depend on the current episode, query, input, time step, domain, or interaction loop.

Dynamicity mode	Operational form	Representative papers
Adaptive cardinality	Spawn, delete, or infer prototype count from data	IMP (Allen et al., 2019), CHDQR-Dynamic (Cengiz et al., 2024)
Query-conditioned adaptation	Rectify or attend prototypes using the current query	DPA (Liu et al., 2024)
Input-adaptive structure	Deform part layouts or steer with instance-specific prototype projections	Deformable ProtoPNet (Donnelly et al., 2021), PDS (Zheng et al., 1 Aug 2025)
Temporal or sequential matching	Align prototype sequences to time-varying data	DPNet (Trinh et al., 2020), DP-DTW (Chang et al., 2021)
Memory-based evolution	Update prototype banks across tasks, clients, or domains	FedMLP (Guo et al., 2023), DP-DGAD (Zheng et al., 1 Aug 2025)
Interactive artifact evolution	Treat prototypes as revisable systems under iterative human control	DIDUP (Ma et al., 2024), DynEx (Ma et al., 2024)

This heterogeneity matters because papers differ on what exactly is dynamic. In some works, the prototype set itself is reconstructed from the current support set; in others, a global prototype bank remains fixed in cardinality but is used in an input-adaptive way; in still others, “dynamic” refers primarily to temporally structured inputs, not to online prototype state updates. The most precise reading is therefore mechanism-specific rather than slogan-based.

2. Adaptive prototype formation and prototype cardinality

One major DP line replaces the one-prototype-per-class assumption with data-dependent prototype multiplicity. “Infinite Mixture Prototypes” generalizes prototypical networks by representing each class with a variable number of clusters inferred from the support set in each episode (Allen et al., 2019). The method initializes one cluster per class, computes distances

$d_{i,c} = \begin{cases} \|h_{\phi}({x}_i) - {\mu}_c\|^2 & \text{if (%%%%0%%%% is labeled and %%%%1%%%% or %%%%2%%%% is unlabeled} \ +\infty & \text{otherwise} \end{cases}$

and spawns a new prototype whenever

$\min_c d_{i,c} > \lambda.$

It then refines assignments with Gaussian soft memberships and classifies a query using the closest cluster for each class rather than a single class mean. This yields prototype sets that are episode-conditioned, class-conditional, and nonparametric in cardinality. The strongest evidence for the value of this mechanism appears on Omniglot alphabets, where IMP improves 10-way 10-shot alphabet recognition from $65.6 \pm 0.4$ to $92.0 \pm 0.1$, a roughly 25% absolute gain over prototypical networks, while preserving strong performance on simpler uni-modal character tasks (Allen et al., 2019).

A second cardinality-adaptive line appears in conformalized high-density quantile regression. CHDQR-Dynamic replaces a fixed grid of output bins with a learned prototype set

$$C = \{c_1, c_2, \dots, c_K\}, \qquad c_i \in \mathbb{R}^d,$$

whose Voronoi regions define an adaptive partition of output space (Cengiz et al., 2024). The prototypes are not merely relocated by gradient descent; they are periodically added and deleted according to normalized usage

$$U(c_i) = \frac{1}{|\mathcal{D}_{tr}|} \sum_{y \in \mathcal{D}_{tr}} q_i(y,\tau).$$

Deletion uses

$$C \longleftarrow C - \{c_i\} \qquad \text{if } U(c_i) \le \delta_{del},$$

and addition uses

$$C \longleftarrow \text{concat}(C,\{c_{K+1}\}) \qquad \text{if } U(c_i) \ge \delta_{add},$$

with the new prototype sampled near the overloaded one. This produces a dynamically refined discretization for density estimation and conformal prediction. In 2D experiments, CHDQR-Dynamic uses about half as many prototypes as fixed 2500-bin baselines—1301 on Uncond2d and 1319 on Energy Efficiency—while maintaining comparable or slightly better PINAW and coverage (Cengiz et al., 2024).

These works share a common principle: prototype capacity is itself an adaptive variable. This suggests that one defining hallmark of DP, in the narrow meta-learning and density-estimation sense, is not prototype learning per se but prototype cardinality as a function of local data geometry.

3. Query-conditioned and input-adaptive prototypes

A second DP family keeps the prototype concept but makes the prototype state depend on the current query or input. In few-shot point cloud segmentation, Dynamic Prototype Adaptation starts from support-derived vanilla prototypes and turns them into query-specific prototypes through prototype rectification, prototype-to-query attention, and prototype distillation (Liu et al., 2024). Rectification is written as

$\mathbf{\hat{p}^i = \mathbf{p}^i+\mathbf{A}^i(\mathbf{p}^i)^\top,$

and cross-attentive adaptation as

$\mathbf{\dot{p}^i = \text{softmax}(\mathbf{Q}\mathbf{K}^\top)V +\mathbf{\hat{p}^i},$

followed by an FFN. The resulting task-specific prototypes are then matched against query point features. The paper reports large gains over the previous prototype-adaptation baseline QGPA, including $\min_c d_{i,c} > \lambda.$0 points on S3DIS and $\min_c d_{i,c} > \lambda.$1 points on ScanNet under the 2-way 1-shot setting, and the ablation on S3DIS 2-way 1-shot shows a progression from 50.79 mean IoU without adaptation to 64.49 with Prototype Rectification, 67.54 with Prototype-to-Query Attention, and 70.19 with full DPA including Prototype Distillation (Liu et al., 2024).

Deformable ProtoPNet pushes input-adaptivity into the internal structure of each prototype (Donnelly et al., 2021). Instead of treating a prototype as one rigid latent patch, it decomposes the prototype into prototypical parts and predicts input-dependent offsets for those parts during matching: $\min_c d_{i,c} > \lambda.$2 The prototype is therefore dynamic in a spatial sense: its effective arrangement depends on the input image. This improves both invariance to pose and the granularity of “this looks like that” explanations. Reported accuracies include 86.4% on CUB with 2×2 prototypes and 86.5% on Stanford Dogs, outperforming rigid ProtoPNet variants in the reported comparisons (Donnelly et al., 2021).

Prototype-Based Dynamic Steering for LLMs uses prototypes in activation space rather than feature space of images or point clouds (Zheng et al., 1 Aug 2025). Here, “reasoning prototypes” are k-means centroids of activation differences between Chain-of-Thought and neutral prompts. At inference, the hidden state of a new prompt is projected onto these prototypes to form an instance-specific steering vector: $\min_c d_{i,c} > \lambda.$3 This is dynamic because the steering direction depends on the current hidden state rather than a global difference-of-means vector. The paper reports consistent gains on GSM8K and AQuA-RAT, with GSM8K improving from 68% to 78% in the Neutral condition and from 11% to 21% in the Anti-CoT condition (Zheng et al., 1 Aug 2025).

Across these works, dynamicity resides in conditioning. The prototype bank is not merely learned once; it is reinterpreted, rectified, deformed, or projected in response to the current sample.

4. Temporal, sequential, graph, and memory-based dynamic prototypes

Temporal data introduce another form of dynamicity: prototypes become templates for evolving patterns rather than static snapshots. In deepfake detection, DPNet learns prototype vectors in latent patch space over clip representations formed from one RGB frame plus 10 optical-flow fields (Trinh et al., 2020). Prototype similarity is computed as

$\min_c d_{i,c} > \lambda.$4

so each prototype refers to a short localized dynamic pattern rather than a single-frame feature. The method couples this with temporal logic specifications over prototype activations, thereby linking prototype evidence to formal post hoc verification. On cross-dataset evaluation, DPNet reaches AUC 92.44 on Google DeepFakeDetection and 90.80 on DeeperForensics after training on FF++, while maintaining interpretable prototype-based explanations (Trinh et al., 2020).

ProtoASNet also treats prototypes as spatio-temporal entities, but in echocardiography video classification (Vaseli et al., 2023). It constructs a prototype-specific pooled feature

$\min_c d_{i,c} > \lambda.$5

then compares it to a learned prototype with cosine similarity. Dynamicity here comes from the explicit time dimension in $\min_c d_{i,c} > \lambda.$6 and from prototype-specific spatio-temporal ROI maps $\min_c d_{i,c} > \lambda.$7. The model also introduces uncertainty prototypes $\min_c d_{i,c} > \lambda.$8 with abstention loss. On the private AS dataset, the video version improves study-level balanced accuracy from 73.9 for ProtoASNet (Image) to 80.0 for ProtoASNet (Video), and on TMED-2 the reported accuracy is 79.7% (Vaseli et al., 2023).

For sequence alignment, DP-DTW learns class-specific temporal prototype sequences and matches them through Dynamic Time Warping rather than rigid pointwise comparison (Chang et al., 2021). For time-series classification it minimizes

$\min_c d_{i,c} > \lambda.$9

where $65.6 \pm 0.4$0 is the DTW discrepancy between sample and class prototype. For weakly supervised action segmentation, action prototypes are concatenated according to transcripts and compared against video features with a margin-based discriminative DTW objective. The paper reports that DP-DTW is the best-ranked method across 128 UCR datasets and improves over D$65.6 \pm 0.4$1TW by 5.1% frame accuracy on Breakfast and 22.0% on Hollywood in the segmentation setting (Chang et al., 2021).

Graph and federated settings emphasize memory and cross-domain evolution. FedMLP maintains local class prototypes, global class prototypes, local semantic prototypes, global semantic prototypes, and previous-task local prototypes as a multi-level prototype memory under Dynamic Heterogeneous Federated Learning (Guo et al., 2023). These prototypes are recomputed, aggregated, clustered, and reused across task stages to mitigate catastrophic forgetting and concept drift. On CIFAR-100 global accuracy, FedMLP reaches 61.02 versus 57.09 for FedProto and 37.03 for FedNTD, and the ablation shows that removing any one of the prototype regularizers degrades both local and global balanced accuracy (Guo et al., 2023).

DP-DGAD uses a memory buffer of prototype pairs $65.6 \pm 0.4$2 for normal and anomalous dynamic graph patterns (Zheng et al., 1 Aug 2025). Prototypes are extracted from temporal ego-graph edge embeddings, ranked by a discriminability score

$65.6 \pm 0.4$3

and a cross-domain similarity score

$65.6 \pm 0.4$4

combined as

$65.6 \pm 0.4$5

The buffer is selectively updated to retain general patterns while incorporating new domain-specific ones, and anomaly scoring is distributional rather than purely nearest-prototype. The ablation shows substantial degradation when dynamic prototype anomaly scoring, domain-adaptive buffer ranking, or confidence-guided adaptation are removed (Zheng et al., 1 Aug 2025).

ProtoFlow lies somewhat differently in this landscape (Holm et al., 16 Dec 2025). It learns three prototypes per class over dynamic scene graph embeddings for surgical workflow recognition. Its prototypes are optimized jointly with the encoder and initialized by classwise clustering after graph autoencoder pretraining. The “dynamic” aspect is mainly that prototypes summarize temporally aggregated dynamic scene graphs rather than that the prototypes themselves are explicitly time-indexed or updated online at test time. This nuance is important: ProtoFlow is a prototype model over dynamic graph inputs, not a recurrent prototype state-space model (Holm et al., 16 Dec 2025).

5. Dynamic prototypes as iterative development artifacts

In a broader HCI usage, Dynamic Prototypes refers not to latent class representatives but to prototypes as living artifacts under revision. DIDUP frames code-based UI prototyping as an iterative spiral process rather than a waterfall pipeline (Ma et al., 2024). It introduces three mechanisms: adaptive planning, code injection, and lightweight state management. Adaptive planning allows users to add, update, or remove tasks during implementation; code injection writes minimal snippets and inserts them instead of rewriting the whole codebase; lightweight state management saves the state of the UI code at every task so users can quickly revert to different working states. The paper’s formative comparison against GPT Pilot describes the baseline as “essentially a waterfall model,” whereas DIDUP reportedly let both users create working prototypes easily, add new features, backtrack to make major adjustments, and produce more complete UIs with more grouping and stylistic elements (Ma et al., 2024).

DynEx extends this broader notion by treating exploratory programming as dynamic traversal of a design space before and during code generation (Ma et al., 2024). Its central construct is the Design Matrix, a 3-by-2 representation with dimensions Person, Approach, and Interaction at Idea and Grounding levels. Prototype generation is then staged through requirements inference, specification generation, placeholder data synthesis, task decomposition, sequential code generation, user approval, debugging, rollback, and branching into multiple variants. In a study of 10 programmers, DynEx increased Divergent Thinking from 4.2 to 6.1, Convergent Thinking from 3.4 to 5.9, and Application Complexity from 3.7 to 4.9 relative to a Claude Artifact baseline, while Idea Realization showed no significant difference, 5.2 versus 5.5 (Ma et al., 2024).

This broader usage suggests that the phrase Dynamic Prototypes can also denote an interaction architecture: prototypes are dynamic when plans remain editable, revisions are local and legible, rollback is cheap, and alternative branches can be explored without restarting. That interpretation is distinct from prototype learning in representation space, but it preserves the same core idea of revisable structure.

6. Empirical profile, limitations, and terminological ambiguities

Empirically, the strongest DP gains tend to appear when the target phenomenon is intrinsically multimodal, temporally elastic, query-mismatched, or under domain shift. IMP yields its clearest advantage on multi-modal alphabet recognition rather than uni-modal character recognition (Allen et al., 2019). DPA improves especially when support-query variation is large in point cloud segmentation (Liu et al., 2024). DPNet and ProtoASNet benefit from explicitly temporal prototypes rather than static ones (Trinh et al., 2020, Vaseli et al., 2023). DP-DTW is strongest where temporal warping is intrinsic (Chang et al., 2021). DP-DGAD and FedMLP derive much of their benefit from evolving environments rather than static IID data (Zheng et al., 1 Aug 2025, Guo et al., 2023). CHDQR-Dynamic is most useful when fixed quantization becomes wasteful in multimodal or higher-dimensional output spaces (Cengiz et al., 2024). PDS is strongest when latent reasoning capacity exists but is not fully expressed by the surface prompt, especially in Neutral and Anti-CoT conditions (Zheng et al., 1 Aug 2025).

Recurring limitations are equally visible. Many methods are task-specialized: DPA is built for few-shot point cloud segmentation; ProtoASNet for echocardiography cine videos; DPNet for deepfake detection; DP-DTW for temporal classification and weakly supervised action segmentation. Several papers note under-specified or partly differentiable mechanisms: IMP uses hard cluster spawning and indirect optimization of $65.6 \pm 0.4$6 (Allen et al., 2019), ProtoFlow leaves some clustering and node-level alignment details underspecified (Holm et al., 16 Dec 2025), and DIDUP does not fully specify prompt structures or insertion heuristics (Ma et al., 2024). Evaluations are sometimes small or preliminary, especially in interactive prototyping work (Ma et al., 2024). Query-conditioned or memory-based prototype systems also introduce extra complexity, parameter sensitivity, and possible instability in adaptation (Liu et al., 2024, Zheng et al., 1 Aug 2025).

The term itself is also ambiguous. “A Deep Positive-Negative Prototype Approach to Integrated Prototypical Discriminative Learning” is only partially relevant to Dynamic Prototypes in the narrow sense because its prototypes are global class-wise parameters updated by SGD rather than query-conditioned or episodically constructed (Zarei-Sabzevar et al., 5 Jan 2025). “Differentially Private Prototypes for Imbalanced Transfer Learning” is prototype-based but not a dynamic-prototype method; here DP means Differential Privacy, and prototypes are computed once from a fixed private dataset or selected once from a public pool (Wahdany et al., 2024). “Dynamic Maintenance of Monotone Dynamic Programs and Applications” is unrelated to prototype learning; there, DP means dynamic programming (Henzinger et al., 2023). Even among central papers, “dynamic” can refer either to evolving prototype state, as in FedMLP or DP-DGAD, or primarily to dynamic input structure, as in ProtoFlow and ProtoASNet.

The most stable encyclopedic conclusion is therefore structural rather than lexical. Dynamic Prototypes denotes a family of methods in which prototype representations are made responsive to data geometry, query context, temporal evolution, domain shift, or iterative human redesign. What unifies the family is not a single algorithm, but a rejection of the fixed-prototype assumption. In the narrow learning-theoretic sense, the most canonical DP mechanisms are adaptive cardinality, query-conditioned adaptation, and memory-updated prototype evolution (Allen et al., 2019, Liu et al., 2024, Guo et al., 2023, Zheng et al., 1 Aug 2025). In the broader systems sense, DP extends to interactive prototyping workflows in which the prototype remains interruptible, revisable, and branchable throughout development (Ma et al., 2024, Ma et al., 2024).