Class-Aware Prompting Strategy

• Class-aware prompting is a strategy that injects class-specific information into neural network inputs using learnable tokens, templates, or principles.
• The approach improves model discrimination, generalization to novel classes, and robustness under adversarial conditions through explicit class conditioning.
• Applications span vision-language tasks, few-shot learning, adversarial defense, and continual learning, yielding measurable accuracy and performance gains.

A class-aware prompting strategy is a paradigm in neural network adaptation, prompting, and inference that introduces explicit, class-conditioned information—whether in the form of learnable embeddings, prompt tokens, templates, principles, or functional code—into the model’s context or representation. This explicit class conditioning enables finer discrimination, improved generalization to novel classes, robustness under adversarial or distributional shift, and interpretability of model decisions. Substantial advances in vision-language modeling, few-shot and continual learning, adversarial defense, data-free quantization, text classification, segmentation, and fine-grained classification leverage class-aware prompting as a central methodological innovation.

1. Core Concepts and Formal Definitions

Class-aware prompting refers to strategies that encode or inject class-conditional information into a model’s prompt or input space. Formally, let $C$ denote the set of candidate classes. The strategy may involve the following components:

• Class-aware prompt templates: Compositional templates incorporating class labels or semantics, e.g., $["a photo of a", \text{class name}]$.
• Learnable prompt tokens: Class-specific embeddings $p_c$ (or tuples/blocks thereof) that are trained jointly with downstream objectives.
• Class-conditional queries: In transformer architectures, per-class query vectors $z^k$ sampled from a generative space (e.g., Gaussian mixture for each class).
• Text-to-text objectives: Losses that force learned prompts $t^r_c$ to be classified as specific class $c$ against hand-crafted templates $t^{h,l}_c$.
• Principle-based guides: Explicit, class-wise “principles” or rules extracted (often via LLMs) that capture discriminative descriptions for each class.
• Mixup-class embedding: Linear interpolation or fusion of the embeddings of multiple class labels, as in data-free quantization, to promote diversity and robustness.
• Class-aware phase/amplitude prompts: In frequency-space defenses, prompts per class are learned for both phase and amplitude spectra.

This explicit class awareness is contrasted with class-agnostic prompts, which do not encode or condition on label information and thus cannot easily support nuanced, class-specific adaptation or discrimination.

2. Model Architectures and Optimization Schemes

2.1 Language-Aware Soft Prompting (LASP)

In LASP (Bulat et al., 2022), the core is a text-to-text cross-entropy loss guiding class-aware soft prompt tokens $[p_1,\dots,p_M]$. For each class $c$, these tokens are appended to the class name embedding $w_c$ to form a prompt $r_c=[p_1,\dots,p_M,w_c]$, whose text encoder output $t^r_c$ is optimized to align with all hand-crafted templates of class $c$. The model combines this text-to-text loss $\mathcal{L}_{TT}$ with the usual vision–language contrastive loss, weighted to emphasize the class-awareness. Extensions include grouped prompt specialization and inclusion of “virtual classes” for zero-shot generalization.

2.2 Dynamic and Modular Class-Conditioned Approaches

• Class-Conditional Prompting Machine (CPM) (Chen et al., 2024) alternates transformer attention between class-agnostic and class-conditional query sets, with the latter directly sampled per class and trained with matching-free audio/visual contrastive objectives.
• Prompt-and-Transfer (PAT) (Bi et al., 2024) for few-shot segmentation uses text-initialized, per-class prompt tokens refined by self-attention and image semantics via a Part Mask Generator and Semantic Prompt Transfer; prompts for background and all significant parts are dynamically updated throughout the encoder.

2.3 Principle- and Path-based Prompt Decomposition

• Principle-Based Multi-Agent Prompting (Wei et al., 11 Feb 2025) for text classification uses LLMs to synthesize class-discriminative principles from demonstration samples, summarized into a reusable, class-aware prompt supplied to classifier agents.
• Class/Path-Aware Prompting in Testing (Ryan et al., 2024) decomposes code coverage generation into explicit path-constraint prompts: for every execution branch or outcome, a tailored prompt specifies the relevant constraints and expected behaviors, isolating the search space to class-specific (type-specific) test synthesis.

2.4 Deep Prompt Specialization

• INCPrompt (Wang et al., 2024) in continual learning maintains task- or class-specific “key-learner” heads and prompt-generating modules, extending the prompt set as new classes arrive. Each head learns to recognize its corresponding class/task via a triplet loss, and prompts are injected into the transformer attention as additional K/V tokens, conferring class-conditioned memory without interference or rehearsal.

2.5 Mixup-Class Prompting

In data-free quantization (Park et al., 29 Jul 2025), mixup-class prompts are constructed by linearly interpolating embeddings of two or more class labels, resulting in prompt $v_\text{mix}=\sum_i\lambda_i v_i$ that is used as the text condition for image synthesis. This prompts generative models to create calibration data that enhances robustness and generalization in quantization.

3. Losses, Calibration, and Training Paradigms

The efficacy of class-aware prompting depends on targeted loss functions and optimization schemes:

Approach	Main Loss Formulation	Calibration/Regularization
LASP (Bulat et al., 2022)	Text-to-text CE + V/L CE	LayerNorm+Bias realignment
CPM (Chen et al., 2024)	Mask/class CE, Focal/Dice, InfoNCE	Alternating query modes
PAT (Bi et al., 2024)	Pixelwise similarity + reg/dis losses	Prompt alternation + mask
Principle-based (Wei et al., 11 Feb 2025)	Macro-F1 (classification)	LLM-driven consolidation
Mixup-class (Park et al., 29 Jul 2025)	PTQ loss/calibration	Gradient-norm analysis
INCPrompt (Wang et al., 2024)	Task/class CE, triplet+L1 (heads)	Frozen backbone, disjoint

The class-aware losses typically enforce that each prompt—whether represented as tokens, principle guides, or query vectors—acts as a discriminative bridge between the model’s internal representations and explicit class semantics.

4. Applications and Empirical Findings

Class-aware prompting is a cross-cutting paradigm that has driven empirical advances in multiple fields:

• Zero-shot and novel class generalization: LASP achieves up to +5% accuracy gain on novel classes via text-to-text loss and grouped prompt specialization (Bulat et al., 2022).
• Few-shot and cross-domain segmentation: PAT excels in few-shot, cross-domain, and zero-shot segmentation across 11 benchmarks, leveraging dynamically updated class-specific prompt tokens (Bi et al., 2024).
• Audio-visual segmentation: CPM’s dual-mode query structure and contrastive objectives yield SOTA mIoU and F₁ across AVSBench-Objects/Semantics (Chen et al., 2024).
• Adversarial robustness: Phase/amplitude-aware, class-specific spatial frequency prompts raise robust accuracy from near 0% to >37% (AutoAttack) without inference overhead (Xu et al., 6 Feb 2025).
• Test generation and regression: SymPrompt increases correct test synthesis 5-fold and line coverage by 26% for CodeGen2, by decomposing the generation problem into class/path-specific sub-prompts (Ryan et al., 2024).
• Class-incremental and continual learning: INCPrompt reduces forgetting to 6.64% (Split CIFAR-100) and 6.47% (ImageNet-R), outperforming rehearsal-free baselines by explicitly encoding class or task identity in the prompt set (Wang et al., 2024).
• Fine-grained classification: In fine-grained scenarios (e.g., CUB-200-2011, Food101), multimodal class-aware prompting (MP-FGVC) yields 0.5–0.9% top-1 accuracy gains with subcategory-specific vision prompts and discrepancy-aware text tokens (Jiang et al., 2023).
• Data-free quantization: Mixup-class prompting pushes low-bit PTQ accuracy above prior methods; e.g., on ResNet50 W2A4, 70.78% vs. 70.35% for GenQ, with additional generalization guarantees (Park et al., 29 Jul 2025).
• Ordinal grading: Ranking-aware prompting (CLIP-DR) explicitly models ordinal class structure with multi-directional prompt blocks and achieves SOTA F1/AUC in diabetic retinopathy grading (Yu et al., 2024).

5. Open Variants: Multimodal, Ranking, and Mixup-Class Prompting

Recent innovations extend class-aware prompting along multiple technical dimensions:

• Multimodal and discrepancy-aware prompting: MP-FGVC (Jiang et al., 2023) employs both vision-side and text-side class-aware prompts to enforce cross-modal semantic alignment, critical in fine-grained visual recognition.
• Ranking-aware prompting: CLIP-DR (Yu et al., 2024) introduces paired prompt embeddings binding neighboring ordinal classes, combined with ranking losses and Gaussian smoothing in the similarity matrix for robust ordinal classification.
• Principle-based and multi-agent guides: LLM-powered principle extraction offers a scalable, interpreter-friendly, and cost-efficient approach for class definition in text domains, outperforming standard few-shot and chain-of-thought (Wei et al., 11 Feb 2025).
• Mixup-class: Data-free quantization demonstrates that fusing class semantics at the prompt level via mixup yields better synthetic calibration sets than either single-class or pixel-mixup methods (Park et al., 29 Jul 2025).

6. Limitations, Ablations, and Generalization

Ablation studies across multiple works consistently demonstrate the following:

• Label supervision is critical: Unlabeled principle extraction, or omission of class/constraint information, degrades downstream performance by 2–8% on macro-F1 (Wei et al., 11 Feb 2025).
• Increasing demonstration/test path granularity beyond a moderate point can yield diminishing or negative returns, especially in LLM principle extraction and path-wise prompt splitting (Ryan et al., 2024, Wei et al., 11 Feb 2025).
• Prompt grouping, alternation, or mixture (as in grouped LASP or Mixup-Class) consistently offers small but robust accuracy or generalization gains, with proper regularization for overfitting.
• Prompt lengths and initialization are robust, with optimal sizes typically being small (e.g., LASP $M=4$ tokens, CLIP-DR $m=8$), and random/zero initialization providing similar stability (Bulat et al., 2022, Yu et al., 2024).

Contemporary class-aware prompting pipelines are efficient: most strategies introduce minimal additional memory or computational overhead—indeed, multi-agent principle-based prompting and grouped/classwise tokens allow re-use across test instances (Wei et al., 11 Feb 2025, Bulat et al., 2022). A plausible implication is that scaling to large class sets or domains is feasible, contingent upon careful prompt management and modularization.

7. Outlook and Generalization Across Domains

Class-aware prompting has proven foundational for state-of-the-art adaptation in vision-language, segmentation, text classification, code coverage, robustness, and quantization. The strategy is readily extensible to ordinal (ranking-aware) tasks, modality fusion, and continual learning frameworks.

A plausible implication is that any foundation model with unified joint embedding or cross-modal capacity (e.g., CLIP, ALIGN, Florence) can benefit from explicit class-aware prompt design, whether by direct supervision, prompt fusion/grouping, or meta-level principle extraction. Moreover, prompt smoothing, regularization, and dynamic alternation are critical for robust generalization in low-data and cross-domain settings.

Class-aware prompting is evolving toward unified, modular strategies that encode semantic structure, task hierarchy, and even ranking relations into the prompting substrate, yielding robust, interpretable, and high-performing models across application domains.