CCPL: Concept-Constrained Prompt Learning

• CCPL is a prompt-learning paradigm that constrains soft prompts using explicit concept-level semantics to enhance generalization and interpretability.
• It regularizes optimization through frozen prototypes, multi-granularity alignment, structured decomposition, and retrieval-mediated guidance across modalities.
• Empirical results demonstrate improved transfer, robustness, and explainability in vision-language tasks, medical diagnosis, text classification, and code search.

Concept-Constrained Prompt Learning (CCPL) denotes a family of prompt-learning methods in which learnable prompts are not optimized as unconstrained latent vectors, but are anchored, regularized, decomposed, or retrieved through explicit concept-level semantics. In the most explicit formulation, CCPL is a few-shot CLIP adaptation framework that aligns learnable class prompts with frozen concept prototypes in text space (Sang et al., 21 Jun 2026). Closely related formulations appear as concept-guided prompt learning for vision-language generalization (Zhang et al., 2024), clinically grounded explainable prompt learning for diagnosis (Bie et al., 2024), concept decomposition for interpretable continuous prompts in text classification (Chen et al., 2024), and contrastive prompt learning for code search with fine-grained NL–PL interaction (Zhang et al., 2023). Across these settings, the shared objective is to preserve or expose semantically meaningful structure that ordinary soft-prompt optimization tends to obscure.

1. Definition and conceptual scope

The exact phrase “Concept-Constrained Prompt Learning” is used explicitly in few-shot CLIP adaptation, where prompt tuning is regularized by frozen class-level concept prototypes and applied to the base-to-new transfer problem (Sang et al., 21 Jun 2026). In that setting, the motivation is precise: class-only prompt optimization can overfit base-class supervision, move text embeddings away from CLIP’s pretrained language semantics, and degrade transfer to unseen classes. Concept constraints are introduced as semantic anchors that keep learned prompts near conceptually meaningful directions in text space.

Related papers broaden the operational meaning of CCPL without always using the term. “Concept-Guided Prompt Learning for Generalization in Vision-LLMs” constrains prompts through retrieved transferable visual concepts such as colors, shapes, sizes, and materials (Zhang et al., 2024). “XCoOp: Explainable Prompt Learning for Computer-Aided Diagnosis via Concept-guided Context Optimization” frames prompt learning around clinical concepts so that soft prompts correspond to disease-relevant findings rather than arbitrary embeddings (Bie et al., 2024). “Concept Based Continuous Prompts for Interpretable Text Classification” replaces a prompt matrix with a concept embedding matrix and coefficient matrix, effectively constraining prompt behavior to a compact concept basis (Chen et al., 2024). In code search, the CCPL connection is interpretive rather than terminological: the prompts function as trainable semantic constraints for aligning NL and PL representations under multimodal contrastive learning (Zhang et al., 2023).

This suggests that CCPL is best understood as a general design principle rather than a single architecture. The unifying criterion is that concept structure is made operational in prompt optimization: concepts may appear as frozen prototypes, hand-crafted descriptors, retrieved cache items, decomposed bases, or modality-specific semantic constraints.

2. Core formulations and optimization patterns

A central distinction between CCPL and ordinary soft-prompt tuning is whether prompts are treated as free optimization variables. In CoOp-style learning, shared context tokens are optimized mainly for downstream accuracy. In CCPL-style methods, those same tokens are additionally tied to concept representations, either directly in text space, indirectly through cross-modal retrieval, or structurally through a decomposition constraint (Sang et al., 21 Jun 2026).

The canonical explicit objective is the text-space cosine consistency loss used in few-shot CLIP adaptation. For base classes, the learned class-prompt embedding $\mathbf{t}^{\mathrm{cls}_c(\theta)}$ is aligned with a frozen concept prototype $\mathbf{t}^{\mathrm{con}_c}$ through

$$\mathcal{L}_{\mathrm{text}}= \frac{1}{|\mathcal{C}_{\mathrm{base}}|} \sum_{c \in \mathcal{C}_{\mathrm{base}}} \left(1-\cos\!\left(\mathbf{t}^{\mathrm{cls}_c(\theta)},\mathbf{t}^{\mathrm{con}_c}\right)\right),$$

and the default training objective is

$$\mathcal{L}_{\mathrm{CCPL}}= \mathcal{L}_{\mathrm{ce}}+\lambda_{\mathrm{text}}\mathcal{L}_{\mathrm{text}},$$

with default $\lambda_{\mathrm{text}}=0.5$, concept dropout $p_{\mathrm{drop}}=0.3$, weak concept-guided fusion $\alpha=0.1$, and no KL-based prediction consistency term (Sang et al., 21 Jun 2026).

A second pattern is multi-granularity prompt alignment. XCoOp aligns soft prompts with clinical prompts both at the token level and at the full-prompt level. The token-level loss $\mathcal{L}_T$ and prompt-level loss $\mathcal{L}_P$ are combined as the soft-hard prompt alignment module $\mathcal{L}_{\mathrm{PPA}}$, and then integrated with a global-local image-prompt alignment loss $\mathbf{t}^{\mathrm{con}_c}$0 so that prompt semantics are grounded in both full-image and local patch evidence (Bie et al., 2024). In this formulation, concept constraints are not merely regularizers on prompt magnitude or smoothness; they explicitly pull the learned prompt space toward medically meaningful descriptors.

A third pattern is concept-space reconstruction. In concept decomposition for text classification, a prompt matrix $\mathbf{t}^{\mathrm{con}_c}$1 is approximated by $\mathbf{t}^{\mathrm{con}_c}$2, where $\mathbf{t}^{\mathrm{con}_c}$3 is a concept embedding matrix and $\mathbf{t}^{\mathrm{con}_c}$4 is a coefficient matrix:

$\mathbf{t}^{\mathrm{con}_c}$5

The paper states that for any $\mathbf{t}^{\mathrm{con}_c}$6 and any $\mathbf{t}^{\mathrm{con}_c}$7, there exist $\mathbf{t}^{\mathrm{con}_c}$8 and $\mathbf{t}^{\mathrm{con}_c}$9 such that $$\mathcal{L}_{\mathrm{text}}= \frac{1}{|\mathcal{C}_{\mathrm{base}}|} \sum_{c \in \mathcal{C}_{\mathrm{base}}} \left(1-\cos\!\left(\mathbf{t}^{\mathrm{cls}_c(\theta)},\mathbf{t}^{\mathrm{con}_c}\right)\right),$$0, and then optimizes a fidelity term based on KL divergence together with a task loss (Chen et al., 2024). Here the constraint is structural: prompt expressivity is preserved while interpretation is forced through a small human-readable concept set.

A fourth pattern is retrieval-mediated concept guidance. CPL builds a text concept dictionary of $$\mathcal{L}_{\mathrm{text}}= \frac{1}{|\mathcal{C}_{\mathrm{base}}|} \sum_{c \in \mathcal{C}_{\mathrm{base}}} \left(1-\cos\!\left(\mathbf{t}^{\mathrm{cls}_c(\theta)},\mathbf{t}^{\mathrm{con}_c}\right)\right),$$1 visual concept terms, constructs a visual concept cache from CLIP-aligned image features, retrieves Top-$$\mathcal{L}_{\mathrm{text}}= \frac{1}{|\mathcal{C}_{\mathrm{base}}|} \sum_{c \in \mathcal{C}_{\mathrm{base}}} \left(1-\cos\!\left(\mathbf{t}^{\mathrm{cls}_c(\theta)},\mathbf{t}^{\mathrm{con}_c}\right)\right),$$2 concepts with $$\mathcal{L}_{\mathrm{text}}= \frac{1}{|\mathcal{C}_{\mathrm{base}}|} \sum_{c \in \mathcal{C}_{\mathrm{base}}} \left(1-\cos\!\left(\mathbf{t}^{\mathrm{cls}_c(\theta)},\mathbf{t}^{\mathrm{con}_c}\right)\right),$$3, and refines text features through a Transformer-decoder projector that maps multi-level visual features into text space (Zhang et al., 2024). In this design, concept constraints are dynamic and image-conditional rather than fixed per class.

3. CLIP adaptation and vision-language generalization

In CLIP adaptation, CCPL addresses a specific failure mode of prompt tuning: improvement on base classes at the expense of unseen-class transfer. The explicit CCPL framework preserves CLIP’s frozen image and text encoders, learns only 16 shared context tokens, and uses class-specific concept banks to form frozen concept prototypes for regularization (Sang et al., 21 Jun 2026). The paper studies a 4-shot base-to-new protocol on DTD, EuroSAT, and OxfordPets, using automatically generated fallback splits rather than the official Zhou et al. splits.

The reported base/new/harmonic mean results show the intended transfer effect. On DTD, CoOp scores 72.3 / 55.2 / 62.6, while CCPL scores 73.8 / 55.3 / 63.2, a harmonic mean gain of +0.6. On EuroSAT, CoOp scores 79.8 / 56.3 / 66.0 and CCPL scores 79.7 / 60.6 / 68.9, a gain of +2.9 driven by improved new-class accuracy. On OxfordPets, CoOp scores 94.9 / 97.7 / 96.3 and CCPL scores 94.9 / 97.5 / 96.2, a near-neutral change of $$\mathcal{L}_{\mathrm{text}}= \frac{1}{|\mathcal{C}_{\mathrm{base}}|} \sum_{c \in \mathcal{C}_{\mathrm{base}}} \left(1-\cos\!\left(\mathbf{t}^{\mathrm{cls}_c(\theta)},\mathbf{t}^{\mathrm{con}_c}\right)\right),$$4 (Sang et al., 21 Jun 2026). The ablations indicate that removing text regularization reduces harmonic mean from 68.9 to 67.6 on EuroSAT, and that stronger inference fusion can improve new-class transfer further but with an explicit base–new trade-off controlled by $$\mathcal{L}_{\mathrm{text}}= \frac{1}{|\mathcal{C}_{\mathrm{base}}|} \sum_{c \in \mathcal{C}_{\mathrm{base}}} \left(1-\cos\!\left(\mathbf{t}^{\mathrm{cls}_c(\theta)},\mathbf{t}^{\mathrm{con}_c}\right)\right),$$5.

Concept-guided prompt learning for broader vision-language generalization implements a related but more retrieval-heavy strategy. Rather than building frozen class prototypes, CPL mines transferable visual concepts from CLIP’s own latent alignment by pairing concept words with matched training-image features in a cache (Zhang et al., 2024). Retrieved concept words are then used to form concept-guided prompts, while a projector refines text features with multi-level visual context from all visual encoder layers. The final classifier remains CLIP-style cosine similarity, but the text branch is enriched by both concept retrieval and visual-to-text refinement.

Empirically, CPL reports strong gains in base-to-novel generalization across 11 datasets. Its average scores are Base 84.38, Novel 78.03, and HM 81.08, with improvements of +1.69 over CoOp on base, +2.89 over the prior best baseline on novel, and +2.53 on harmonic mean. The paper highlights particularly strong harmonic mean gains on FGVCAircraft, DTD, EuroSAT, and UCF101, and reports the best target average of 68.07 in cross-dataset transfer, exceeding MaPLe’s 66.30 by 1.77% (Zhang et al., 2024). In ablation, concept-guided prompting contributes the largest component gain, with the sequence CLIP baseline 60.33, +CGP 64.38, +CGP+P 66.35, and +CGP+P+TA 66.92.

Taken together, these works identify two related but distinct CCPL regimes in VLMs. One regime uses class-level concept anchors to regularize prompt semantics during few-shot adaptation; the other retrieves transferable concepts and conditions prompt construction on image-specific evidence. Both are motivated by the claim that class names alone are semantically sparse and that unconstrained prompt optimization can distort CLIP’s pretrained cross-modal geometry.

4. Explainable prompt learning in medical diagnosis

In medical image classification, concept-constrained prompting is motivated less by transfer alone than by explainability requirements in high-stakes decision settings. XCoOp argues that standard soft-prompt learning methods such as CoOp and CoCoOp learn unexplainable latent vectors, which is inadequate when clinicians require explanations that are understandable, plausible, and faithful to the decision process (Bie et al., 2024). Its proposed solution is to align images, learnable prompts, and clinical concept-driven prompts at multiple granularities.

For each disease class, XCoOp constructs a hand-crafted clinical prompt from concept annotations when they exist, and otherwise elicits disease-specific visual concepts from GPT-4 using prompts such as “What are the most useful visual concepts to distinguish [disease name] in a dermoscopic image, chest X-ray, etc.?” The resulting prompts can take forms such as “a photo of {melanoma, with irregular pigment network, dots and globules, blue-whitish veil, and vascular structures}” (Bie et al., 2024). Learnable soft prompts are then initialized from templates such as “a photo of a [disease name]” and optimized subject to concept alignment.

The alignment has two layers. Token-level alignment matches the token embeddings of soft prompts and clinical prompts through cosine-similarity-based cross-entropy, while prompt-level alignment matches the global text features of full prompt sequences through the CLIP text encoder. These terms form the soft-hard prompt alignment module. XCoOp then adds global-local image-prompt alignment, combining global image features with averaged local patch similarities to the prompt text feature. This architecture directly encodes the claim that diagnosis depends on local lesion patterns as well as overall image evidence (Bie et al., 2024).

The explainability outputs are both textual and visual. Textual interpretation is obtained by measuring the distance between learned soft prompts and clinical prompts, with lower distance indicating better interpretability. Visual explanations are obtained by visualizing image–prompt similarity, and t-SNE plots of token embeddings show disease-wise clustering that the paper interprets as coherent clinical semantics (Bie et al., 2024).

The experimental results support both predictive and interpretive claims. XCoOp is reported to achieve the best performance on Derm7pt, SkinCon, Pneumonia, and IU X-Ray, improving over the second-best method by about 1.2% to 3.4% AUC and 1.2% to 2.0% accuracy. On prompt interpretation distance, KgCoOp records 1.982 average, LASP 3.011, and XCoOp 1.104. On SkinCon, the ablation sequence is LASP 78.31 AUC, + clinical concept-driven prompts 79.93, + image-prompt alignment 80.46, and + soft-hard prompt alignment 81.12. The faithfulness analysis uses knowledge intervention with random knowledge, general knowledge, clinical-concept-based knowledge, and intervened knowledge, and performance drops when clinical concepts are replaced or perturbed. The model is also reported to converge in about 35 epochs across different backbones (Bie et al., 2024).

Within the broader CCPL landscape, XCoOp is the clearest example of concept constraints serving interpretability as a first-class objective rather than a by-product of regularization.

5. Text classification and code search extensions

Concept constraints also appear in NLP settings where the primary aim is to interpret continuous prompts or to improve cross-modal semantic alignment. In interpretable text classification, concept decomposition replaces opaque prompt embeddings with a small set of human-readable concepts plus learned coefficients (Chen et al., 2024). Candidate concepts are generated by GPT-4o through class-specific prompts such as “describe the concepts of a positive review” or analogous requests for AGNews categories, with 50 samples used for diversity. The resulting candidate pools are then pruned by a monotone submodular objective

$$\mathcal{L}_{\mathrm{text}}= \frac{1}{|\mathcal{C}_{\mathrm{base}}|} \sum_{c \in \mathcal{C}_{\mathrm{base}}} \left(1-\cos\!\left(\mathbf{t}^{\mathrm{cls}_c(\theta)},\mathbf{t}^{\mathrm{con}_c}\right)\right),$$6

where $$\mathcal{L}_{\mathrm{text}}= \frac{1}{|\mathcal{C}_{\mathrm{base}}|} \sum_{c \in \mathcal{C}_{\mathrm{base}}} \left(1-\cos\!\left(\mathbf{t}^{\mathrm{cls}_c(\theta)},\mathbf{t}^{\mathrm{con}_c}\right)\right),$$7 promotes discriminative diversity and $$\mathcal{L}_{\mathrm{text}}= \frac{1}{|\mathcal{C}_{\mathrm{base}}|} \sum_{c \in \mathcal{C}_{\mathrm{base}}} \left(1-\cos\!\left(\mathbf{t}^{\mathrm{cls}_c(\theta)},\mathbf{t}^{\mathrm{con}_c}\right)\right),$$8 promotes coverage. Because the objective is monotone submodular, the paper uses the standard greedy algorithm.

After selection, the concept-constrained prompt $$\mathcal{L}_{\mathrm{text}}= \frac{1}{|\mathcal{C}_{\mathrm{base}}|} \sum_{c \in \mathcal{C}_{\mathrm{base}}} \left(1-\cos\!\left(\mathbf{t}^{\mathrm{cls}_c(\theta)},\mathbf{t}^{\mathrm{con}_c}\right)\right),$$9 is trained with

$$\mathcal{L}_{\mathrm{CCPL}}= \mathcal{L}_{\mathrm{ce}}+\lambda_{\mathrm{text}}\mathcal{L}_{\mathrm{text}},$$0

where $$\mathcal{L}_{\mathrm{CCPL}}= \mathcal{L}_{\mathrm{ce}}+\lambda_{\mathrm{text}}\mathcal{L}_{\mathrm{text}},$$1 is a KL-divergence fidelity term preserving the behavior of the original prompt and $$\mathcal{L}_{\mathrm{CCPL}}= \mathcal{L}_{\mathrm{ce}}+\lambda_{\mathrm{text}}\mathcal{L}_{\mathrm{text}},$$2 is the task loss (Chen et al., 2024). The reported outcome is that concept decomposition achieves similar accuracy to P-tuning across SST-2, IMDB, and AGNews, often matches or approaches the Vocab-1500 discrete explanation baseline using only a few concepts, and yields the best concept correlation in 7 out of 9 settings. The appendix further indicates that performance is often strongest with only 1–10 concepts, reinforcing the idea that a compact concept basis can be sufficient for faithful prompt reconstruction.

In code search, the paper “Contrastive Prompt Learning-based Code Search based on Interaction Matrix” does not explicitly frame itself as CCPL, but the mechanism is closely aligned with the same principle (Zhang et al., 2023). The method uses a dual-encoder architecture with separate prompt-template generators and MLP reparameterization branches for PL and NL, removes weight sharing between these prompt branches because of modality heterogeneity, freezes the original encoders, and trains only prompts and reparameterization layers. Prompt learning therefore acts as a parameter-efficient task-conditioning mechanism.

The semantic constraint is reinforced by symmetric contrastive learning and a fine-grained interaction matrix

$$\mathcal{L}_{\mathrm{CCPL}}= \mathcal{L}_{\mathrm{ce}}+\lambda_{\mathrm{text}}\mathcal{L}_{\mathrm{text}},$$3

which is aggregated by row and column max pooling into a final similarity score $$\mathcal{L}_{\mathrm{CCPL}}= \mathcal{L}_{\mathrm{ce}}+\lambda_{\mathrm{text}}\mathcal{L}_{\mathrm{text}},$$4 (Zhang et al., 2023). The model achieves overall MRR 0.789 on CodeSearchNet across Ruby, JavaScript, Go, Python, Java, and PHP, compared with 0.778 for CoCoSoDa, and also reports R@1 = 0.701, R@5 = 0.902, and R@10 = 0.940. Efficiency claims include around 80% of GPU memory of fine-tuning methods for a fixed batch size, only 19% of trainable parameters compared to fine-tuning, and about 22% faster training. This suggests that concept-like prompt constraints can be valuable even when the “concepts” are not human-authored attributes but latent semantic relations between natural-language intent and program functionality.

6. Empirical regularities, misconceptions, and boundary conditions

A recurring empirical pattern across CCPL variants is that concept constraints are most beneficial when the concept bank aligns naturally with the semantics of the target dataset. The explicit CCPL paper is unusually direct on this point: EuroSAT benefits substantially because scene-level descriptions such as vegetation, water, road structure, and land cover correspond closely to visual discriminants, whereas OxfordPets is a boundary condition because template concepts such as “Siamese cat texture” do not capture subtle breed-level cues (Sang et al., 21 Jun 2026). The paper therefore treats fine-grained categories as a current limit rather than assuming concept guidance is uniformly helpful.

A second regularity is that concept quality matters as much as optimization design. XCoOp assumes that meaningful clinical concepts can be written as prompt text, either by experts or by an LLM; it also implies that poor or incomplete LLM-generated concepts can degrade explanation quality and that expert validation may be required before deployment in high-stakes settings (Bie et al., 2024). Concept decomposition similarly acknowledges noise in selected concepts, heuristic filtering, and cases where explanations are not obviously aligned with the input text, which the authors attribute in part to shortcut links in prompt behavior (Chen et al., 2024).

A common misconception is that CCPL is merely prompt regularization in a generic sense. The surveyed methods show a stronger claim. In the explicit few-shot CLIP formulation, the regularizer aligns prompt embeddings to frozen class-level concept prototypes; in XCoOp, it aligns prompts to clinical prompts at token and prompt levels and further grounds them in local image evidence; in concept decomposition, it constrains the prompt to live in the span of selected concept embeddings; and in CPL, it retrieves concept tokens from a CLIP-native cache and refines text features with multi-level visual context (Sang et al., 21 Jun 2026). These are semantic constraints, not only optimization stabilizers.

Another misconception is that concept constraints automatically guarantee faithful interpretability. The evidence is more limited. XCoOp reports lower prompt interpretation distance, meaningful image-prompt visualizations, and sensitivity to knowledge intervention, which supports a stronger faithfulness claim than most prompt-learning papers (Bie et al., 2024). By contrast, text-side concept decomposition reports plausible and coherent concept summaries but also explicit failure cases. This suggests that CCPL can improve the inspectability of prompts and sometimes their faithfulness, but the two should not be conflated without task-specific validation.

The present literature therefore supports a measured conclusion. CCPL is a technically coherent prompt-learning paradigm with demonstrated benefits in transfer, interpretability, and parameter efficiency, but its success depends on the adequacy of the concept source, the semantic granularity of the task, and the fidelity with which concept constraints reflect the model’s actual decision variables. The strongest current evidence lies in CLIP adaptation under base-to-new evaluation, clinically grounded explainable diagnosis, and concept-level reinterpretation of continuous prompts, while fine-grained categories, noisy generated concepts, and limited evaluation scope remain persistent constraints (Sang et al., 21 Jun 2026).