Large vision-language models have been used to create cross-modal retrieval systems, including CLIP, that

have achieved large performance improvements but often act as black boxes, which makes it more difficult to

use these models and apply them in more critical areas. This non-disclosure is a great impediment to the

responsible implementation of such systems in high stakes applications. As a measure to counter this

shortcoming, we suggest an explainability-based design with an embedded post-hoc interpretation modules as

part of the CLIP retrieval pipeline. The framework provides sensible, dual- mode accounts of bidirectional

retrieval tasks; to begin with, it produces visual heatmaps that both outline the regions in the image that have

the strongest impact on a retrieval decision and to a second end, it deactivates word-level attribution to

explainability mechanisms should be augmented to multimodal retrieval systems in order to foster trustful,

responsible AI solutions. Based on the increased transparency, the approach prepares the foundation of more

solid and trustworthy human-AI cooperation.

Index Terms: Explainable AI (XAI), CLIP, Cross-Modal Re- trieval, Multimodal Learning, Interpretability,

Visual Attribution, Text Attribution, Human-AI Collaboration

The rapid evolution of artificial intelligence has ushered in a new era of sophisticated multimodal systems

capable of understanding and connecting information across different data types, such as images and text.

Among these, cross-modal retrieval—the task of finding relevant content in one modality (e.g., images) using

alignment comes at a cost: interpretability. CLIP, like many other deep learning models, functions as a

complex, non-linear “black box.” While it can retrieve a relevant image for a query like” a dog playing fetch,”

it provides no inherent explanation for why that particular image was selected over others, or which visual

features (the dog, the frisbee, the grass) or words in the caption were most decisive. This opacity is not merely

a technical curiosity; it is a significant barrier to user trust, system debugging, bias detection, and the ethical

deployment of AI in sensitive domains such as healthcare, security, and media.

The burgeoning field of Explainable AI (XAI) aims to bridge this gap between performance and transparency.

Tech- niques like Grad-CAM [?] for visual models and LIME [?] for tabular and text data have been

developed to shed light on model decisions. Yet, their direct application to multimodal, contrastive models

like CLIP in a retrieval context remains a nuanced and underexplored challenge. Explaining a classification

standard CLIP-based retrieval pipeline with two dedicated, plug-and-play interpretation modules. The

framework provides:

Visual Explanations: Generating similarity-aware heatmaps that localize and highlight the specific regions

within a retrieved image that were most influential for matching a given text query.

Textual Explanations: Employing an occlusion-based attribution method to score and highlight the

importance of individual words in a retrieved caption relative to a query image.

experiments confirm that the addition of these explanation modules does not degrade the intrinsic retrieval

performance of the base CLIP model, as measured by standard metrics like Recall@K. The framework thus

delivers meaningful, intuitive transparency “for free,” transforming a powerful but opaque retrieval tool into

Our study falls at the crossroads of three running areas of interest: (1) vision-language pre-training and cross-

modal re- trieval; (2) explainable artificial intelligence of visual models; and (3) explicability of multimodal

and retrieval systems. An overview of the relevant advances is given here and the specific gap that our

framework intends to fill outlined.

The quest to bridge the semantic gap between vision and language has driven significant innovation. Early

shared space, a method both elegant and powerful.

Following CLIP, subsequent efforts have focused on scaling (e.g., Florence [5]), incorporating more granular

objectives (e.g., BLIP [3] for generative and understanding tasks), or expanding modalities. These models

have set the state-of- the-art for zero-shot and fine-tuned cross-modal retrieval on benchmarks like Flickr30k

and MS-COCO. Our framework is agnostic to the specific base model but utilizes CLIP as a canonical, high-

performing example to build upon, focusing not on improving the embedding quality itself but on making its

retrieval decisions interpretable.

attribute importance.

While highly effective for classification and detection tasks, these methods are not directly transferable to

class. In contrast, explaining re- trieval requires elucidating a relative similarity score between two distinct

modalities—why is this image more similar to this text than others? This necessitates a fundamental

rethinking of the attribution target.

Explaining multimodal systems presents unique challenges, as explanations must often bridge modalities

themselves. Initial work has explored generating textual rationales for visual question answering (VQA),

as seen in VQA-X [?] and related efforts, where the goal is to produce a natural language justification for

Other works have adapted gradient-based techniques to the contrastive loss function to derive saliency maps

[?]. While promising, these approaches often remain tied to specific model architectures or training

modifications.

Our work distinguishes itself by proposing a lightweight, post-hoc framework that can be seamlessly

attached to a pre- trained, frozen CLIP model without any retraining or architectural change. We combine

a novel similarity-aware visual attribution method that operates directly on the embedding space with a

straightforward yet effective occlusion-based textual attribution.

This design philosophy prioritizes practi- cal deployability and user accessibility, aiming to make the

“black-box” retrieval process immediately more transparent to end-users and researchers alike.

thereby directly addressing the transparency gap in deployable multimodal AI systems.

Our framework, illustrated in Figure 1, consists of four inter- connected modules: (1) Multimodal Encoding,

(2) Embedding Indexing, (3) Similarity-based Retrieval, and (4) Explainability Generation.

The architecture maintains CLIP’s frozen pa- rameters while adding lightweight explanation modules that

operate in post-hoc manner without affecting the core retrieval mechanism.

Fig. 1: End-to-end architecture of the proposed explainability driven framework. Dashed lines indicate

optional paths for explanation generation.

we consider n-gram contexts:

Fig. 3: Visual explanation for the text query “a brown dog running through a grassy field.”

Textual Explanation Case Study: Figure 4 demonstrates the textual attribution for an image query depicting a

crowded urban street at night. The top-retrieved caption is “A busy city street illuminated by neon signs

at night.” Our occlusionbased attribution highlights “busy,” “city,” “street,” “neon signs,” and “night” as high-

importance words. Articles and prepositions receive negligible scores. This breakdown reveals that the

Fig. 4: Textual explanation for an image of a neon-lit city street.

Two aspects of our framework are validated by the ex- perimental results. In terms of numbers, it maintains

the robust CLIP baseline’s retrieval accuracy, guaranteeing that interpretability and performance are not

compromised. Qualitatively, the explanations produced are not only algorithmically sound but also

immediately comprehensible to a human observer and semantically meaningful.

attribution from classification tasks to the retrieval domain. By providing a causal, perturbation-based

Our framework’s dual success—maintaining baseline CLIP performance while offering human-

interpretable explana- tions—is confirmed by experimental validation. This section explores the

implications of these findings in greater detail, looks at the inherent drawbacks of our methodology, and

con- siders how important explainability is to building confidence in AI systems.

The qualitative effectiveness of our visual and textual attri- butions offers more than just proof-of-

functionality; it provides a lens to audit and understand CLIP’s internal reasoning. For instance, in Figure

3, the model’s focus on the ”dog” and ”grass” validates that it has semantically grounded the textual

concepts in specific visual features. Similarly, the textual attribution in Figure 4 reveals that CLIP prioritizes

concrete nouns and descriptive adjectives over grammatical filler words. This alignment between