LLEXICORP: CNN Explanations via CRP & LLM

• LLEXICORP is a modular system that couples CRP with a multimodal LLM to transform quantitative CNN outputs into natural-language explanations.
• It automates concept extraction, labeling, and contextualization, overcoming the manual bottlenecks in traditional concept-based explainability.
• The system grounds its narrative explanations in measured relevance scores and saliency maps, enhancing transparency and usability for end users.

Searching arXiv for the primary paper and closely related CRP/XAI context. LLEXICORP, short for “LLM EXplaIns COncept Relevance Propagation,” is a modular system for end-user explainability of convolutional neural networks that couples Concept Relevance Propagation (CRP) with a multimodal LLM to transform concept-level attributions, saliency maps, and representative images into natural-language explanations (Kůr et al., 4 Nov 2025). In the formulation introduced in “LLEXICORP: End-user Explainability of Convolutional Neural Networks” (Kůr et al., 4 Nov 2025), the system addresses a specific bottleneck in concept-based XAI: CRP can identify relevant channels, localize them in the input, and retrieve representative images, but semantic interpretation, concept naming, and explanation synthesis have largely remained manual.

1. Problem setting and motivation

LLEXICORP is positioned at the intersection of concept-based attribution and scalability (Kůr et al., 4 Nov 2025). The motivating observation is that, in the top layers of CNNs, individual channels can often be attributed to human-understandable concepts such as “lizard tails,” “circular holes,” or “wooden sticks.” CRP can identify which channels were most relevant for a prediction, show heatmaps on the input image indicating where each concept appeared, and provide representative images for each concept. What it does not provide by itself is an end-user-oriented narrative.

The workflow limitation is described as threefold: experts must visually inspect activation images and heatmaps to infer what each channel represents, manually name the concepts, and mentally combine concept relevance scores and heatmaps into a coherent explanation (Kůr et al., 4 Nov 2025). This process is time-consuming and non-scalable, expert-dependent, and poorly suited for auditors, customers, and non-technical stakeholders who need textual justifications rather than raw heatmaps.

LLEXICORP addresses this by coupling CRP with a multimodal LLM, specifically GPT-4o in the reported implementation, so that the system can automatically assign descriptive names to concept prototypes, contextualize each concept with respect to the original image and prediction, and summarize per-concept explanations into a coherent textual explanation (Kůr et al., 4 Nov 2025). The stated objective is end-user explainability with faithfulness: explanations should remain grounded in CRP’s quantitative outputs while being accessible and scalable.

2. Concept Relevance Propagation as the explanation substrate

The explanatory backend assumed by LLEXICORP is CRP, treated as a black-box explanation method but described formally enough to define the interface exposed to the LLM (Kůr et al., 4 Nov 2025). For a selected convolutional layer with filter set

$$C = \{ \text{all convolution filters in a selected layer} \},$$

a concept is identified with a specific filter or channel $c \in C$.

For an input image $I$ and predicted class $y$, CRP uses Layer-wise Relevance Propagation to propagate the prediction score backward through the network and determine which channels contributed most to that prediction. It identifies a subset

$\{c_1, \ldots, c_n\} \subseteq C$

with corresponding relevance values

$p_1, \ldots, p_n,$

where $p_i$ quantifies how relevant concept $c_i$ is for predicting $I$ as $y$ (Kůr et al., 4 Nov 2025). These values are typically normalized to sum to 1 for the chosen concept set, so they can be read as percentages.

For each selected concept $c \in C$0, CRP provides a heatmap $c \in C$1 over the input image and a set of representative images

$c \in C$2

from a validation set where the concept is highly relevant, together with corresponding saliency maps (Kůr et al., 4 Nov 2025). The paper gives the practical interpretation directly: channel relevance scores answer how important a concept is for a specific prediction, heatmaps answer where that concept is recognized in the input, and representative images answer what other images look like this concept.

LLEXICORP does not expose raw activations or network weights to the LLM. The model sees only processed CRP outputs: saliency overlays, relevance scores, prediction confidence, concept labels produced in the naming stage, and the original image (Kůr et al., 4 Nov 2025). This separation is central to the system’s notion of faithfulness.

3. Architecture and prompting strategy

The implementation reported in the paper uses a pretrained VGG16 batch-normalized variant in PyTorch, a standard CRP implementation from Anders et al. via Zennit/CoRelAy, and OpenAI GPT-4o as the multimodal LLM (Kůr et al., 4 Nov 2025). The pipeline has four stages: concept extraction, concept labeling, contextualization, and global summary.

In the concept extraction stage, CRP identifies the top-$c \in C$3 relevant concepts $c \in C$4, scalar relevances $c \in C$5, and heatmaps $c \in C$6, while also retrieving representative images $c \in C$7 and their heatmaps for each concept (Kůr et al., 4 Nov 2025). In the concept labeling stage, a “naming agent” prompts the multimodal LLM with the representative images and their saliency maps for a concept and asks it to produce a descriptive label $c \in C$8 capturing the common visual pattern. In the contextualization stage, an “explanation agent” receives the original image, predicted class $c \in C$9, confidence $I$0, concept heatmap $I$1, relevance $I$2, and concept label $I$3, and produces a contextualized explanation $I$4. In the final summarization stage, the LLM receives only the textual outputs $I$5 and produces a concise final explanation (Kůr et al., 4 Nov 2025).

The prompting strategy explicitly separates naming from explanation (Kůr et al., 4 Nov 2025). For concept labeling, the LLM is told that the common pattern must be recognizable by a convolutional filter. This constraint is intended to prevent the model from jumping to overly high-level semantic interpretations. The example outcomes reported in the paper include labels such as “circular holes or openings with a distinct rim” and “wooden cylindrical objects arranged in parallel or patterns.”

For contextualization, the prompt is organized around two decision tasks. The first is the concept–pattern relation, with four categories: direct recognition, feature recognition, co-occurrence recognition, and misidentification. The second is the concept–prediction relation, with four categories: exact classification, compositional association, contextual association, and misassociation (Kůr et al., 4 Nov 2025). This two-stage decision process is described as an “algorithmic” phrasing that reduces free-form storytelling and encourages a faithful mapping from CRP outputs to text.

The summarization step is additionally constrained by a no-new-information instruction: the LLM must create a concise, non-technical summary using only the previously generated textual descriptions (Kůr et al., 4 Nov 2025). This design choice is tied directly to the human evaluation.

4. Representation of relevance and explanation styles

The system encodes CRP outputs for the LLM in two forms: visual saliency maps and textual or numeric relevance information (Kůr et al., 4 Nov 2025). Prediction confidence $I$6 and concept relevance shares $I$7 are provided as plain numeric values in the prompt so that the LLM can express relative importance in the final explanation.

The paper gives the practical relevance-share formula as

$I$8

where $I$9 is the total relevance assigned to channel $y$0 for the prediction (Kůr et al., 4 Nov 2025). In the lizard example used throughout the paper, the final summary refers directly to the top five concept percentages: 40.62% for “lizards,” 21.35% for “elongated shapes with distinct edges,” 21.29% for “elongated green leaves or stems,” 8.60% for “rough, scaly texture,” and 8.14% for “circular holes or openings.”

LLEXICORP supports at least two levels of abstraction in its generated text (Kůr et al., 4 Nov 2025). At the low-level or technical end, prompts can request explicit mention of concepts, relevance percentages, prediction confidence, and the distinctions among direct, feature, and co-occurrence recognition, as well as compositional versus contextual association. At the high-level or non-technical end, the summarization step asks for concise narrative language without introducing new information. The paper’s final summary for the “American chameleon” image illustrates this second mode, integrating all five concepts and their relevance percentages into a single explanation.

This suggests that the system is not a new attribution method in itself but a translation layer over concept-based attribution. A plausible implication is that its main technical novelty lies less in the discovery of concepts than in the controlled linguistic rendering of CRP semantics.

5. Evaluation, case studies, and empirical findings

The reported evaluation is qualitative and is conducted on ImageNet validation images using VGG16 with CRP (Kůr et al., 4 Nov 2025). The authors fix a random seed, sample 8 images, and for each image generate per-concept explanations for the top-5 concepts and a final summary.

The American chameleon example is the paper’s principal case study (Kůr et al., 4 Nov 2025). VGG16 classifies the image as “American chameleon” with 86.11% confidence. CRP identifies five concepts, and the LLM labels them as “lizards,” “elongated shapes with distinct edges,” “elongated green leaves or stems,” “rough, scaly texture,” and “circular holes or openings with a distinct rim.” The contextualization stage distinguishes structural features such as body parts and texture from contextual cues such as surrounding foliage.

A second case concerns an image predicted as cheeseburger in which CRP identifies a concept corresponding to “wooden cylindrical objects arranged in parallel” and localizes it on French fries (Kůr et al., 4 Nov 2025). The explanation classifies this as feature recognition and contextual association: the fries visually resemble the concept prototypes and commonly co-occur with cheeseburgers.

The user questionnaire involves 21 participants and evaluates both per-concept descriptions and final summaries over the 8 sampled images (Kůr et al., 4 Nov 2025). The aggregate results are as follows.

Evaluation item	Agree	Not sure / Disagree
Visual pattern	78%	10% / 11%
Highlighted areas	61%	13% / 26%
Reasonable presence	67%	15% / 18%
Useful explanation	65%	15% / 21%
Only existing info	83%	10% / 7%
Helpful summary	75%	15% / 10%
More helpful than individual concepts	33%	35% / 33%

The top-ranked concept performs substantially better than lower-ranked concepts, with 94% agreement on visual pattern, 85% on highlighted areas, 92% on reasonable presence, and 92% on useful explanation (Kůr et al., 4 Nov 2025). The conditional tables further indicate that when respondents accept the visual pattern and localization, they overwhelmingly judge the explanation as reasonable and useful.

The paper interprets these findings as evidence that concept descriptions and summaries are often perceived as faithful and useful, especially when CRP produces convincing prototypes and localization (Kůr et al., 4 Nov 2025). At the same time, the lower agreement for highlighted areas shows that explanation quality remains dependent on the underlying attribution quality.

6. Faithfulness, limitations, and broader interpretations

The paper argues for faithfulness through four mechanisms: grounding in CRP outputs, structured prompts, separation of tasks, and human-study evidence (Kůr et al., 4 Nov 2025). All explanations are derived from concept rankings $y$1, spatial heatmaps $y$2, and prototypes; naming is separated from contextualization; and summaries are generated only from previously produced text. The 83% agreement score on the no-new-information criterion is used as empirical support for this design.

Several limitations are explicit (Kůr et al., 4 Nov 2025). Localization reliability is imperfect, with only 61% agreement overall on correctly identified highlighted areas. Lower-relevance concepts receive weaker agreement scores and may clutter explanations without adding clarity. Final summaries are often helpful as consolidation, but only 33% of participants found them more helpful than the individual concept descriptions. The experiments are confined to ImageNet, VGG16, and CRP, so generalization to other architectures, tasks, or domains is not evaluated. The paper also notes that LLMs can still over-interpret patterns or choose overly specific names even under constrained prompting.

Within the XAI landscape, LLEXICORP is presented as a combination of concept-based interpretability, relevance propagation, and LLM-based narrative generation (Kůr et al., 4 Nov 2025). Relative to original CRP workflows, it automates concept labeling, contextualization, and summarization. Relative to saliency methods such as Grad-CAM, SHAP, LIME, D-RISE, and traditional LRP, it adds concept-level decomposition plus natural-language rendering. Relative to concept methods such as TCAV and ACE, it adds automatic naming and explanation. Relative to prior LLM-XAI systems, it is specifically tailored to CNNs, multimodal inputs, and CRP semantics.

In adjacent literature, the name “LLEXICORP” also appears as an inferred label for broader lexical or corpus-oriented resource design rather than for CNN explainability. This broader reading is not the primary referent of the 2025 paper, but the supplied literature suggests several neighboring interpretations. XLEnt is a large automatically mined cross-lingual entity dataset with 164 million distinct cross-lingual entity pairs across 120 languages aligned with English, built with the LSP-Align method (El-Kishky et al., 2021). LEXpander is a one-hop lexicon expansion method over colexification networks, evaluated on English and German LIWC-style categories and sentiment variables (Natale et al., 2022). “Language Lexicons for Hindi-English Multilingual Text Processing” defines Hindi-English language lexicons with two probabilistic language-strength scores for each word over a combined vocabulary of 36,429 entries (Ansari et al., 2021). “Adaptation of a Lexical Organization for Social Engineering Detection and Response Generation” presents an extensible, LCS-based lexical organization aligned to PERFORM, GIVE, LOSE, and GAIN for ask and framing detection (Bhatia et al., 2020). “LeXFiles and LegalLAMA” releases a multinational English legal corpus of approximately 5.8 million documents and 18.8 billion tokens together with a legal probing benchmark (Chalkidis et al., 2023). These works do not redefine LLEXICORP itself, but they suggest that the term can plausibly function as an umbrella shorthand for large lexical, cross-lingual, or legal corpora in neighboring discussions.

As a named system, however, LLEXICORP denotes the CRP-plus-LLM pipeline introduced for end-user explainability of CNNs (Kůr et al., 4 Nov 2025). Its central contribution is to convert concept rankings, prototypes, and saliency maps into controlled natural-language explanations while preserving an explicit link to the quantitative structure of CRP.