In this work, we focus on integrating LLMs into a neuro-symbolic framework to enhance the quality of explanations associated with IF-THEN rules generated by neuro-fuzzy inference systems. To address the challenge posed by the lack of a reference ground truth in explanation tasks, we propose a quantitative evaluation based on linguistic and semantic quality metrics, aiming to assess the clarity, coherence, and relevance of the generated text. We systematically compare a selection of LLMs varying in size and architectural family, and investigate the impact of diferent prompting strategies-including zero-shot, persona-based, and fact-checking approaches-on the resulting explanations. The proposed framework is applied to a real-world case study on EEG-based seizure detection, illustrating its potential in high-stakes medical contexts where transparency and reliability are critical. The findings show that quantitative metrics alone are insuficient to capture the true quality of explanations, highlighting the critical role of both model selection and prompt design in generating efective, trustworthy, and human-aligned explanations.
In recent years, Large Language Models (LLMs) have emerged as powerful AI tools trained on massive
text corpora, enabling them to encode and generalize linguistic knowledge. Their broad applicability
stems from this rich pretraining, allowing efective use across diverse domains, such as healthcare
[
However, in the medical domain, the adoption of AI necessitates transparent and interpretable
explanations to foster trust, ensure accountability, and enable safe integration into clinical practice.
Clinicians must be able to understand and justify AI-driven decisions, particularly in high-stakes
scenarios [
Although LLMs are inherently black-box models, they are increasingly being adopted in the state of the art for generating explanations. Current research explores the use of in-context learning to inject both domain-specific and explainability-related knowledge into LLMs, enabling them to generate responses that combine user-friendly narratives for non-experts with technical insights for specialists Italy (G. Zaza)
CEUR
ceur-ws.org
[
In contrast to black-box machine learning models, fuzzy logic has long proven valuable for
explainability, as it is inherently transparent and enables the representation of human perceptions through
linguistic terms that map numerical values onto the imprecise concepts characteristic of human
reasoning [
In a recent study [
In this work, we extend our previous study by comparing LLMs from diferent families and sizes. We quantitatively evaluate the generated explanations using metrics that assess textual quality without relying on a ground truth, which is typically unavailable in this type of task. Furthermore, we investigate the efectiveness of diferent prompt patterns, including zero-shot prompting, persona-based prompting, and fact-checking–oriented prompting. Particularly, this work aims to address the following research questions, focusing on the efectiveness and quality of LLM-generated explanations in the context of neuro-symbolic systems applied to the same predictive task: RQ1 Which LLMs generate the most efective explanations, based on linguistic quality criteria such as lfuency, coherence, and well-formedness? RQ2 Is there a prompt pattern, among zero-shot, persona, persona + fact-checking, that yields higherquality explanations in terms of linguistic clarity, coherence, and contextual relevance? The rest of the paper is organized as follows: Section 2 presents the experimental setup, Section 3 discusses the results, and Section 4 outlines the conclusions and directions for future work.
To address the research questions outlined above, we conducted a case study on EEG-based epileptic
seizure classification. Building on the pipeline introduced in [
2. Explanatory phase: This phase focuses on transforming symbolic information into interpretable textual descriptions. The extracted fuzzy rules are reformulated by replacing generic labels with semantically meaningful names and then used to construct targeted prompts for LLMs, generating human-centered and understandable explanations. 3. Evaluation phase: In this phase, the explanations generated by the LLMs are analyzed and compared using linguistic quality metrics.
Unlike our previous work, this study does not focus extensively on the user-centered aspects of the
pipeline, even if these remain relevant and will be addressed in future developments involving domain
expert engagement. Instead, we concentrate on the evaluation module and the construction of the fuzzy
rule base, which are central to the current investigation.
2.1. Data
Epilepsy is one of the most prevalent neurological disorders worldwide, afecting over 50 million people,
with approximately one-third of patients continuing to experience frequent seizures despite treatment
with multiple antiepileptic drugs [
We used the Epilepsy2 dataset [
The EEG signals underwent preprocessing through Fourier transformation, which enabled the decomposition of each signal into five principal frequency bands. These bands are known to reflect diferent neurological and cognitive states, and their analysis aids in highlighting patterns associated with epileptic activity: • Alpha (8–12 Hz): Typically observed during states of calm wakefulness with closed eyes, linked to a relaxed mental state. • Beta (12–30 Hz): Characteristic of focused attention and active mental engagement, often present during movement and cognitive tasks. • Gamma (30–50 Hz): Involved in complex cognitive processing, including sensory perception and higher-order brain functions.
(a) (b)
• Delta (0.5–4 Hz): Dominant during deep sleep, indicating phases of physical and neural restoration. • Theta (4–8 Hz): Associated with drowsiness, light sleep, and introspective activities such as meditation or imagination.
Since the original training set was imbalanced, with 1,630 non-epileptic and 6,432 epileptic instances, resulting in two balanced classes of 6,432 instances each. This step was crucial not only to ensure fair model training but also to prevent the majority class from dominating the generated explanations, which could lead to misleading or biased interpretations.
This phase represents the core of the modeling process, where raw EEG data are preprocessed and balanced to train an Adaptive Neuro-Fuzzy Inference System, which is employed to perform the classification task.
ANFIS is a four-layer feed-forward architecture that embeds fuzzy logic into a trainable neural network. ANFIS models the input space using Gaussian membership functions and learns a set of zero-order Takagi-Sugeno (TS) rules, where each rule associates a combination of fuzzy conditions with a constant output value. Formally, the TS rules can be expressed as:
IF ( 1 is 1 ) AND … AND ( is ) THEN ( is ), where = 1, … , represents the rule index, is the total number of rules, is the number of fuzzy variables, denotes a fuzzy set for the input variable in the -th rule, and is the fuzzy singleton defining the output class.
The system computes the degree of membership, the rule activation (using a product t-norm), and the final output through a weighted sum of the rule consequents. Parameters are optimized via backpropagation, allowing the network to adapt while maintaining a transparent, rule-based structure.
In the context of the seizure classification, for each of the five EEG features describing the data (Alpha, Beta, Gamma, Delta, Theta), ANFIS automatically generated three Gaussian fuzzy sets from the data, corresponding to the linguistic terms Low, Medium, and High, as shown in Figure 2b. The consequent of each rule is a numerical singleton representing the class label (Epileptic or Non-Epileptic). Figure 2b shows an excerpt of the fuzzy rule base, generated after training the neuro-fuzzy network.
The total number of rules in ANFIS depends on the number of fuzzy variables and the number of fuzzy terms used to describe them. In this case, the rule base consists of 243 rules, since the total is obtained by raising the number of linguistic terms to the power of the number of fuzzy variables.
After training the model and generating the fuzzy rule base, the activation degree of each rule
is computed for every test instance. These raw values are then normalized using an ℓ1 (unit-sum)
normalization [
()
∑ = 1 ()
,
where () denotes the raw activation of rule for instance , and is the total number of fuzzy
rules ( = 243 ). For each instance, the rule with the highest normalized activation is identified as the
DominantRule, following a winner-takes-all strategy [
DominantRule() = arg max ̃ ().
We identified three examples from the epileptic class and three from the non-epileptic class using k-means clustering, selecting the instances closest to the cluster centroids. This procedure yields a representative and compact subset of instances that captures the central regions of each cluster. These instances are then used in the subsequent explanation phase based on fuzzy rules.
After analyzing the data and identifying the most active fuzzy rules in key regions of the decision space, we initiated the explanatory phase. The goal of this phase is to transform the symbolic information generated by the system—such as fuzzy rules—into interpretable textual descriptions expressed in natural language.
This phase consists of two complementary steps: 1. Fuzzy rule explication, i.e., translating raw rules into a semantically meaningful and readable format, accessible also to clinicians and non-technical users; 2. Automatic explanation generation, carried out through Large Language Models instructed to provide coherent, human-centered, and context-aware descriptions.
In the first step, the fuzzy rules generated by the ANFIS model are initially expressed in a generic form, with references to numerical variables (x0, x1, …, x4) and unlabeled membership functions (mf0, mf1, mf2). To improve interpretability and clarify the semantic meaning of each rule, we introduced an explicitation phase. In this step, symbolic placeholders were reformulated by replacing the numerical variables with their corresponding EEG frequency bands (x0 = Delta, x1 = Theta, x2 = Alpha, x3 = Beta, x4 = Gamma). Similarly, the membership functions were mapped to meaningful linguistic labels, with mf0 corresponding to Low, mf1 to Medium, and mf2 to High. Finally, the output labels were made explicit by associating the vector [1.0, 0.0] with the class Non-Epileptic and [0.0, 1.0] with the class Epileptic. explanations.
For the second step, dedicated to the automatic generation of explanations, we employed
diferent LLMs and tested three prompt patterns to compare their efectiveness in producing meaningful
Prompt engineering plays a crucial role in guiding language models to produce accurate, coherent, and
contextually meaningful explanations. It involves the deliberate design of input prompts to align model
outputs with user intent and task-specific requirements [
The zero-shot prompt involves presenting the language model with a task-specific instruction without providing any in-prompt examples. As illustrated in Figure 3a, the prompt includes a brief clinical context (e.g., the patient’s condition), a statement indicating that the result was produced using a fuzzy inference system, and, most importantly, the fuzzy rule that was most activated for the given instance. This structure enables the LLM to grasp the key concepts expressed by the fuzzy rule that contribute to the decision. The model must rely exclusively on its pre-trained knowledge to interpret the prompt and generate a coherent and informative explanation.
This approach enables us to assess the model’s generalization capabilities and its ability to follow natural language instructions without requiring fine-tuning or in-context examples. The expected output is a descriptive, user-friendly explanation of the selected fuzzy rule, enriched with relevant medical knowledge—especially concerning the condition and symptoms described in the rule—thus enhancing the system’s interpretability for end-users.
In this work, we also investigate the efectiveness of the persona pattern (Figure 3b), a prompt engineering strategy in which the language model is instructed to adopt a specific role, identity, or perspective, referred to as a “persona”, to influence the tone, style, and content of its responses. By embedding a persona within the prompt, users can control the level of formality and tone, simulate domain-specific expertise, and enhance coherence and consistency, especially in multi-turn interactions. This pattern is particularly valuable in tasks involving explainability, such as the medical domain. In our setting, we instructed the language model to assume the role of a physician and generate explanations in a communication style that is more familiar and accessible to end users (e.g., healthcare professionals or patients). Our objective is to assess whether adopting this pattern leads to more efective and user-centered explanations compared to the zero-shot baseline.
Finally, we experimented with a fact-checking pattern (Figure 3c) designed to enhance the transparency of the generated output. Since language models are known to produce plausible but factually unsupported statements, this strategy extends the persona pattern, where the model adopts the role of a physician, by explicitly requiring a concise list of key scientific or clinical facts at the end of each explanation. This “fact list” allows users to verify the reliability of the information against external sources independently. Although it does not eliminate the risk of errors, this strategy introduces an additional layer of verifiability and accountability in the use of LLMs in the medical domain.
We conducted a comparative analysis of 14 LLMs, selected to represent a wide range of model families, architectural designs, and training strategies. We aimed to assess both intra-family variations—by comparing models with diferent architectures within the same family—and inter-family diferences across distinct LLM frameworks. All selected models are open-source and were chosen based on their compatibility with local execution, given the computational constraints of running experiments on a standard laptop equipped with an 11th-generation Intel Core i7 CPU, 16 GB RAM, an NVIDIA RTX 3050 GPU with 4 GB of VRAM, and a 2 TB SSD. Below is a brief overview of each model considered in the study: • LLaMA 3 (8B and 70B, 8192 context) [31]: Advanced models from Meta’s LLaMA family, optimized for extended context handling and strong performance in generative and reasoning tasks. • LLaMA 3.1-8B-Instant [31]: A latency-optimized variant designed for rapid response generation without significant loss in output quality. • LLaMA 3.3-70B-Versatile [31]: A robust, general-purpose model aimed at multitask adaptability and long-form generation coherence. • Allam-2-7B [32]: A lightweight yet expressive model, suitable for low-latency applications and edge computing scenarios. • DeepSeek R1 Distill LLaMA-70B [33]: A distilled version of the LLaMA-70B designed to reduce inference cost while maintaining accuracy, particularly in summarization and reasoning. • LLaMA-4 Maverick 17B (128e) [34]: An extended-embedding model fine-tuned for complex tasks, with emphasis on explainability and controllable outputs. • LLaMA-4 Scout 17B (16e) [35]: Tailored for eficient interpretation and classification tasks, with optimizations for linguistic clarity.
(a) (b) (c)
• Qwen-QWQ-32B [36]: A multilingual model by Alibaba, demonstrating robustness in noisy settings and solid performance on semantic tasks. • Compound-Beta [37] and Compound-Beta-Mini [38]: Experimental models focused on merging generative and logic-based components to improve transparency and interpretability. • Google Flan-T5 (Small, Base, Large) [39]: Instruction-tuned encoder-decoder models based on T5, known for their eficiency in structured tasks such as classification, translation, and explanation.
To quantitatively assess the quality of the generated explanations, we defined a multifaceted evaluation framework including linguistic, semantic, and computational aspects. The dimensions considered were lexical diversity, readability, coherence, information coverage, and generation time. All metrics were computed using the Spacy library [40], complemented with custom implementations for semantic similarity and coverage analysis.
Lexical diversity was assessed using two widely adopted indicators. The first is the Type-Token Ratio (TTR), which captures the richness of vocabulary as the proportion of unique words (types) to the total number of words (tokens) in a text: |types| TTR = . (3)
|tokens| A higher TTR reflects a more varied and informative lexical composition. To account for text length sensitivity, we also employed the Maas Index, a logarithmic transformation of TTR that adjusts for increasing token counts:
Maas = log(|tokens|) − log(|types|) (log(|tokens|))2
(4) Lower values of the Maas Index indicate greater lexical diversity, ofering a more robust assessment across explanations of varying lengths.
Readability was quantified through the Flesch Reading Ease Score (FRES), which evaluates the comprehensibility of a text based on average sentence length and syllable density. The formula is defined as:
FRES = 206.835 − (1.015 × ASL) − (84.6 × ASW), where ASL is the average sentence length and ASW is the average number of syllables per word. Higher FRES values denote more readable texts, a key factor in user-centered medical explanation.
To evaluate coherence, we adopted a metric based on cosine similarity between adjacent sentence embeddings, capturing the logical flow of information across the explanation. Given a text composed of sentences with embeddings 1, … , , the coherence score is computed as:
CohS =
1 − 1 −1 ∑ cos( , +1 ), =1 where cos( , +1 ) denotes the cosine similarity between sentence and sentence + 1 . Higher values suggest smoother transitions and a more logically consistent narrative.
We also evaluated the coverage of information through two complementary metrics. The first is the Embedding-based Coverage Score (ECS), which measures the alignment between the generated explanation and the prompt by averaging the cosine similarity between the explanation embedding and each reference sentence embedding: (5) (6) (7) (8) (9) −1 ∑ cos( text, ref ), =1 while coverage was defined as:
Diversity =
TTR + (1 − MAAS) 2
, Coverage = ECS + TCS 2 where text is the embedding of the LLM-generated output, and ref is the embedding of the -th sentence in the prompt. The second is the Token-based Coverage Score (TCS), which quantifies lexical overlap between the prompt and the generated explanation. After removing punctuation and stop words, the TCS is computed as:
TCS = | text ∩ ref|
, | text ∪ ref| where text and ref are the sets of tokens from the output and the prompt, respectively. Higher TCS values indicate that more relevant information from the prompt is preserved in the generated explanation.
Finally, we recorded the generation time, defined as the number of seconds elapsed from prompt submission to completion of the model’s output. This metric reflects computational eficiency, which is especially relevant for real-time or interactive applications.
In this section, we present the quantitative results obtained from the explanations generated by the 14 considered LLMs, using three prompting patterns across six examples (three from the positive class and three from the negative class). The evaluation relies on the measures previously described, with average values reported to provide a general overview of the performance of each model.
To facilitate the interpretation of the results, we summarized the tabular data into radar charts, highlighting the five analyzed dimensions: lexical diversity, readability, coherence, information coverage, and generation time. Lexical diversity was computed as:
All metrics were normalized to the range [
Although the primary goal of this work is not to assess the predictive capability of the classifier, we report the quantitative classification results to confirm that the model adequately captures the underlying data distribution, even on the test set. This ensures that the derived IF–THEN explanations are consistent with the information contained in the data. A data split was applied to guarantee a balanced representation of the classes and to obtain a more reliable performance evaluation. Specifically, the original dataset was partitioned into two subsets: 70% of the data was used for training, while the remaining 30% was reserved for testing. On the test set, the model achieved precision, recall, and F1-scores of 0.95 for both classes, along with an overall accuracy of 0.95, indicating that the classifier provides a faithful representation of the data.
We conducted a detailed analysis of the generated explanations by examining each quality dimension individually. Table 1 reports the results obtained with the zero-shot pattern. With respect to lexical diversity, the Flan-T5 models produced the richest and most varied vocabulary. However, this apparent advantage is ofset by their poor coherence, revealing a disconnect between lexical richness and logical structure. In contrast, instruction-tuned LLaMA variants ofered a more balanced profile, striking a balance between acceptable diversity and greater structural consistency. In terms of readability, Flan-T5 models again stand out, generating text that is particularly simple and easy to follow. However, this comes at the cost of lower semantic grounding, as reflected in their weak coherence. Many LLaMA and Compound models instead yielded lower readability, likely due to their tendency toward more technical or formal language. Coherence, understood as the logical continuity of the explanation and its alignment with the prompt, was strongest in the Compound models and some LLaMA variants, which consistently produced semantically consistent and well-structured content. Flan-T5 models, by contrast, consistently underperformed in this respect. Coverage of information, assessed through both embedding-based and token-based measures, was highest for models such as Compound and LLaMA-4 variants, which better integrated relevant concepts into their generated output. While Flan-T5 models achieved high token overlap, this often reflected surface-level repetition rather than substantive incorporation of information. Finally, the analysis of inference times highlights clear diferences in computational eficiency. Larger models generally required longer generation times, confirming the correlation between parameter size and latency. Nonetheless, some optimized LLaMA models achieved relatively low inference times despite their scale, indicating that architectural refinements and implementation strategies play a significant role. Smaller models, as expected, remained the fastest, suggesting their suitability for real-time applications.
Taken together, the results show that no single LLM clearly dominates across all evaluation dimensions. Flan-T5 models excel in readability and lexical richness but struggle with coherence. Compound and LLaMA models achieve stronger coherence and information coverage, though sometimes at the cost of readability or eficiency. Overall, LLaMA-based models appear to ofer the most balanced trade-of, whereas the optimal choice depends on which explanatory dimension is prioritized for the application context.
Table 2 reports the quantitative results obtained with the persona pattern. The analysis reveals that Flan-T5 models exhibit very high lexical diversity and readability; however, this comes at the expense of extremely low coherence and limited coverage, which undermines the overall reliability of their explanations. In contrast, the LLaMA family, including both medium- and large-scale variants, achieves a more balanced profile: their lexical diversity is slightly lower, but they compensate with stronger coherence and more stable information coverage, even though their readability remains modest. The Compound models behave similarly, ofering good coherence and coverage but with longer generation times. Qwen and Deepseek distinguish themselves for higher readability compared to most LLaMA models, although they lag in coherence and coverage.
Taken together, these results confirm that no single LLM excels across all dimensions: Flan-T5 stands out only for readability and diversity. At the same time, LLaMA- and Compound-based models appear better suited for generating coherent and semantically grounded explanations, even if their readability is less immediate. Overall, the persona pattern favors models with balanced behavior, suggesting that LLaMA variants ofer the most stable trade-of.
Table 3 presents the results obtained by combining the persona pattern with fact-checking. The analysis shows that Flan-T5 models continue to achieve the highest lexical diversity and readability. However, their performance is undermined by very low coherence and limited coverage, which reduces the reliability of the generated explanations. LLaMA-based models, particularly the larger variants, ofer a more balanced profile, maintaining moderate lexical diversity and readability while achieving stronger coherence and stable coverage of information. The Compound models also perform well in terms of coherence and coverage, but are penalized by slower inference times. Qwen and Deepseek display higher readability than most LLaMA variants, though their coherence and coverage remain weaker.
Overall, no single model clearly dominates across all dimensions; Flan-T5 stands out only in terms of diversity and readability. In contrast, LLaMA and Compound models provide more coherent and semantically grounded explanations, confirming their suitability when factual accuracy and logical consistency are prioritized.
Observing the radar charts in Figure 4, a clear diference emerges between the results obtained with the zero-shot pattern (4a) and those based on the persona pattern (4b and 4c). In the zero-shot setting, Flan-T5-large (light violet line) stands out for lexical diversity and coverage, suggesting that these models generate explanations with richer vocabulary and broader reference to the content of the prompt. Flan-T5-small and Flan-T5-base (dark violet lines) excel in readability, producing outputs that are simple and easy to follow, although this fluency does not necessarily translate into depth or precision of reasoning. By contrast, LLaMA3-8B-8192 (light blue line) achieves the best results in coherence and generation time, indicating a stronger ability to preserve logical flow across sentences while also being computationally eficient. The remaining LLMs are concentrated mainly around the coherence dimension, reflecting an efort to maintain internal consistency; however, they perform more modestly in terms of coverage, readability, and diversity, which limits the richness and accessibility of their explanations.
With the persona pattern, however, the distribution of LLMs in the radar chart changes substantially, supporting the hypothesis that diferent prompt patterns emphasize diferent explanatory qualities. Flan-T5-small (dark violet line) emerges with strong performance in time, diversity, readability, and partially in coverage, showing that it can generate accessible and lexically rich explanations with good alignment to the prompt, though still lacking in coherence. Most of the other LLMs cluster around coherence and coverage, prioritizing structured and content-grounded explanations, albeit at the expense of readability and lexical variety. LLaMA3-8B-8192 (light blue line) once again shows the highest values.
When comparing the two radars based on the persona and persona+fact‑checking patterns (Figures 4b and 4c), the distributions appear largely similar. However, it is noteworthy that Flan‑T5‑small (dark violet line) now spans almost all dimensions, achieving high or near‑maximum values in every aspect except coherence, suggesting that fact‑checking reinforces its strengths in diversity, readability, and coverage without fully remedying its weakness in logical continuity.
While the quantitative metrics outlined above provide an objective and comparable assessment of model performance, they do not fully capture the nuances of linguistic quality or the practical usability of the generated explanations. To complement and validate the numerical results, a focused qualitative analysis was conducted to gain a deeper understanding of the observed behaviors. Given the extensive amount of data, it was not feasible to examine all prompts and responses exhaustively; instead, the analysis concentrated on the most representative cases identified through the quantitative evaluation.
Figures 5a and 5b illustrate an example of a zero-shot prompt for a non-epileptic and an epileptic case, respectively. These examples will serve as references for analyzing the outputs of diferent models. Please note that we do not report the prompts generated with the other patterns, since they follow the same rule presented here and rely on the templates already introduced.
Figure 6 shows three examples of explanations generated for both epileptic and non-epileptic cases in a zero-shot setting using three diferent LLMs. Figures 6a and 6b illustrate the explanations generated for the non-epileptic and epileptic cases using LLaMA3-8b-8192 in a zero-shot learning setting. The explanations are organized into four sections: the first reports the decision produced by the fuzzy rulebased system, clarifying its meaning and context of application; the second provides a textual description of the rule that led to this decision; the third expands on this rule by analyzing the frequency band values, their respective intensities, and their relationship with the patient’s states, thereby demonstrating the model’s ability to exploit the knowledge embedded in its training; and the fourth concludes with a comprehensive summary of the reasoning process, emphasizing the key elements of the inferential outcome. This is consistent with the quantitative evaluation, which indicated a high degree of coherence.
Figures 6c and 6d show the explanations generated by Flan-T5-base. Despite the quantitative metrics indicating high readability, elevated lexical diversity, and null coherence, the explanations consist merely of repetitions of the fuzzy rules. The LLM neither contributes additional knowledge nor structures the explanation as required, resulting in poor and uninformative outputs. This excessive brevity and lack of narrative connection not only confirm the low coherence but also reveal that the observed lexical diversity stems from the use of non-repetitive technical terms rather than from genuine richness and variety of prose. Such overly concise and poor explanations highlight the model’s limitations in producing meaningful interpretative content.
Finally, Figures 6e and 6f illustrate the explanations obtained with DeepSeek-R1-Distill-LLaMA70B. In this case, the average values of the five quantitative dimensions shown in the radar chart are confirmed by the qualitative analysis. The explanations are very lengthy and include an initial informal part, where the model explicitly verbalizes the reasoning process it should follow to generate the explanation. This introduces redundancy and confusion, ultimately weakening the interpretability of the output. In the actual explanation, outside the ‘think’ tags, the fuzzy rule itself is not reported; however, references to the frequency bands and their connection to the predicted state are included, drawing on the knowledge embedded in the LLM, along with a mention of the fuzzy inference system. Without the preliminary reasoning section, the explanations would be considerably more efective. Notably, in the epileptic case, the model digresses by providing a general description of fuzzy logic itself.
From this initial analysis, it emerges that while quantitative metrics help measure the quality of explanations, they do not always align with our perception of what constitutes a “good” explanation, as exemplified by the case of Flan-T5-base.
Figure 7 presents two examples of LLM-generated explanations for non-epileptic and epileptic cases. We report the explanations obtained with LLaMA3-8b-8192 in Figures 7a and 7b, since the radar chart showed that, in addition to maintaining coherence in the zero-shot setting, this model also gains coverage when adopting the persona pattern. The structure of the explanations, however, varies between cases. In the non-epileptic explanation, the model first reports the rule, then describes the relationships between the frequency band values and the output, and even provides details on the functioning of the fuzzy inference system. In the epileptic explanation, the structure changes: the rule is not explicitly reported, and the focus shifts to a detailed account of the band values and their relation to the predicted condition, with only a brief reference to the fuzzy inference system.
We did not report the explanations generated by Flan-T5; however, despite the radar chart indicating very high values of diversity and readability, as in the previous case the explanations are extremely short and often consist only of the fuzzy rule—or in some cases not even that (e.g., “The fuzzy rule is based on a combination of EEG frequency band features”). By contrast, we included the explanations obtained with Compound-Beta-Mini, since the radar chart shows a significant increase in both coverage and coherence compared to the zero-shot setting. As illustrated in Figures 7c and 7d, the explanations are structured into two sections. The first defines the output, describes the use of the fuzzy system, reports the rule, and explains the correlations between the frequency bands and the decision. The second provides a concise explanation of what a fuzzy inference system is and why it can be helpful in the medical domain. In both examples, we also observe a notable shift in tone, which becomes more formal and aligned with medical discourse.
Finally, Figure 8 reports two examples of explanations for the epileptic case generated using the (a) (b) persona pattern with fact-checking, obtained with LLaMA3-8B-8192 and Compound-Beta-Mini. In both cases, we observe that the explanations conclude with bulleted lists that provide factual references to support the preceding statements. This strategy is intended to enhance expert trust in the analyses produced by automatic systems and, consequently, in the generated explanations.
For the sake of brevity, the explanations generated by Flan-T5 are not reported; however, in this case, the model reproduces the prompt without providing a list of references. This confirms that, despite the quantitative metrics suggesting optimal performance, the practical outcome is unsatisfactory. Similar issues were observed with DeepSeek, whose explanations again contain a section labeled ‘think’, and which is unable to generate a factual list. The same limitation applies to Qwen. LLaMA, on the other hand, does provide factual statements but fails to include references. Therefore, the integration of fact-checking with these LLMs does not efectively increase user trust in the generated explanations.
This study presented a workflow for generating structured textual explanations from IF–THEN rules produced by a neuro-fuzzy inference system, using Large Language Models as the explanatory layer. Applied to a case study on EEG-based seizure classification, the workflow enabled a systematic evaluation of 14 LLMs from diferent families and sizes under three prompting strategies: zero-shot, persona, and persona combined with fact-checking.
The results demonstrate that explanation quality cannot be fully captured by quantitative metrics alone. While Flan-T5 achieved high scores in diversity and readability, its explanations were superficial and uninformative, in contrast to LLaMA and Compound models, which ofered more coherent and content-grounded outputs at the cost of readability or eficiency. Among the prompting strategies, the persona pattern emerged as the most efective, improving both quantitative performance and qualitative richness. Fact-checking added a degree of transparency but revealed uneven support across diferent models.
Building on these findings, there is a clear need for novel quantitative metrics that go beyond surfacelevel linguistic indicators to more efectively capture coherence, informativeness, and contextual alignment. Equally important is the involvement of domain experts, both in assessing the explanations and in guiding the explanation process according to their needs, which is essential for the development of reliable, trustworthy, and human-centered explainable AI systems. Furthermore, when dealing with text generated by LLMs, an in-depth investigation of hallucinations is required to avoid the inclusion of false or misleading content—an aspect of particular relevance in the medical domain.
Gi.C. and G.Z. acknowledge the support of the project FAIR – Future AI Research (PE00000013), Spoke 6 – Symbiotic AI (CUP H97G22000210007) under the NRRP MUR program funded by NextGenerationEU. Ga.C, Gi.C. and G.V, acknowledge the support of the PNRR TT project ARIAS / M.A.M.M.A. (CUP B533D22000980006), which partially funded this research within the project FAIR - Future AI Research (PE00000013), Spoke 6—Symbiotic AI (CUP H97G22000210007) under the NRRP MUR program funded by the NextGeneration EU. A Ph.D. fellowship funds A.G.V.’s research within the Italian “D.M. n. 630, April 24, 2024” – under the NRRP, Mission 4, Component 2, Investment 3.3 – Ph.D. project “Explainability of Artificial Intelligence systems for applications in the medical field”, co-supported by Bristol-Myers Squibb s.r.l. (CUP B91I24000160007).
Generative AI tools, specifically OpenAI’s ChatGPT and Grammarly, were used exclusively for grammar correction and language refinement. The authors conceived, wrote, and validated all content. et al., The Prompt Report: A Systematic Survey of Prompting Techniques, arXiv preprint arXiv:2406.06608 (2024). [30] J. White, Q. Fu, S. Hays, M. Sandborn, C. Olea, H. Gilbert, A. Elnashar, J. Spencer-Smith, D. C.
Schmidt, A prompt pattern catalog to enhance prompt engineering with chatgpt, arXiv preprint arXiv:2302.11382 (2023). [31] G. Aaron, et al., The llama 3 herd of models, 2024. arXiv:2407.21783. [32] B. M Saiful, at al., ALLam: Large language models for arabic and english, in: The Thirteenth
International Conference on Learning Representations, 2025. [33] L. Aixin, et al., Deepseek-v3 technical report, 2025. arXiv:2412.19437. [34] Meta, Llama 4 Maverick 17B-128E-Instruct-FP8, https://www.llama.com/docs/llama-4-maverick, 2025. Version: 17B active parameters, 400B total parameters, 128 experts. [35] Meta, Llama 4: A New Foundation Model, https://www.llama.com/docs/llama-4-scout, 2025.
Version: Scout (17B). [36] Y. An, et al., Qwen3 technical report, 2025. arXiv:2505.09388. [37] Groq, Compound Beta, a Compound AI System, https://console.groq.com/docs/compound/systems/ compound-beta, 2025. Version: 2025-07-23 (Stable); Powered by Llama 4 Scout and Llama 3.3 70B. [38] Groq, Compound Beta Mini, a Compound AI System, https://console.groq.com/docs/compound/ systems/compound-beta-mini, 2025. Version: 2025-07-23 (Stable); Powered by Llama 4 Scout and Llama 3.3 70B. [39] W. Hyung, et al., Scaling instruction-finetuned language models, 2022. arXiv:2210.11416. [40] M. Honnibal, I. Montani, S. V. Landeghem, A. Boyd, spaCy: Industrial-strength Natural Language Processing in Python, 2020.