Knowledge distillation has become a standard technique for compressing large language models into efficient student models, but existing methods often struggle to balance prediction accuracy with explanation quality. Recent approaches such as Distilling Step-by-Step (DSbS) introduce explanation supervision, yet they apply it in a uniform manner that may not fully exploit the different learning dynamics of prediction and explanation. In this work, we propose a task-structured curriculum learning (TSCL) framework that structures training into three sequential phases: (i) prediction-only, to establish stable feature representations; (ii) joint prediction–explanation, to align task outputs with rationale generation; and (iii) explanation-only, to refine the quality of rationales. This design provides a simple but effective modification to DSbS, requiring no architectural changes and adding negligible training cost. We justify the phase scheduling with ablation studies and convergence analysis, showing that an initial prediction-heavy stage followed by a balanced joint phase improves both stability and explanation alignment. Extensive experiments on five datasets (e-SNLI, ANLI, CommonsenseQA, SVAMP, and MedNLI) demonstrate that TSCL consistently outperforms strong baselines, achieving gains of +1.7–2.6 points in accuracy and 0.8–1.2 in ROUGE-L, corresponding to relative error reductions of up to 21%. Beyond lexical metrics, human evaluation and ERASER-style faithfulness diagnostics confirm that TSCL produces more faithful and informative explanations. Comparative training curves further reveal faster convergence and lower variance across seeds. Efficiency analysis shows less than 3% overhead in wall-clock training time and no additional inference cost, making the approach practical for real-world deployment. This study demonstrates that a simple task-structured curriculum can significantly improve the effectiveness of knowledge distillation. By separating and sequencing objectives, TSCL achieves a better balance between accuracy, stability, and explanation quality. The framework generalizes across domains, including medical NLI, and offers a principled recipe for future applications in multimodal reasoning and reinforcement learning.

Large Language Models (LLMs) have significantly advanced human-computer interaction by improving natural language understanding and generation. However, their vulnerability to adversarial prompts–carefully designed inputs that manipulate model outputs–presents substantial challenges. This paper introduces a classification-based approach to detect adversarial prompts by utilizing both prompt features and prompt response features. Eleven machine learning models were evaluated based on key metrics such as accuracy, precision, recall, and F1-score. The results show that the Convolutional Neural Network–Long Short-Term Memory (CNN-LSTM) cascade model delivers the best performance, especially when using prompt features, achieving an accuracy of over 97% in all adversarial scenarios. Furthermore, the Support Vector Machine (SVM) model performed best with prompt response features, particularly excelling in prompt type classification tasks. Classification results revealed that certain types of adversarial attacks, such as “Word Level” and “Adversarial Prefix”, were particularly difficult to detect, as indicated by their low recall and F1-scores. These findings suggest that more subtle manipulations can evade detection mechanisms. In contrast, attacks like “Sentence Level” and “Adversarial Insertion” were easier to identify, due to the model’s effectiveness in recognizing inserted content. Natural Language Processing (NLP) techniques played a critical role by enabling the extraction of semantic and syntactic features from both prompts and their corresponding responses. These insights highlight the importance of combining traditional and deep learning approaches, along with advanced NLP techniques, to build more reliable adversarial prompt detection systems for LLMs.

The purpose of this study is to develop a reliable method for distinguishing between AI-generated, paraphrased, and human-written texts, which is crucial for maintaining the integrity of research and ensuring accurate information flow in critical fields such as healthcare. To achieve this, we propose HybridGAD, a novel hybrid model that combines Long Short-Term Memory (LSTM), Bidirectional LSTM (Bi-LSTM), and Bidirectional Gated Recurrent Unit (Bi-GRU) architectures with an attention mechanism. Our methodology involves training this hybrid model on a dataset of radiology abstracts, encompassing texts generated by AI, paraphrased by AI, and written by humans. The major findings of our analysis indicate that HybridGAD achieves a high accuracy of 98%, significantly outperforming existing state-of-the-art models. This high performance is attributed to the model’s ability to effectively capture the contextual nuances and structural differences between AI-generated and human-written texts. In conclusion, HybridGAD not only enhances the accuracy of text classification in the field of radiology but also paves the way for more advanced medical diagnostic processes by ensuring the authenticity of textual information. Future research will focus on integrating textual and visual data for comprehensive radiology assessments and improving model generalization with partially labeled data. This study underscores the potential of HybridGAD in transforming medical text classification and highlights its applicability in ensuring the integrity and reliability of research in healthcare and beyond.