Frontiers in Artificial Intelligence最新文献

Introduction: Local interpretability methods such as LIME and SHAP are widely used to explain model decisions. However, they rely on assumptions of local continuity that often fail in recursive, self-modulating cognitive architectures.

Methods: We analyze the limitations of local proxy models through formal reasoning, simulation experiments, and epistemological framing. We introduce constructs such as Modular Cognitive Attention (MCA), the Cognitive Leap Operator (Ψ), and the Internal Narrative Generator (ING).

Results: Our findings show that local perturbations yield divergent interpretive outcomes depending on internal cognitive states. Narrative coherence emerges from recursive policy dynamics, and traditional attribution methods fail to capture bifurcation points in decision space.

Discussion: We argue for a shift from post-hoc local approximations to embedded narrative-based interpretability. This reframing supports epistemic transparency in future AGI systems and aligns with cognitive theories of understanding.

The use of Large Language Models (LLMs) such as ChatGPT is a prominent topic in higher education, prompting debate over their educational impact. Studies on the effect of LLMs on learning in higher education often rely on self-reported data, leaving an opening for complimentary methodologies. This study contributes by analysing actual course grades as well as ratings by fellow students to investigate how LLMs can affect academic outcomes. We investigated whether using LLMs affected students' learning by allowing them to choose one of three options for a written assignment: (1) composing the text without LLM assistance; (2) writing a first draft and using an LLM for revisions; or (3) generating a first draft with an LLM and then revising it themselves. Students' learning was measured by their scores on a mid-course exam and final course grades. Additionally, we assessed how the students rate the quality of fellow students' texts for each of the three conditions. Finally we examined how accurately fellow students could identify which LLM option (1-3) was used for a given text. Our results indicate only a weak effect of LLM use. However, writing a first draft and using an LLM for revisions compared favourably to the 'no LLM' baseline in terms of final grades. Ratings for fellow students' texts was higher for texts created using option 3, specifically regarding how well-written they were judged to be. Regarding text classification, students most accurately predicted the 'no LLM' baseline, but were unable to identify texts that were generated by an LLM and then edited by a student at a rate better than chance.

Effective diabetes care relies on communication, patient empowerment, and lifestyle management. However, rising prevalence and workforce shortages challenge current care models. Large language models (LLMs) have the potential to support healthcare delivery by providing personalized health information. While prior studies show promising results, few have compared LLM-generated responses with those from healthcare professionals in chronic disease contexts, particularly from end-users' perspectives. This study compared GPT-4o and healthcare professional responses to diabetes-related questions, evaluating them on knowledge, helpfulness, and empathy. It also explored correlations between these qualities and differences based on participants' educational background. Using a cross-sectional experimental design, 1,810 evaluations were collected through an online questionnaire (November 2024-January 2025). Participants rated responses on 5-point Likert scales for knowledge, helpfulness, and empathy. For all metrics combined, GPT-4o received higher ratings in 46.7% of evaluations (95% CI: 28.8%-64.5%), while healthcare professionals were preferred in 23.3% (95% CI: 8.2%-38.5%). Participants with lower education levels rated GPT-4o significantly higher across all dimensions, while those with ≥4 years of higher education rated it higher for empathy and helpfulness. Quality measures were strongly correlated. Although differences were statistically significant, the observed effect sizes were small and should be interpreted as modest in practical terms. These findings assess perceived quality and accessibility of healthcare communication from end-user perspectives and suggest that LLMs may enhance the perceived quality and accessibility of healthcare communication, particularly among individuals with lower educational attainment. Further research is needed to determine their appropriate role in clinical practice, including objective assessment of clinical accuracy.

Summarization of texts have been considered as essential practice nowadays with the careful presentation of the main ideas of a text. The current study aims to provide a methodology of summarizing complex texts such as argumentative discourse. Extractive and abstractive summarization techniques have recently gained significant attention. Each has its own limitations that reduce efficiency in the coverage of the main points of the summary, but by combining them, we can use the positive points of each to improve both summarization performance and summary generation quality. This paper presents a novel extractive-abstractive text summarization method that ensures coverage of the main points of the entire text. It is based on combining Bidirectional Encoder Representations from Transformers (BERT) and transfer learning. Using a dataset comprising two UK parliamentary debates, the study shows that the proposed method effectively summarizes the main points. Comparing extractive and abstractive summarization, the experiment used Recall-Oriented Understudy for Gisting Evaluation (ROUGE) sets of metrics and achieved scores of 30.1, 9.60, and 27.9 for the first debate, and 36.2, 11.80, and 31.5 for the second, using ROUGE-1, ROUGE-2, and ROUGE-L metrics, respectively.

Understanding how large language models (LLMs) process emotionally sensitive content is critical for building safe and reliable systems, particularly in mental health contexts. We compare the performance of LLMs of different sizes on two key tasks: trinary classification of emotional safety (safe vs. unsafe vs. borderline) and multi-label classification using a six-category safety risk taxonomy. To support this, we construct a novel dataset by merging several human-authored mental health datasets (> 15K samples) and augmenting them with emotion re-interpretation prompts generated via ChatGPT. We evaluate four LLaMA models (1B, 3B, 8B, 70B) across zero-shot and few-shot settings. Our results show that larger LLMs achieve stronger average performance, particularly in nuanced multi-label classification and in zero-shot settings. However, lightweight fine-tuning allowed the 1B model to achieve performance comparable to larger models and BERT in several high-data categories, while requiring < 2GB VRAM at inference. These findings suggest that smaller, on-device models can serve as viable, privacy-preserving alternatives for sensitive applications, offering the ability to interpret emotional context and maintain safe conversational boundaries. This work highlights key implications for therapeutic LLM applications and the scalable alignment of safety-critical systems.

The growing interest in utilizing clinical blood biomarkers for non-invasive diagnostics has transformed the approach to early detection and prognosis of respiratory diseases. Biomarker-driven diagnostics offer cost-effective, rapid, and scalable alternatives to traditional imaging and clinical assessments. In this study, we conducted a retrospective analysis of 913 patients from a local respiratory clinic in Hail region, evaluating the diagnostic relevance of 15 blood biomarkers across four respiratory conditions: COVID-19, pneumonia, asthma, and other complications. Through data-driven analysis, statistical correlation assessments, and machine learning classification models (decision tree classifiers), we identified significant biomarker interactions that contributed to disease differentiation. Notably, CRP and HGB demonstrated a strong negative correlation (-55%), supporting the well-established role of systemic inflammation in anemia of chronic disease. Additionally, Ferritin and LDH exhibited a positive correlation (+50%), indicating metabolic stress and cellular injury in severe respiratory illnesses. Other significant correlations included Creatinine and ESR being negatively associated with RBC, while GGT and ALT were positively correlated (+49%). Additionally, bilirubin and HGB were positively correlated (+49%), collectively reflecting systemic inflammatory and metabolic responses associated with respiratory pathology. The machine learning model demonstrated high predictive accuracy, with the following performance metrics: COVID-19: Precision (0.94), Recall (0.96), F1-score (0.95). Pneumonia: Precision (0.97), Recall (0.71), F1-score (0.85). Asthma: Precision (1.00), Recall (0.95), F1-score (0.97). Other Complications: Precision (0.88), Recall (0.90), F1-score (0.90). These findings validate the diagnostic potential of biomarker panels in respiratory disease classification, offering a novel approach to integrating statistical and computational modeling for clinical decision-making. By leveraging biomarker relationships and machine learning algorithms, this study contributes to the development of personalized, non-invasive, and cost-effective diagnostic tools for respiratory diseases, ultimately improving patient outcomes and healthcare efficiency.

Accurate prediction of human crowd behavior presents a significant challenge with critical implications for autonomous systems. The core difficulty lies in developing a comprehensive computational framework capable of effectively modeling the spatial-temporal dynamics through three essential components: feature extraction, attention propagation, and predictive modeling. Current spatial-temporal graph convolutional networks (STGCNs), which typically employ single-hop neighborhood message passing with optional self-attention mechanisms, exhibit three fundamental limitations: restricted receptive fields due to being confined to limited propagation steps, poor topological extensibility, and structural inconsistencies between network components that collectively lead to suboptimal performance. To address these challenges, we establish the theoretical connection between graph convolutional networks and personalized propagation neural architectures, thereby proposing attention diffusion-prediction network (ADP-Net). This novel framework integrates three key innovations: (1) Consistent graph convolution layers with immediate attention mechanisms; (2) Multi-scale attention diffusion layers implementing graph diffusion convolution (GDC); and (3) Adaptive temporal convolution modules handling multi-timescale variations. The architecture employs polynomial approximation for GCN operations and implements an approximate personalized propagation scheme for GDC, enabling efficient multi-hop interaction modeling while maintaining structural consistency across spatial and temporal domains. Comprehensive experiments on standardized benchmarks (ETH/UCY and Stanford Drone Dataset) show cutting-edge results, with enhancements of 4% for the average displacement error (ADE) and 26% for the final displacement error (FDE) metrics when contrasted with prior approaches. This advancement provides a robust theoretical framework and practical implementation for crowd behavior modeling in autonomous systems.

The proliferation of digital platforms has enabled fraudsters to deploy sophisticated camouflage techniques, such as multi-hop collaborative attacks, to evade detection. Traditional Graph Neural Networks (GNNs) often fail to capture these complex high-order patterns due to limitations including homophily assumption failures, severe label imbalance, and noise amplification during deep aggregation. To address these challenges, we propose the Hypergraph-based Contrastive Learning Network (HCLNet), a novel framework integrating three synergistic innovations. Firstly, multi-relational hypergraph fusion encodes heterogeneous associations into hyperedges, explicitly modeling group-wise fraud syndicates beyond pairwise connections. Secondly, a multi-head gated hypergraph aggregation mechanism employs parallel attention heads to capture diverse fraud patterns, dynamically balances original and high-order features via gating, and stabilizes training through residual connections with layer normalization. Thirdly, hierarchical dual-view contrastive learning jointly applies feature masking and topology dropout at both node and hyperedge levels, constructing augmented views to optimize self-supervised discrimination under label scarcity. Extensive experiments on two real-world datasets demonstrate HCLNet's superior performance, achieving significant improvements over the baselines across key evaluation metrics. The model's ability to reveal distinctive separation patterns between fraudulent and benign entities underscores its practical value in combating evolving camouflaged fraud tactics in digital ecosystems.

Body Dysmorphic Disorder (BDD) is increasingly recognized in the aesthetic practice, yet it remains underdiagnosed and often misunderstood. With its high prevalence, particularly in cosmetic consultations, BDD poses significant ethical and clinical challenges. Aesthetic providers must be vigilant in identifying at-risk individuals and prioritizing psychological well-being alongside procedural outcomes. Artificial Intelligence (AI), with its capacity to analyze behavioral patterns, automate screening tools, and detect subtle indicators of cognitive distortion, presents a new frontier in managing BDD. However, integrating AI into clinical practice requires caution to prevent reinforcing appearance-focused biases and to ensure privacy and fairness. This commentary discusses the opportunities, limitations, and ethical considerations of leveraging AI to assist clinicians in detecting BDD, fostering safer patient outcomes, and advancing the compassionate practice of aesthetic medicine. AI should not accelerate aesthetic procedures but promote reflective, ethically sound decision-making. When integrated responsibly, it can enhance recognition of BDD, support psychological safety, and preserve patient trust through transparency, data protection, and clinician oversight.