Search Results (565)

The safe integration of Large Language Models in clinical environments requires strict adherence to verified medical evidence. As part of the PARROT AI project, this study provides a systematic evaluation of how prompting strategies affect the reliability of Retrieval-Augmented Generation (RAG) pipelines using the MedQA USMLE benchmark ($N=500$). Four prompting strategies were examined: Baseline (zero-shot), Neutral, Expert Chain-of-Thought (Expert-CoT) with structured clinical reasoning, and StrictCitations with mandatory evidence grounding. The experiments covered six modern model architectures: Command R (35B), Gemma 2 (9B and 27B), Llama 3.1 (8B), Mistral Nemo (12B), and Qwen 2.5 (14B). Evaluation was conducted using the Deterministic RAG Evaluator, providing an objective assessment of grounding through the Unsupported Sentence Ratio (USR) based on TF-IDF and cosine similarity. The results indicate that structured reasoning in the Expert-CoT strategy significantly increases USR values (reaching 95–100%), as models prioritize internal diagnostic logic over verbatim context. In contrast, the StrictCitations strategy, while maintaining high USR due to the conservative evaluation threshold, achieves the highest level of verifiable grounding and source adherence. The analysis identifies a statistically significant Verbosity Signal ($r=0.81,p<0.001$), where increased response length serves as a proxy for model uncertainty and parametric leakage, a pattern particularly prominent in Llama 3.1 and Gemma 2. Overall, the findings demonstrate that prompting strategy selection is as critical for clinical reliability as model architecture. This work delivers a reproducible framework for the development of trustworthy medical AI assistants and highlights citation-enforced prompting as a vital mechanism for improving clinical safety. Full article

Background: Large Language Models (LLMs) are reshaping medical research workflows. Objective: This narrative review synthesizes evidence on LLM applications across systematic reviews, scientific writing, and clinical research. Methods: We reviewed literature from 2023–2025 examining LLM applications in medical research, identified through PubMed, Scopus, Web of Science, arXiv, medRxiv, and Google Scholar. Studies reporting empirical findings, methodological evaluations, or systematic analyses of LLM applications were included; editorials and commentaries without empirical data were excluded. Results: In systematic reviews, LLMs achieve 80–94% data extraction accuracy and 40% reduction in screening workload, but show only slight-to-moderate agreement (κ = 0.16–0.43) in risk-of-bias assessment. In scientific writing, hallucination rates of 47–55% for fabricated references and over 90% prevalence of demographic bias require rigorous verification. For clinical research, LLMs assist with statistical coding and protocol development but require human validation. Critically, excessive reliance on automated tools may cause cognitive offloading that compromises analytical capabilities. Conclusions: LLMs are powerful but unstable tools requiring constant verification. Success depends on maintaining human-in-the-loop approaches that preserve critical thinking while leveraging AI efficiency. Full article

Conventional image stitching pipelines predominantly rely on homographic alignment, whose planar assumption often breaks down in dual-camera configurations capturing non-planar scenes, producing geometric warping, bulging, and structural distortion. To address these limitations, this paper presents SENA (Seamlessly Natural), a geometry-driven image stitching approach with three complementary contributions. First, we propose a hierarchical affine-based warping strategy that combines global affine initialization, local affine refinement, and a smooth free-form deformation field regulated by seamguard adaptive smoothing. This multi-scale design preserves local shape, parallelism, and aspect ratios, thereby reducing the hallucinated distortions commonly associated with homography-based models. Second, SENA incorporates a geometry-driven adequate zone detection mechanism that identifies regions with reduced parallax directly from the disparity consistency of correspondences filtered by RANSAC, without relying on semantic segmentation or depth estimation. Third, within this zone, anchor-based seamline cutting and segmentation enforce one-to-one geometric correspondence between image pairs, reducing ghosting and smearing artifacts. Extensive experiments demonstrate that SENA achieves 26.2 dB PSNR and 0.84 SSIM, obtains the lowest BRISQUE score (33.4) among compared methods, and reduces runtime by 79% on average across resolutions. These results confirm improved structural fidelity and computational efficiency while maintaining competitive alignment accuracy. Full article

This paper defends the thesis that LLM hallucinations are best explained as a truth representation problem: Current models lack an internal representation of propositions as truth-bearers, so truth and falsity cannot constrain generation in the way factual discourse requires. It begins by surveying leading explanations—computational limits on self-verification, deficiencies in training data as truth sources, and architectural factors—and argues that they converge on the same underlying representational deficit. Next, it reconstructs the philosophical background of current LLM design, showing how optimization for fluent continuation aligns with coherence-style evaluation and with broadly structuralist, relational semantics, before turning to David Chalmers’s recent attempt to secure propositional interpretability by drawing on Davidson/Lewis-style radical interpretation and by locating propositional content in “middle-layer” structures; it argues that this approach downplays the ubiquity of hallucination and inherits instability from post-training edits. Finally, the paper offers a positive proposal: Atomic propositions should be represented in the basic vector layer, reviving a logical atomist program as a principled route to reducing hallucination. Full article

Pareidolia is often framed as a viewer-side illusion: a tendency to perceive meaningful forms—especially faces—in ambiguous inputs. This Concept Paper argues that pareidolia can also be deliberately engineered and therefore provides a tractable entry point into the neurophysiology of visual creativity. We propose a unifying construct in which engineered pareidolia functions as externally scaffolded mental imagery: minimal visual constraints recruit internally generated templates and top-down inference while remaining anchored to sensory input. To strengthen theoretical rigor, we define necessary and sufficient features that distinguish this construct from adjacent accounts (scaffolded cognition; perceptual scaffolding; bistable perception). Using Arcimboldo’s composite portraits and Dürer’s embedded face in View of the Arco Valley, plus a canonical Renaissance example (Leonardo’s Bacchus/Saint John the Baptist), we outline distinct “design regimes” that modulate cue validity, attentional release, and interpretive switching. We then connect engineered pareidolia to creativity research by linking pareidolia design and detection to measurable constructs in divergent/creative perception, including but not limited to Torrance-style domains, and we propose feasible behavioral and neurophysiological paradigms that control for artistic skill and clinical status. Finally, we distinguish benign pareidolia from hallucination, discuss clinical resonance in dementia with Lewy bodies where pareidolia can be quantified, and outline an empirically testable research program that reframes pareidolia as a bridge between imagination, perception, and creativity. Full article

Large language model-based virtual patients (LLM-VPs) are an emerging simulation tool for health professions education, but their design and integration into curricula are not well characterized. This scoping review mapped how LLM-VPs are being used for simulated clinical learning across health professions. Following a protocol registered on OSF, we searched MEDLINE, EMBASE, CENTRAL, Scopus, and Web of Science to 11 April 2025, per PRISMA-ScR guidelines, and included 21 studies that used LLMs to generate virtual patients for simulated clinical encounters. Data were extracted on technical design, fidelity domains, curricular integration, human factors, and Technology Acceptance Model constructs, and synthesized narratively. Most studies (n = 11) were pilot or feasibility evaluations with small samples (median 21) and used GPT-based models with dynamic text chat. Integration was limited to 10 studies that operated as pilots, 7 as electives, and 3 as core curricular components. The outcomes focused on Level 2 learning (clinical reasoning and preclinical OSCE performance), with predominantly self-report assessments. No studies reported Level 3 or 4 outcomes. Fidelity was strongest in cognitive, socio-cultural, and emotional domains, and 11 studies reported hallucinations or inaccurate outputs. LLM-VPs appear feasible and well-received but remain early-stage, underscoring the need for fidelity-aligned design and more rigorous, longitudinal evaluations. Full article

High-quality Fourier Transform Infrared (FTIR) imaging usually needs extensive signal averaging to reduce noise and drift, which severely limits clinical speed. Deep learning can accelerate imaging by reconstructing spectra from rapid, single-scan inputs. However, separating noise and baseline drift simultaneously without ground truth is an ill-posed inverse problem. Standard black-box architectures often rely on statistical approximations that introduce spectral hallucinations or fail to generalize to unstable atmospheric conditions. To solve these issues, we propose a physics-informed cascade Unet that separates denoising and baseline correction tasks using a new, deterministic Physics Bridge. This architecture forces the network to separate random noise from chemical signals using an embedded SNIP layer to enforce spectroscopic constraints instead of learning statistical approximations. We benchmarked this approach against a standard single Unet and a traditional Savitzky–Golay smoothing followed by SNIP baseline correction workflow. We used a dataset of human hypopharyngeal carcinoma cells (FaDu). The cascade model outperformed all other methods, achieving a 51.3% reduction in RMSE compared to raw single-scan inputs, surpassing both the single Unet (40.2%) and the traditional workflow (33.7%). Peak-aware metrics show that the cascade architecture eliminates spectral hallucinations found in standard deep learning. It also preserves peak intensity with much higher fidelity than traditional smoothing. These results show that the cascade Unet is a robust solution for diagnostic-grade FTIR imaging. It enables imaging speeds 32 times faster than current methods. Full article

Background: Precision nutrition (PN) aims to personalize dietary guidance by accounting for inter-individual variability across biological, metabolic, lifestyle, and environmental factors influencing nutritional needs and health outcomes. While traditional Artificial Intelligence (AI) has advanced nutritional research through systems like automated dietary assessment, these models often operate rigidly. Generative AI (GenAI) introduces the capacity for adaptive interventions for enhanced PN. However, the scope and maturity of its applications remain insufficiently characterized. Objective: This review examined original works applying GenAI in PN, focusing on application, methodology, and limitations. Methods: A systematic search was conducted in PubMed, ACM Digital Library, and Scopus. Inclusion criteria focused on original works deploying GenAI models in PN contexts. Included works were further formally assessed based on data used, validation, transparency, bias, and security and privacy. Results: 21 eligible studies were identified, all published after 2024. The literature indicated a surge in large language model-based systems for personalized dietary recommendations, followed by applications in data foundation building and food effect understanding. A recurrent limitation was questionable evaluation on synthetic data and hallucinations, necessitating a human-expert-in-the-loop, especially in high-stakes clinical settings. Additionally, only 4 of 21 reviewed studies incorporated biological content or biological inputs, and fewer approached biologically grounded PN within implemented personalization workflows using metabolic and/or genomic variables. Conclusions: Although GenAI research in PN is expanding rapidly, most applications remain personalized at a user-preference level rather than including biological determinants. The need for standardized reporting, stronger genome-informed modeling, and consistent human-in-the-loop validation protocols is further highlighted to advance towards holistic PN. Full article

Generating domain models from requirements is a vital and complex challenge in automated software engineering. Although large language models (LLMs) have exhibited significant competence in this area, their propensity for hallucination frequently results in models that are redundant, inconsistent, or structurally unsound. To enhance the quality of automatically generated models, this paper introduces MoRe, an LLM-based approach to domain model generation with self-refinement. Within our approach, an LLM is first tasked with producing an initial domain model draft. Subsequently, a hybrid refinement—combining LLMs with a rule-based scanner—is employed to identify and correct common issues in the model. An empirical study was conducted using 30 domain modeling problems and four open-source LLMs. The results indicate that MoRe significantly improves the quality of generated domain models. This paper advocates for incorporating a self-refinement phase as a standard component in any automated modeling workflow. Full article

Peer review is the cornerstone of scholarly publishing and, in medicine, the ultimate guarantor of the reliability of clinical evidence that informs guidelines, therapeutic strategies, and patient care. However, the current peer review system is increasingly strained by bias, abuse, and reviewer overload. Favoritism toward prominent authors, editorial “nepotism,” coercive citation practices, superficial evaluations, and even documented cases of idea theft from confidential manuscripts undermine the trustworthiness of the scientific literature upon which clinical decisions depend. In this paper, we argue that artificial intelligence (AI) and large language models (LLMs) offer a transformative opportunity to strengthen the integrity and efficiency of medical peer review. AI-driven tools can perform rapid consistency checks, detect statistical errors or plagiarism, and enforce compliance with ethical and methodological standards across thousands of manuscripts. Early implementations of AI-guided review platforms, plagiarism detectors, and citation-anomaly algorithms demonstrate that machine assistance can make reviews more thorough, objective, and reproducible. At the same time, we acknowledge the limitations of AI, including hallucinations, a lack of human judgment, and risks to confidentiality if misused. To address these concerns, we propose a hybrid model in which AI handles routine screening and technical tasks under strict safeguards, while human experts retain final responsibility for scientific evaluation. This human–AI partnership may represent an essential step toward improving the quality, fairness, and reliability of the clinical evidence base. Full article

Large language models (LLMs) hold promise for automated extraction of structured biological information from scientific literature, yet their reliability in some domain-specific tasks, such as DNA probe parsing remains underexplored. We developed a verification-focused, schema-guided extraction pipeline that transforms unstructured texts from scientific articles into a normalized database of oligonucleotide probes, primers, and associated metadata. The system combined multi-turn JSON generation, strict schema validation, sequence-specific rule checks, and a post-processing recovery module that rescues systematically corrupted nucleotide outputs. Benchmarking across nine contemporary LLMs revealed distinct accuracy–hallucination trade-offs, with the context-optimized Qwen3 model achieving the highest overall extraction efficiency while maintaining low hallucination rates. Iterative prompting substantially improved fidelity but introduced notable latency and variance. Across all models, stable error profiles and the success of the recovery module indicated that most extraction failures stem from systematic and correctable formatting issues rather than semantic misunderstandings. These findings highlight both the potential and the current limitations of LLMs for structured scientific data extraction. The research provides a reproducible benchmark and extensible framework for future large-scale curation of molecular biology datasets. Full article

Volume electron microscopy (vEM) provides nanometer-scale 3D imaging, yet its axial (z) resolution is often much lower than the in-plane ($xy$) resolution, yielding anisotropic volumes that hinder segmentation and connectomic reconstruction. We present a two-stage cross-axial super-resolution framework for isotropic reconstruction that combines a conditional diffusion model and domain-specific self-supervised pretraining of a vision transformer (ViT). First, the student–teacher self-distillation paradigm of DINOv3 is adopted to learn representations from large sets of high-resolution $xy$ sections, capturing vEM-specific texture statistics and ultrastructural patterns. Second, a conditional diffusion denoiser is trained with supervised anisotropic degradation simulated by z-downsampling, while a perceptual loss based on frozen ViT feature distances constrains generated slices to match real-section distributions. These constraints recover axial high-frequency details and reduce hallucinated textures and inter-slice drift, improving cross-slice consistency. Experiments on two public vEM datasets show improved fidelity, perceptual quality, and membrane-boundary continuity over interpolation and learning-based baselines. Full article

Green AI contributes to digital sustainability in higher education by encouraging computationally efficient technologies and responsible digital practices. Despite growing interest in sustainable AI, empirical evidence remains limited on how Gen Z students develop socially responsible intentions toward the use of sustainability-aligned AI, particularly within a single host-country higher-education context. This study examines these intentions among students enrolled in Saudi Arabia, using a culturally diverse sample of Saudi and international students while treating national origin as a demographic characteristic rather than a basis for cross-national comparison. The research also addresses emerging concerns related to AI hallucinations and ethical governance in educational settings. An integrated framework is employed that combines the instrumental appraisal logic of UTAUT with responsibility-oriented constructs. The model includes Sustainable Performance Value (SPV), Responsible Use Ease (RUE), Ethical Social Norms (ESN), Institutional Ethical Support (IES), Responsible AI Competence (RAC), AI Hallucination Awareness (AHA), and Green Digital Responsibility (GDR) as predictors of Socially Responsible Intentions (SRI). Data were collected through an anonymous survey of 1159 higher-education students residing and studying within the Saudi higher-education system. The study design reflects one institutional context rather than a multi-country comparison. The findings show strong explanatory and predictive capability (R² = 0.64; Q² = 0.43). SPV, RAC, AHA, and GDR are the strongest predictors of SRI, while RUE shows a moderate association and IES provides contextual support; ESN is not significant. The results highlight the importance of values, competence, and risk awareness in shaping the responsible use of AI. Implications focus on governance and curriculum strategies that support sustainability-aligned engagement with AI in higher education. Full article

Algorithmic management systems increasingly coordinate work, allocate resources, and support decisions in corporate, public sector, and research environments. Yet many such systems remain opaque: they optimize and score effectively but struggle to communicate rationales that are contextual, auditable, and defensible under emerging governance expectations. Large language models (LLMs) can help bridge this gap by translating quantitative signals into human-readable explanations and enabling interactive clarification. However, LLM integration also introduces new risks—hallucinated rationales, bias amplification, prompt-based security failures, and automation dependence—that must be governed rather than merely engineered. This article proposes a governance-oriented architecture for LLM-augmented algorithmic management. The model combines the following elements: an algorithmic decision core; an LLM-based cognitive interface for explanation and dialogue, and a verification and governance layer that enforces policy constraints, provenance, audit trails, and human-in-command oversight. The framework is developed through targeted conceptual synthesis and normative alignment with key governance instruments (e.g., the EU AI Act, GDPR, and ISO/IEC 42001). It is illustrated through cross-domain scenarios and complemented by a demonstrative synthetic-trace simulation that highlights transparency–latency trade-offs under verification controls. Using the demonstrative simulation (n = 120 decision events), the framework illustrates a mean baseline latency of 100.3 ms and a mean LLM-augmented latency of 115.8 ms (≈15.5% increase), a mean explanation validity proxy of 85.6%, and a simulated constraint-satisfaction rate of 94.2% (113/120 events), with failed cases routed to review. These values are presented as design-level indicators of operational plausibility and governance trade-offs, not empirical performance benchmarks or state-of-the-art comparisons. The paper contributes a conceptual and governance-oriented architectural blueprint for integrating generative AI into organisational decision systems without sacrificing accountability, compliance, or operational reliability. Full article

Generative recommendation systems based on Large Language Models leverage their reasoning capabilities to capture users’ latent interests. However, aligning continuous user behavioral embeddings with the discrete semantic space of LLMs remains a challenge. Direct alignment often leads to semantic mismatch and hallucination issues. Furthermore, existing methods typically rely on multi-stage training strategies to adapt to variations in feature distributions, thereby limiting training efficiency. To address the aforementioned issues, we propose SBT-Rec, a structured behavioral tokenization framework. Specifically, we first design a hierarchical discrete structure discovery module, utilizing a recursive residual quantization mechanism to decompose continuous behavioral vectors into discrete behavioral atoms to resolve modality discrepancies. Second, the multi-scale behavioral semantic reconstruction module reconstructs behavioral representations via residual superposition, thereby reducing data noise. Third, a residual-aware modality distribution aligner is introduced to transform behavioral features into input tokens compatible with the LLM via non-linear mapping. Finally, based on structured discrete representations, we propose a single-stage behavioral-semantic adaptive optimization strategy, achieving end-to-end parameter-efficient fine-tuning. Experiments on the MovieLens, LastFM, and Steam datasets demonstrate that SBT-Rec outperforms existing baseline models in terms of recommendation accuracy, training efficiency, and noise robustness. Full article