ABSTRACT Neuromorphic computing is emerging as a promising paradigm for sustainable edge intelligence by enabling event‐driven, low‐latency and energy‐aware computation close to sensors. However, the field remains fragmented across device technologies, mixed‐signal and digital hardware platforms, spiking neural network models, software frameworks, event‐stream processing tools, interoperability standards and benchmarking practices. This review provides a cross‐layer synthesis of contemporary neuromorphic computing platforms with particular emphasis on their relevance to scalable and sustainable edge deployment. The article organizes the neuromorphic ecosystem into interconnected layers spanning materials and devices, hardware architectures, software and interoperability tools, benchmark resources and application domains. Mixed‐signal platforms are analysed in terms of analog efficiency, accelerated neural dynamics, biological plausibility, calibration requirements, variability and reproducibility challenges. Digital neuromorphic processors are examined with respect to programmability, deterministic execution, routing fabrics, memory organization, software integration and deployment readiness. The review further discusses software frameworks, simulators, event‐data libraries, hardware‐mapping tools and intermediate representations that support model development, portability and cross‐platform evaluation. A central finding is that neuromorphic systems cannot be compared meaningfully using isolated metrics such as neuron count, chip power, latency, or throughput alone; fair evaluation requires explicit reporting of workload, event rate, model topology, mapping strategy, software stack, measurement boundary and deployment context. Accordingly, the article proposes a benchmarking and reporting perspective for neuromorphic edge intelligence that links accuracy, latency, energy efficiency, robustness and reproducibility. Thus, this review clarifies current progress, unresolved challenges and future directions for sustainable edge sensing, robotics, healthcare monitoring, smart infrastructure, industrial automation and distributed intelligent systems.

The emergence of Large Language Models (LLMs) has profoundly reshaped computational linguistics, enabling unprecedented reasoning, context awareness, and semantic understanding capabilities. Integrating these sophisticated models into Internet-of-Things (IoT) ecosystems holds transformative potential for enabling intelligent, autonomous, and contextually-aware applications. This article begins with an extensive state-of-the-art survey of existing literature on the integration of LLMs within IoT environments, establishing foundational insights into current capabilities, limitations, and deployment frameworks. Subsequently, the manuscript contributes a comprehensive analysis of lightweight LLMs and embedding models suitable for resource-constrained IoT platforms while introducing a taxonomy of sub-billion–parameter ( < 1B), mid-range (1B–2B), and exact 2B–parameter LLMs—spanning families such as Qwen, Llama, SmolLM, and IBM’s Granite—as well as embedding models under 1B parameters optimized for low-latency retrieval. Comparative assessments elucidate trade-offs in model size, inference latency, context windows, energy consumption, and performance across models categorized by parameter count. Next, a diverse spectrum of prospective use cases—including home healthcare, smart agriculture, industrial optimization, and environmental monitoring—demonstrates the practical efficacy of deploying tailored LLM-IoT frameworks for real-world problem-solving. Later, the article systematically explores key challenges that must be addressed to fully realize the integration of LLMs within IoT contexts, encompassing resource constraints, heterogeneous data processing, privacy and security risks, latency requirements, model interpretability, and ethical considerations. Finally, we outline critical directions for future research, advocating advancements in IoT-specific model architectures, multimodal sensor fusion strategies, real-time adaptive inference methods, energy-aware inference scheduling, and privacy-preserving federated learning paradigms.

Disease outbreaks and quarantine enforcement demand contact-free logistics solutions that eliminate human exposure while maintaining reliable supply delivery. This paper presents LLMQDelivery, an LLM-driven autonomous quadcopter logistics framework designed to minimize human contact during quarantine or biohazard situations through fully automated, language-guided delivery and retrieval of essential supplies. The system combines persistent spatial awareness, physicsaccurate manipulation, and natural language reasoning within a unified control architecture implemented in CoppeliaSim via a Python-based controller. A Large Language Model (LLM) serves as the cognitive interface, interpreting free-form instructions such as “Deliver a red medical kit to zone two” into executable multistage flight plans, eliminating the need for operators to physically interact with contaminated environments. Experimental validation involving eighteen supply objects achieved a 100% delivery success rate with centimeter-level placement accuracy and 98.5% natural language command success across 67 diverse prompts. The results demonstrate that fusion of deterministic motion control and stochastic language reasoning enables practical, cognitively autonomous UAV systems-representing a new paradigm for safe, scalable, and AI-augmented quarantine logistics.

Deploying large language models (LLMs) on resource-constrained edge platforms introduces strict latency, token-budget, and computational constraints that fundamentally shape pipeline behavior. In localized settings, context assembly must complete within sub-second windows while generation latency and token throughput remain bounded by hardware limitations. This article presents an empirical performance characterization of an end-to-end local LLM pipeline that integrates context retrieval, memory selection, summarization, ordering, isolation, and inference-parameter configuration within an explicitly formalized performance framework. Our contributions are fourfold. First, we formalize context engineering as a structured performance modeling problem over tunable modules {Φ _R,Φ _M,Φ _S,Φ _O,Φ _I,Φ _G} , and define measurable cost functions C_cost and G_cost that capture retrieval size, context length, sampling hyperparameters, and hardware-related generation effects. This formulation provides a quantitative lens for analyzing trade-offs among latency, token usage, and computational overhead. Second, we implement and instrument a modular pipeline on four representative local models–granite3.1-moe, llama3.2:1b, gemma3:1b, and qwen3:1.7b–and report detailed runtime measurements. End-to-end latency varies from 5.62 s (granite3.1-moe) to 14.12 s (qwen3:1.7b), prompt evaluation time ranges from 0.206 s to 0.507 s , and token throughput spans 50.3 tokens/s down to 30.2 tokens/s . Context-processing micro-stages execute in under μs each, establishing generation as the dominant performance bottleneck. Third, we conduct comprehensive statistical analyses–including analysis of variance (ANOVA), eta-squared effect size estimation, principal component analysis (PCA), and mutual information–to quantify model-dependent performance variation. Results confirm statistically significant model effects on generation metrics (e.g., F = 198.5 , p < 10^-29 for throughput), while revealing negligible variance across micro-context stages. Fourth, we demonstrate that Gaussian mixture clustering of execution traces reveals distinct “context-dominated” and “generation-dominated” regimes. These empirically observed operating modes provide structural insight into performance behavior and offer a foundation for future adaptive scheduling and resource-aware control strategies. In addition, response correctness across simulated evaluation prompts is assessed using standard classification metrics such as precision, recall, and F1-score, providing an auxiliary measure of output reliability under constrained inference settings. Collectively, the findings establish a quantitative framework for analyzing the interaction between context engineering strategies and generation efficiency in localized LLM deployments. The code is available at https://github.com/ParthaPRay/contxt_engineering_large_language_model .

We present HIAPLLM, a privacy-preserving hybrid edge–cloud framework for real-time air-quality advisories tailored to the Indian context using Large Language Models (LLMs). At the network edge, a Raspberry Pi 4 B hosts a Gradio/Folium interface enabling users to input any Indian geocoordinates. Upon each request, the Pi issues a 5 s-timeout call to the Open-Meteo application programming interface (API), retrieving PM_2.5 (5.2–82.9 g/m^3 ), PM_10 (8.6–172.8 g/m^3 ), CO (0.10–0.80 mg/m^3 ), NO_2 (3.6–46.5 g/m^3 ), SO_2 (0.3–40.7 g/m^3 ), O_3 (35–160 g/m^3 ), aerosol optical depth (AOD: 0.14–1.10), dust loading (0–238 units), and CH_4 (1 316–1 614 ppm). These concentrations are mapped via piecewise linear breakpoints into air quality index (AQI) sub-indices and qualitative bands. The Pi then constructs a structured LLM prompt (e.g., “ PM_2.5 = 82.9 g/m^3 ; Category: Poor”) and forwards it over LAN to a laptop running ten quantized LLMs (0.6–7 B parameters) under the Ollama engine, eliminating external cloud dependency and preserving raw location privacy. We profile model load times (0.80–2.46 s) and inference latencies (6.26–94.05 s), and find a strong inverse relationship between token throughput and total latency (slope − 0.516 s per token/s, R^2=0.552 ). Multivariate analyses (Principal Component Analysis (PCA) capturing 42 PM_10 : 8.75× 10^3 , dust: 3.66× 10^3 , total_duration_s: 3.84× 10^10 ) expose extreme collinearity, addressed via dimension reduction. HIAPLLM’s runtime of O(N+P)+T_LLM and space O(N+P+m) ensures scalable, transparent, and low-cost deployment suitable for India’s connectivity-constrained, privacy-sensitive settings. Code and data are available at https://github.com/ParthaPRay/Air_Pollution_LLM_Advisor .

Deploying large language models (LLMs) on resource-constrained edge devices remains a significant systems challenge due to limited CPU capacity, restricted memory, absence of hardware accelerators, and the need for low-latency, privacy-preserving operation without cloud connectivity. Existing approaches primarily emphasize model compression, energy profiling, or hybrid cloud–edge architectures, but lack systematic evaluation of end-to-end LLM behavior when exposed as concurrent, web-accessible services on single-device edge platforms. In particular, the interaction between quantized LLM inference, asynchronous web serving, and request-level concurrency remains poorly understood. This paper presents SanicLLMPiAB, a fully localized and internet-independent framework for evaluating quantized LLM inference on a Raspberry Pi 4B using an asynchronous Sanic web server, the Ollama local LLM runtime, and Apache Bench–driven stress testing. Unlike prior model-centric studies, the proposed framework enables fine-grained, per-request observability by capturing token-level execution metrics, end-to-end latency, CPU utilization, memory stability, and accuracy under controlled concurrency. All inference is executed locally, ensuring data privacy and eliminating network variability. Experimental evaluation of four quantized LLMs (Qwen 0.5B, TinyLLaMA, TinyDolphin, and StableLM2) demonstrates that Qwen 0.5B achieves the best balance of scalability and robustness, sustaining stable throughput (11–12 tokens/s), sub-second end-to-end latency at higher concurrency, and 100 https://github.com/ParthaPRay/Sanic-Ollama-RaspberryPi-API .

The expanding role of intelligent systems in biomedical science marks a shift from passive analysis towards active participation in discovery and care. Recent scholarship has begun to frame these systems not merely as models, but as agents capable of planning, interaction, and adaptation. This letter reflects on such developments, acknowledging their conceptual clarity and practical ambition, while raising questions about evaluation, responsibility, human judgement, and long-term scientific culture. The intent is to encourage careful reflection as these technologies move closer to real-world integration.

This letter comments on the recently published AICellType platform for large language model (LLM)-based cell type annotation in single-cell and spatial transcriptomics. While appreciating the authors' systematic benchmarking and practical contribution, concerns are raised regarding the continued dependence on proprietary commercial LLMs such as Claude 3.5 Sonnet for biomedical annotation. Greater emphasis is suggested on open-source biomedical LLMs, multimodal vision-language models, local deployment, reproducibility, privacy preservation, and regulatory compliance to ensure more transparent, reliable, and sustainable medical annotation systems for translational bioinformatics.

ReliaPy is an open-source Python/Gradio workbench designed to make reliability analysis accessible, transparent, reproducible, and publicationoriented. It integrates four major workflows: life-data analysis, reliability-growth modeling, repairable-system analysis, and accelerated life testing. Users can upload CSV or Excel data, apply Weibull, lognormal, Crow–AMSAA, power-law, log-linear, piecewise nonhomogeneous Poisson process, Arrhenius, and power-law stress models, and obtain parameter tables, diagnostic plots, engineering interpretations, and downloadable CSV and high-resolution PNG outputs. The platform supports censored data, automatic column inference, change-point detection, recurrent-event visualization, and use-stress life prediction. Its modular architecture separates data processing, computation, presentation, and export, enabling deployment in Google Colab, Hugging Face Spaces, or local environments. Validation against a NIST censored life-test dataset produced Weibull estimates differing by less than 0.005% from reference values. ReliaPy supports students, researchers, and small engineering laboratories. It is intended for education, rapid analysis, reproducible demonstrations, and early-stage reliability method development rather than replacing expert engineering judgment.

This letter responds to the important work of Bastaninejad et al., highlighting how advanced large language models (LLMs) such as ChatGPT and Gemini perform in the emotionally and technically demanding landscape of revision rhinoplasty consultations. While the study demonstrates the remarkable communicative, empathetic, and informational strengths of general-purpose LLMs, it also exposes their probabilistic limitations and the absence of domain-specific safeguards required for high-stakes surgical decision making. The findings underscore the urgent need for specialized Surgical-LLMs and Bio-LLMs—ethically aligned, clinically validated, and fine-tuned on high-quality operative, anatomical, and perioperative datasets. Future research should expand question complexity, include multi-turn dialog, and diversify evaluators. With responsible development, multimodal, privacy-preserving surgical LLM ecosystems could meaningfully augment pre-consultation education, risk communication, and patient support. Level of Evidence V This journal requires that authors assign a level of evidence to each article. For a full description of these Evidence-Based Medicine ratings, please refer to the Table of Contents or the online Instructions to Authors www.springer.com/00266 .

Existing AI risk assessment frameworks, developed for static predictive models, struggle to address the dynamic, tool-driven failures emerging from autonomous agents and foundation models. Unanticipated harms—ranging from biased decision logic and privacy breaches to adversarial exploits and unsustainable resource consumption—highlight a critical need for a methodology that spans end-to-end, multi-stage pipelines. In response, this article introduces a unified five-domain taxonomy of risks—agentic operation, training data risks, inference behavior risks, output risks, and non-technological factors—each further classified into traditional AI vulnerabilities, those amplified by generative capabilities, and novel hazards unique to generative agents as inspired by IBM AI risk atlas. Through detailed case studies—covering function-calling hallucinations, dynamic bias amplification, privacy leakage across toolchains, and edge-deployment constraints—we examine how specific failure modes arise, interact, and propagate. Building on this analysis, we characterize ten core operational challenges and recommend targeted monitoring and mitigation strategies. Finally, we propose twelve future research and engineering directions—spanning adaptive fairness loops, federated privacy collaboration, schema-guided invocation, policy-as-code governance, and Green AI practices—to guide the design of resilient, transparent, and sustainable agentic systems. Our framework equips researchers, practitioners, and policymakers with a structured foundation for systematic risk assessment, targeted intervention, and continual oversight.

The integration of artificial intelligence into web browsers is transforming traditional browsing from passive navigation into active, intent-driven interaction. Recent advancements in large language models (LLMs), agent-based orchestration, and retrieval-augmented systems have enabled browsers to perform complex tasks such as conversational search, automated workflows, and context-aware decision support. However, existing research remains fragmented, focusing on isolated aspects such as extensions, security, or specific applications, with no unified understanding of AI-powered browser architectures and system-level trade-offs. This paper presents a comprehensive survey of AI-powered web browsers, addressing this gap by systematically analyzing their architectural design, operational workflow, and functional capabilities. We propose a layered reference architecture capturing key components including interaction interfaces, orchestration mechanisms, multi-model inference, retrieval pipelines, and persistent memory. Furthermore, we examine the end-to-end workflow that transforms user intent into executable actions and structured outputs. A detailed taxonomy of emerging platforms is developed, followed by a multi-dimensional comparative analysis across autonomy, privacy, productivity, and deployment strategies. Our analysis reveals significant trade-offs between local, hybrid, and cloud-assisted architectures, as well as variations in agentic capabilities and system specialization. Despite rapid progress, critical challenges remain, including reliability issues due to LLM hallucination, security risks from autonomous execution, privacy concerns in cloud-based systems, and the lack of standardized evaluation frameworks. This survey establishes a structured foundation for understanding AI-powered browsers and highlights key research directions toward building reliable, privacy-preserving, and scalable systems.

Edge deployments of large language models (LLMs) often suffer from significant latency due to the overhead of high-level client runtimes on resource-constrained hardware. To address this challenge, we conducted a side-by-side performance analysis of four quantised LLMs – Llama 3.2:1b, Gemma 3:1b, Granite 3.1-MoE:1b, and Qwen 2.5:0.5b – on a Raspberry Pi 4 Model B (8 GB LPDDR4, quad-core ARM Cortex-A72) using both Python and Rust API clients. Each model was served via a local Ollama inference server, and a fixed suite of twenty prompts – covering factual retrieval, arithmetic reasoning, translation, code synthesis, and creative generation – was executed sequentially with a two-second inter-request delay, yielding 160 measurements per client. Rust markedly reduces cold-start delays: mean model load times fall from 1 648.7 ms (Python) to 52.8 ms (Rust) for Llama 3.2:1b, and from 607.0 ms to 171.3 ms for Qwen 2.5:0.5b. Corresponding end-to-end latencies decrease by 1.4-2.0 s across models. In warm-start conditions, both clients deliver nearly identical decoding throughput – ≈2.7 tokens/s for Llama 3.2:1b, 4.4 tokens/s for Gemma 3:1b, 7.4 tokens/s for Granite 3.1-MoE, and 8.6 tokens/s for Qwen 2.5:0.5b – indicating that runtime overhead is negligible once models are loaded. Rigorous statistical testing, including paired t-tests, Mann-Whitney U tests, and bootstrap confidence intervals, confirms that Rust’s coldstart advantages are highly significant (p < 0.01). At the same time, throughput differences in steady-state inference are not statistically meaningful. We discuss limitations in platform specificity, quantisation approaches, and prompt diversity, and outline future work on heterogeneous accelerators, adaptive scheduling, and ondevice fine-tuning. Finally, we highlight practical applications in smart agriculture, healthcare monitoring, industrial IoT, autonomous robotics, and offline educational tools. This benchmark furnishes actionable guidelines for selecting client languages and quantised models in edge AI scenarios.

The emergence of agentic generative artificial intelligence (GenAI) systems marks a fundamental shift in the reliability landscape of AI-enabled systems. Unlike traditional predictive models that advise human decision-makers, agentic GenAI systems integrate large language models with planning logic, memory, and tool invocation capabilities, enabling them to autonomously select and execute actions within operational environments. This transition from analysis to action introduces a new class of reliability risks, where probabilistic reasoning directly affects system state, resource utilization, data integrity, and regulatory compliance. Conventional model-centric reliability metrics—such as accuracy and robustness—are insufficient for evaluating systems that operate across continuous decision–action loops. This article advances a reliability-by-design framework for agentic GenAI, drawing on the AI Risk Atlas and aligning engineering practices with the lifecycle obligations of the European Union Artificial Intelligence Act (EU AI Act). It reframes agentic risks—including epistemic overconfidence, tool-calling hallucinations, loss of traceability, privacy leakage, and runaway feedback loops—through a system-level reliability lens. The proposed approach emphasizes structured hazard analysis, operational boundary definition, action provenance logging, graduated autonomy, and continuous monitoring. Further, the article outlines measurable, action-oriented reliability metrics tailored to acting systems and presents a practical blueprint for constructing an EU AI Act-ready assurance case.

ABSTRACT Artificial intelligence (AI) is increasingly reshaping ophthalmology because the specialty depends heavily on structured imaging, quantitative measurements, and repeatable diagnostic workflows. This review provides a clinically grounded and translationally oriented synthesis of AI in ophthalmology, covering methodological foundations, ophthalmic imaging modalities, public datasets, disease‐specific applications, evaluation metrics, deployment barriers, and future directions. Unlike reviews that mainly summarize algorithmic performance by disease category or model type, this article organizes ophthalmic AI through an integrated framework that emphasizes clinical use cases, evidence maturity, translational readiness, and real‐world implementation requirements. The review examines applications across population screening, referral triage, disease grading, progression monitoring, prognosis, treatment guidance, workflow support, and automated reporting. Major disease domains include diabetic retinopathy, glaucoma, age‐related macular degeneration, cataract, infectious keratitis, and keratoconus. Particular attention is given to the distinction between retrospective proof‐of‐concept studies, external validation, multicenter evaluation, prospective trials, and real‐world deployment. The review also interprets evaluation metrics from a clinical perspective, highlighting the importance of threshold selection, sensitivity, specificity, false referral burden, missed disease, calibration, uncertainty, segmentation adequacy, robustness, and generalization. Key translational challenges include dataset bias, domain shift, interpretability, privacy, regulatory oversight, infrastructure constraints, workflow integration, and post‐deployment monitoring. Emerging paradigms such as multimodal AI, foundation models, generative AI, and edge‐based point‐of‐care systems are discussed cautiously, with emphasis on hallucination risk, clinical grounding, accountability, and the gap between benchmark performance and deployment readiness. Overall, the review argues that the next phase of ophthalmic AI should move beyond high accuracy values toward prospective validation, external generalization, clinician‐centered design, calibrated uncertainty, and accountable integration into real‐world eye‐care pathways.

This correspondence responds to the recent commentary on my article proposing a transparent, hybrid generative AI framework for patient selection in cosmetic surgery. The commentary rightly emphasizes the importance of explicit task specification, external and temporal validation, and clear threshold-to-action mapping to ensure safe and clinically meaningful deployment. I elaborate on how reasoning-capable large language models, specialty medical models, and retrieval-augmented generation pipelines can produce auditable, guideline-anchored suitability assessments, while acknowledging the need for stronger calibration, stratified reporting, and workflow-linked decision pathways. I also affirm the necessity of regulatory rigor, independent validation, privacy safeguards, and bias monitoring as prerequisites for real-world adoption. This exchange highlights a shared commitment to developing calibrated, ethical, and clinically respectful AI systems that enhance surgical judgment, protect patients, and support proportionate, evidence-aligned care in aesthetic practice. Level of Evidence V This journal requires that authors assign a level of evidence to each article. For a full description of these Evidence-Based Medicine ratings, please refer to the Table of Contents or the online Instructions to Authors www.springer.com/00266 .