Large Language Models (LLMs), trained on extensive datasets using advanced deep learning architectures, have demonstrated remarkable performance across a wide range of language tasks, becoming a cornerstone of modern AI technologies. However, ensuring their trustworthiness remains a critical challenge, as reliability is essential not only for accurate performance but also for upholding ethical, cultural, and social values. Careful alignment of training data and culturally grounded evaluation criteria are vital for developing responsible AI systems. In this study, we introduce the EPT (Evaluation of Persian Trustworthiness) metric, a culturally informed benchmark specifically designed to assess the trustworthiness of LLMs across six key aspects: truthfulness, safety, fairness, robustness, privacy, and ethical alignment. We curated a labeled dataset and evaluated the performance of several leading models - including ChatGPT, Claude, DeepSeek, Gemini, Grok, LLaMA, Mistral, and Qwen - using both automated LLM-based and human assessments. Our results reveal significant deficiencies in the safety dimension, underscoring the urgent need for focused attention on this critical aspect of model behavior. Furthermore, our findings offer valuable insights into the alignment of these models with Persian ethical-cultural values and highlight critical gaps and opportunities for advancing trustworthy and culturally responsible AI. The dataset is publicly available at: https://github.com/Rezamirbagheri110/EPT-Benchmark.

In subjective assessment of image and video quality, observers rate or compare selected stimuli. Before calculating the mean opinion scores (MOS) for these stimuli from the ratings, it is recommended to identify and deal with outliers that may have given unreliable ratings. Several methods are available for this purpose, some of which have been standardized. These methods are typically based on statistics and sometimes tested by introducing synthetic ratings from artificial outliers, such as random clickers. However, a reliable and comprehensive approach is lacking for comparative performance analysis of outlier detection methods. To fill this gap, this work proposes and applies an empirical worst-case analysis as a general solution. Our method involves evolutionary optimization of an adversarial black-box attack on outlier detection algorithms, where the adversary maximizes the distortion of scale values with respect to ground truth. We apply our analysis to several hard and soft outlier detection methods for absolute category ratings and show their differing performance in this stress test. In addition, we propose two new outlier detection methods with low complexity and excellent worst-case performance. Software for adversarial attacks and data analysis is available.

This paper explores the multilingual commonsense generation abilities of Large Language Models (LLMs). To facilitate this investigation, we introduce MULTICOM, a novel benchmark that extends the COCOTEROS dataset to four languages: English, Spanish, Dutch, and Valencian. The task involves generating a commonsensical sentence that includes a given triplet of words. We evaluate a range of open-source LLMs, including LLaMA, Qwen, Gemma, EuroLLM, and Salamandra, on this benchmark. Our evaluation combines automatic metrics, LLM-as-a-judge approaches (using Prometheus and JudgeLM), and human annotations. Results consistently show superior performance in English, with significantly lower performance in less-resourced languages. While contextual support yields mixed results, it tends to benefit underrepresented languages. These findings underscore the current limitations of LLMs in multilingual commonsense generation. The dataset is publicly available at https://huggingface.co/datasets/gplsi/MULTICOM.

Questionnaire-based surveys are foundational to social science research and public policymaking, yet traditional survey methods remain costly, time-consuming, and often limited in scale. This paper explores a new paradigm: simulating virtual survey respondents using Large Language Models (LLMs). We introduce two novel simulation settings, namely Partial Attribute Simulation (PAS) and Full Attribute Simulation (FAS), to systematically evaluate the ability of LLMs to generate accurate and demographically coherent responses. In PAS, the model predicts missing attributes based on partial respondent profiles, whereas FAS involves generating complete synthetic datasets under both zero-context and context-enhanced conditions. We curate a comprehensive benchmark suite, LLM-S^3 (Large Language Model-based Sociodemographic Survey Simulation), that spans 11 real-world public datasets across four sociological domains. Our evaluation of multiple mainstream LLMs (GPT-3.5/4 Turbo, LLaMA 3.0/3.1-8B) reveals consistent trends in prediction performance, highlights failure modes, and demonstrates how context and prompt design impact simulation fidelity. This work establishes a rigorous foundation for LLM-driven survey simulations, offering scalable and cost-effective tools for sociological research and policy evaluation. Our code and dataset are available at: https://github.com/dart-lab-research/LLM-S-Cube-Benchmark

This study examined the use of machine learning and domain specific enrichment on patient generated health data, in the form of free text meal logs, to classify meals on alignment with different nutritional goals. We used a dataset of over 3000 meal records collected by 114 individuals from a diverse, low income community in a major US city using a mobile app. Registered dietitians provided expert judgement for meal to goal alignment, used as gold standard for evaluation. Using text embeddings, including TFIDF and BERT, and domain specific enrichment information, including ontologies, ingredient parsers, and macronutrient contents as inputs, we evaluated the performance of logistic regression and multilayer perceptron classifiers using accuracy, precision, recall, and F1 score against the gold standard and self assessment. Even without enrichment, ML outperformed self assessments of individuals who logged meals, and the best performing combination of ML classifier with enrichment achieved even higher accuracies. In general, ML classifiers with enrichment of Parsed Ingredients, Food Entities, and Macronutrients information performed well across multiple nutritional goals, but there was variability in the impact of enrichment and classification algorithm on accuracy of classification for different nutritional goals. In conclusion, ML can utilize unstructured free text meal logs and reliably classify whether meals align with specific nutritional goals, exceeding self assessments, especially when incorporating nutrition domain knowledge. Our findings highlight the potential of ML analysis of patient generated health data to support patient centered nutrition guidance in precision healthcare.

The rapid development of the low-altitude economy emphasizes the critical need for effective perception and intent recognition of non-cooperative unmanned aerial vehicles (UAVs). The advanced generative reasoning capabilities of multimodal large language models (MLLMs) present a promising approach in such tasks. In this paper, we focus on the combination of UAV intent recognition and the MLLMs. Specifically, we first present an MLLM-enabled UAV intent recognition architecture, where the multimodal perception system is utilized to obtain real-time payload and motion information of UAVs, generating structured input information, and MLLM outputs intent recognition results by incorporating environmental information, prior knowledge, and tactical preferences. Subsequently, we review the related work and demonstrate their progress within the proposed architecture. Then, a use case for low-altitude confrontation is conducted to demonstrate the feasibility of our architecture and offer valuable insights for practical system design. Finally, the future challenges are discussed, followed by corresponding strategic recommendations for further applications.

Multispectral and hyperspectral imagery are widely used in agriculture, environmental monitoring, and urban planning due to their complementary spatial and spectral characteristics. A fundamental trade-off persists: multispectral imagery offers high spatial but limited spectral resolution, while hyperspectral imagery provides rich spectra at lower spatial resolution. Prior hyperspectral generation approaches (e.g., pan-sharpening variants, matrix factorization, CNNs) often struggle to jointly preserve spatial detail and spectral fidelity. In response, we propose SpecSwin3D, a transformer-based model that generates hyperspectral imagery from multispectral inputs while preserving both spatial and spectral quality. Specifically, SpecSwin3D takes five multispectral bands as input and reconstructs 224 hyperspectral bands at the same spatial resolution. In addition, we observe that reconstruction errors grow for hyperspectral bands spectrally distant from the input bands. To address this, we introduce a cascade training strategy that progressively expands the spectral range to stabilize learning and improve fidelity. Moreover, we design an optimized band sequence that strategically repeats and orders the five selected multispectral bands to better capture pairwise relations within a 3D shifted-window transformer framework. Quantitatively, our model achieves a PSNR of 35.82 dB, SAM of 2.40°, and SSIM of 0.96, outperforming the baseline MHF-Net by +5.6 dB in PSNR and reducing ERGAS by more than half. Beyond reconstruction, we further demonstrate the practical value of SpecSwin3D on two downstream tasks, including land use classification and burnt area segmentation.

Dust storms harm health and reduce visibility; quick detection from satellites is needed. We present a near real-time system that flags dust at the pixel level using multi-band images from NASA's Terra and Aqua (MODIS). A 3D convolutional network learns patterns across all 36 bands, plus split thermal bands, to separate dust from clouds and surface features. Simple normalization and local filling handle missing data. An improved version raises training speed by 21x and supports fast processing of full scenes. On 17 independent MODIS scenes, the model reaches about 0.92 accuracy with a mean squared error of 0.014. Maps show strong agreement in plume cores, with most misses along edges. These results show that joint band-and-space learning can provide timely dust alerts at global scale; using wider input windows or attention-based models may further sharpen edges.

Early research into data poisoning attacks against Large Language Models (LLMs) demonstrated the ease with which backdoors could be injected. More recent LLMs add step-by-step reasoning, expanding the attack surface to include the intermediate chain-of-thought (CoT) and its inherent trait of decomposing problems into subproblems. Using these vectors for more stealthy poisoning, we introduce ``decomposed reasoning poison'', in which the attacker modifies only the reasoning path, leaving prompts and final answers clean, and splits the trigger across multiple, individually harmless components. Fascinatingly, while it remains possible to inject these decomposed poisons, reliably activating them to change final answers (rather than just the CoT) is surprisingly difficult. This difficulty arises because the models can often recover from backdoors that are activated within their thought processes. Ultimately, it appears that an emergent form of backdoor robustness is originating from the reasoning capabilities of these advanced LLMs, as well as from the architectural separation between reasoning and final answer generation.

The widespread deployment of LLMs across enterprise services has created a critical security blind spot. Organizations operate multiple LLM services handling billions of queries daily, yet regulatory compliance boundaries prevent these services from sharing threat intelligence about prompt injection attacks, the top security risk for LLMs. When an attack is detected in one service, the same threat may persist undetected in others for months, as privacy regulations prohibit sharing user prompts across compliance boundaries. We present BinaryShield, the first privacy-preserving threat intelligence system that enables secure sharing of attack fingerprints across compliance boundaries. BinaryShield transforms suspicious prompts through a unique pipeline combining PII redaction, semantic embedding, binary quantization, and randomized response mechanism to potentially generate non-invertible fingerprints that preserve attack patterns while providing privacy. Our evaluations demonstrate that BinaryShield achieves an F1-score of 0.94, significantly outperforming SimHash (0.77), the privacy-preserving baseline, while achieving 64x storage reduction and 38x faster similarity search compared to dense embeddings.