Large language models have demonstrated impressive performance on natural language tasks, but their decision-making processes remain opaque. Existing explanation methods either suffer from limited faithfulness to the model's reasoning or produce explanations that are difficult for humans to understand. To address these challenges, we propose ProtoSurE, a novel prototype-based surrogate framework that provides faithful and understandable explanations for LLMs. ProtoSurE trains an interpretable-by-design surrogate model that aligns with the target LLM while utilizing sentence-level prototypes as understandable concepts. Extensive experiments show that ProtoSurE consistently outperforms state-of-the-art explanation methods across diverse LLMs and datasets. Importantly, ProtoSurE demonstrates strong data efficiency, requiring relatively few training examples to achieve good performance, making it practical for real-world applications.

Large vision-language models (LVLMs) achieve impressive performance on multimodal tasks but often suffer from hallucination and confidently describe objects or attributes not present in the image. Current training-free interventions struggle to maintain accuracy in open-ended and long-form generation scenarios. We introduce the Confidence-Aware Attention Calibration (CAAC) framework to address this challenge by targeting two key biases: spatial perception bias, which distributes attention disproportionately across image tokens, and modality bias, which shifts focus from visual to textual inputs over time. CAAC employs a two-step approach: Visual-Token Calibration (VTC) to balance attention across visual tokens, and Adaptive Attention Re-Scaling (AAR) to reinforce visual grounding guided by the model’s confidence. This confidence-driven adjustment ensures consistent visual alignment during generation. Experiments on CHAIR, AMBER, and POPE benchmarks demonstrate that CAAC outperforms baselines, particularly in long-form generations, effectively reducing hallucination. Data and code are available at https://github.com/mehrdadfazli/CAAC/.

The outburst of COVID-19 in late 2019 was the start of a health crisis that shook the world and took millions of lives in the ensuing years. Many governments and health officials failed to arrest the rapid circulation of infection in their communities. The long incubation period and the large proportion of asymptomatic cases made COVID-19 particularly elusive to track. However, wastewater surveillance soon became a promising data source in addition to conventional indicators such as confirmed daily cases, hospitalizations, and deaths. Despite the consensus on the effectiveness of wastewater viral load, there is a lack of methodological approaches that leverage viral load to improve COVID-19 forecasting. This paper proposes a deep learning framework to automatically discover the relationship between daily cases and viral load data. We trained a Deep Temporal Convolutional Network (DeepTCN) and a Temporal Fusion Transformer (TFT) model to obtain a global forecasting model. We supplement the daily confirmed cases with viral loads and other socio-economic factors as covariates to the models. Our results suggest that TFT outperforms DeepTCN and learns a better association between viral load and daily cases. We demonstrate that equipping the models with the viral load improves forecasting accuracy and reduces uncertainty. Moreover, viral load is shown to be the second most predictive input, following the containment and health index. Our results reveal the feasibility of training a location-agnostic deep-learning model to capture the dynamics of infection diffusion when wastewater viral load data is available.

Wearable smart devices are capable of capturing a variety of information from their users using a multitude of noninvasive sensing modalities. Using features from the raw measurements of wearable devices, sensor fusion enables us to obtain a holistic picture of the users’ context and monitor their activity state with increased accuracy. Human activity recognition using noninvasive sensors allows us to capture the natural behavior of users in their day-to-day lives. This in-the-wild activity recognition, however, poses several key challenges that must be addressed to create effective classification models. The main challenges are class imbalance, uncertainty in classifier decisions, and large feature spaces. To address them, this study further explores a probabilistic sensor fusion method called Naive Adaptive Probabilistic Sensor (NAPS) Fusion. In doing so, we establish the viability of NAPS Fusion for natural human activity recognition using noninvasive sensing modalities. NAPS Fusion handles dimensionality reduction by creating reduced feature sets and mitigates the class imbalance issue through the use of Synthetic Minority Oversampling Technique (SMOTE). Moreover, NAPS Fusion addresses uncertainty in the decisions of classifiers using a Dempster-Shafer theoretic late fusion framework. Our empirical evaluation demonstrates that NAPS Fusion has broad applications beyond its original design for cognitive state detection. It outperforms similar decision level sensor fusion methods (late fusion using averaging, LFA, and late fusion using learned weights, LFL) in the detection of exercise and sedentary activities such as walking, running, lying down, and sitting. We observe improvements of up to 56% in F1 score and up to 59% in precision with NAPS Fusion over the compared methods.

Using wastewater surveillance as a continuous pooled sampling technique has been in place in many countries since the early stages of the outbreak of COVID-19. Since the beginning of the outbreak, many research works have emerged, studying different aspects of viral SARS-CoV-2 DNA concentrations (viral load) in wastewater and its potential as an early warning method. However, one of the questions that has remained unanswered is the quantitative relation between viral load and clinical indicators such as daily cases, deaths, and hospitalizations. Few studies have tried to couple viral load data with an epidemiological model to relate the number of infections in the community to the viral burden. This paper proposes a stochastic wastewater-based SEIR model to showcase the importance of viral load in the early detection and prediction of an outbreak in a community. We built three models based on whether or not they use the case count and viral load data and compared their simulations and forecasting quality. Our results demonstrate that a simple SEIR model based on viral load data can reliably predict the number of infections in the future. Therefore, wastewater-based surveillance is a promising way of monitoring the spread of COVID19 and can provide city officials with timely information about the circulation of COVID-19 in the community.