LLM based question answering systems have been used in healthcare, education,and customer support, and in these settings user prompts often contain names, health conditions, contact details, or other sensitive clues, which makes privacy protection at inferencetime hard to avoid, especially when the service is deployed as a black box. Many existingdefenses still rely on retraining or relatively rigid filtering rules, so they do not adapt wellwhen contextual sensitivity changes from one interaction to another. This paper proposes aprivacy protection framework for LLM QA at the question answer pair level. The frameworkestimates exposure risk with a probabilistic graphical model, then adjusts mitigation strengththrough threshold based control. In the protection stage, sanitization, abstraction, and calibrated noise are used together, so stronger intervention can be applied to more sensitive inputswhile ordinary interactions are affected less. Experiments on a SafetyBench derived datasetshow that the framework can reduce privacy risk and still keep answer utility at a useful level,with relatively low computational overhead.

Identifying stress states from user text is important for mental health monitoring, but it remains difficult because stress expression is subjective and varied, and language cues across states overlap. In closed-set classification, the model must select a label from a fixed set of labels, as vague or weak evidence often leads to errors. While instruction follows LLMs to accomplish this task, the output can be word-sensitive and use unconstrained reasoning when the evidence is insufficient. We propose StressAgent, an evidence-enhanced dual-agent framework, to improve robustness. StressAgent breaks down closed-set predictions into two steps: a Reasoning and Decision (RD) agent and a Retrieval-Augmented Generation (RAG) agent. The RAG agent retrieves the 10 most similar training instances from the search index and returns their category labels and short text fragments associated with the tags as evidence. The RD agent then integrates the retrieved evidence into a restricted closed-set output format to generate a final label. Experiments were conducted on Qwen2.5-7B-Instruct and Deepseek-Chat to evaluate all combinations of enabled or disabled retrieval and enabled or disabled chains of thought. The results show that retrieval enhancement provides continuous accuracy improvements and explains most improvements. Search using cosonic similarity outperformed Euclidean distance, highlighting the impact of similarity measures on the usefulness of evidence. In the absence of chain of thought, cosine similarity retrieval improved the accuracy of Qwen2.5-7B-Instruct from 61.97% to 71.37%, and Deepseek-Chat from 67.52% to 74.89%. Once retrieval is enabled, the further benefits of chains of thought are limited. Overall, StressAgent provides an interpretable, controllable, evidence-based driven classification paradigm for closed set stresses.

Existing erasing-based data augmentation methods suffer from two key limitations:mask generation is decoupled from image semantics, resulting in random and blind erasing locations; and a uniform erasing intensity is applied across all categories, which exacerbates training signal imbalance under long-tail distributions. To address these issues, this paper proposes the Semantic-aware Category-Adaptive Erasing augmentation method (SCAE): a Grad-CAM-based saliency estimation module is introduced to guide erasing locations via semantic activation distributions, combined with a curriculum learning strategy for progressive hard sample generation; meanwhile, erasing probabilities are adaptively adjusted according to class frequency, and mixup augmentation is incorporated for extreme tail classes. Experiments show that SCAE outperforms existing mainstream methods on CIFAR-100,

In most games, the Non-Player Characters (NPCs) are designed mainly using a rule-based approach for controllability. For achieving realistic NPC behaviors, usually, a lot of manual fine-tuning is needed for different player types and scenarios. However, machine learning can be adopted to enable NPCs to learn and adjust their behavior over time using data and interaction feedback. Although there are related studies using machine learning to improve NPC behaviors for games, they are scattered across different technical platforms, tasks, evaluation methods and metrics. Therefore, it is impossible to make any cross-study comparisons and draw conclusions. This paper introduces a review framework named PMM (Platform–Method–Metric) which is a method-based classification approach. Using this framework, this paper systematically reviews the existing research and also constructs a unified system for measuring player experience. The future work of this research will contribute to the goal of unifying and automating cross-platform evaluation standards.

In integrated navigation systems, information fusion and positioning accuracy depend on the characteristics of inertial systems and sensors, yet obtaining prior knowledge remains challenging in practice. To address issues of varying satellite signal quality and system nonlinearities degrading navigation performance, this paper proposes a Fuzzy Adaptive Interacting Multiple Model with Multiple Kalman Filters (FAIMM-MKF) algorithm. It combines a Fuzzy Controller based on satellite signal quality with an Adaptive Interacting Multiple Model (AIMM), integrating Unscented Kalman Filter (UKF), Iterated Extended Kalman Filter (IEKF), and Square-Root Cubature Kalman Filter (SRCKF) to match vehicle dynamics models. The performance is validated through hardware-in-the-loop experiments. Results show that compared to traditional IMM algorithms, this method significantly improves positioning accuracy in complex environments when satellite signal quality changes.

Facial expression recognition (FER), as the core technology in the cross field of emotional computing and computer vision, has important theoretical significance and practical application value in the real environment. It not only promotes the theoretical deepening of emotional computing, human-computer interaction and other related disciplines, but also provides emotional perception support for multi domain scenes, and helps to realize a more intelligent and humanistic interaction mode. Firstly, this paper briefly describes the research significance and application value of facial expression recognition in real environment; Then, the evolution of FER technology is sorted out, and the technical iteration from traditional convolutional neural network (CNN) to vision transformer (VIT) is analyzed. The unique advantages of the attention mechanism supported by ViT in capturing global features and modeling long-distance dependence are clarified, as well as its specific mechanism in dealing with complex challenges such as light change, face occlusion, pose shift and so on in the real environment; Then, it systematically summarizes the core challenges faced by FER research in the current real environment, and sorts out the frontier research directions and development trends such as multimodal fusion; Finally, the key problems that have not been solved in this field are pointed out, and the future technology development trend is prospected.

With the advancement of AI, empathetic dialogue systems have made great progress, recognizing user feelings and providing responses that are suitable in different situations, which are becoming increasingly valuable. Nevertheless, there are still difficulties in specific areas, including mental health services and smart customer care. The main current limitations are the insufficient depth of emotional understanding in traditional AI, the inability to precisely detect the differences in the intensity of emotions, the inability to dynamically balance emotional responses to contextual semantics, and the inability to sustain emotional resonance in long-term dialogue. Moreover, the current approaches tend to use a single, fixed-weight reward system, which fails to meet the dynamics of the dialogue context. In the meantime, the vast majority of the models use single-turn responses as their optimization goal and ignore the role of long-term dependencies in the dialogue history in determining empathetic effectiveness. To address these challenges, this project suggests a machine learning-based empathetic dialogue generation framework that uses reinforcement learning. Using the "reward-penalty" mechanism and dynamic adjustment of weights to various goals, the system constructs a variety of reward signals to steer models towards sustained dialogue interactions instead of single-response patterns, adapting to various conversational scenarios and emotional variations.

Dancing With The Stars have some problems in balancing fan popularity and technical merit, so we do several tasks to solve it. First, we estimate unknown fan votes with a Sequential Bayesian Inference Model, check its consistency and certainty. Second, we compare three voting rules with a Counterfactual Simulation Framework, find their differences and suggest a better one. Third, we build a model to analyze factors that influence Judge Scores and Fan Votes, see how they impact the two. Fourth, we propose a fairer elimination system called DBEP, test its effect. Finally, we do a sensitivity test on the first task's hyperparameter k to prove the model is robust.

Quantitative finance is advancing rapidly, making the effective use of massive financial data essential for stock return prediction. Traditional prediction methods cannot easily integrate numerical market data with unstructured textual information,which leads to lower predictive accuracy. To address these issues, this paper proposes a stock return prediction model. This model uses a fine-tuned BERT and a multimodal feature fusion method. First, we use an end-to-end fine-tuning technique. We add the BERT model as a trainable module inside a single prediction framework. With a backpropagation mechanism, the BERT parameters and the prediction model are trained together. Then the text feature extraction process is tightly bound to the final return prediction goal. In this way, we solve the goal mismatch problem from old two-stage methods.Second, we design a cross-attention mechanism for feature fusion. This mechanism fixes the separation between text features and number features. Then the two feature types can interact and fuse inside a shared space. We also use rolling window methods for dynamic training and testing. We run comparative experiments with both simulated data and real market data. The results show that our model works better than single-mode benchmark models, as key metrics like mean squared error and mean absolute error all show clear improvement.

The accuracy and efficiency of traffic sign recognition have to be very high to facilitate autonomous driving systems. AE-YOLO is a YOLOv8-based model that is combined with an adaptive early-stopping scheme to recognize traffic signs, which is proposed in this paper. The developed approach decreases redundant training at the cost of competitive recognition performance. Also, systematic hyperparameter optimization is performed to evaluate how model size, input size and stopping parameters affect the trade-off between accuracy and computational time. In order to test the robustness even further, one more set of road sign images acquired through online databases is added to the dataset and the same evaluation procedure is used. Experimental evidence shows that AE-YOLO attains a better trade-off between precision and speed with high level of robustness in various road sign settings. Final accuracy of AE-YOLO is 2.3 percentage points better than the one of the most well-known baselines (four factor groups). The code may be found at: https://github.com/chenjunzhu4-ctrl/Adaptive-Early-Stopping-YOLO-for-Road-Sign-Recognition