Accurate workload prediction is essential to ensure application Quality of Service (QoS), cost efficiency, and compliance with Service Level Agreements (SLAs) during cloud-based deployment. However, existing methods struggle to achieve accurate forecasts across multiple temporal scales and often fail to generalize well with limited historical data. To tackle these problems, we propose LLM4Load-Turbo, a Prompt-Driven LLM Framework with Knowledge Distillation for Efficient Multi-Scale Workload Prediction. Firstly, we design a structured prompt with dataset introduction, task description, and workload features characterization to extract multi-scale features. Secondly, we introduce a cross-modality alignment mechanism combined with label embedding to further enhance predictive accuracy and generalization, effectively mitigating the container cold-start problem. Thirdly, we propose a two-level knowledge distillation strategy, enabling LLM4Load-Turbo to maintain high accuracy while substantially reducing inference latency, memory footprint, and computational cost. Specifically, our framework achieves up to an 89.27% improvement in inference speed and a 93.30% reduction in model parameter scale compared to state-of-the-art baselines. Extensive experiments on four real cloud workload datasets validate the effectiveness of our framework. For multi-scale workload prediction, LLM4Load-Turbo improves up to 50.44%. For container cold-start scenarios, LLM4Load-Turbo improves up to 93.54%. These results demonstrate the potential of LLM4Load-Turbo to enable dynamic and efficient resource management in modern cloud systems.

As cardiovascular diseases and arrhythmias rise globally, pacemakers have become a critical therapeutic option for managing cardiac rhythm disorders. Accurate identification of pacemaker implantation sites is essential for personalized pacing therapy and optimal clinical outcomes. While 12-lead electrocardiogram (ECG) signals provide a non-invasive means to infer implantation locations, they are susceptible to noise and morphological variability, posing challenges for high-accuracy localization. To advance data-driven solutions in this domain, we present PILDE, the first publicly available dataset specifically designed for pacemaker implantation site identification, comprising 12-lead ECG recordings from 733 patients across four distinct implantation locations. Based on this dataset, we propose STELLAR, a novel deep learning framework that integrates a Spatio-Temporal Lead-Harmonic Mechanism to model both the temporal dynamics of ECG waveforms and the spatial coherence across leads. Extensive experiments demonstrate that STELLAR outperforms conventional deep models-including CNN, LSTM, and Transformer baselines-on both the PILDE and PTB-XL datasets. Specifically, STELLAR achieves an average accuracy improvement of 10.45 % on PILDE and 14.19 % on PTB-XL, with significant gains in sensitivity and F1-score for minority classes. These results highlight the robustness and precision of STELLAR in automating implantation site identification, offering a promising tool for pre-procedural planning and clinical decision support. The source code and dataset access information will be made publicly available.

As application migration to the cloud becomes the mainstream way of application deployment, accurate workload prediction is critical to ensure the quality of service (QoS) and cost-efficiency of the applications and meet service level agreements (SLAs) with users. Short-term workload prediction can handle workload fluctuations over a short duration, while long-term prediction can capture trend and periodic changes for workload. However, existing studies are unable to deal with long-term forecasts in multiple scales effectively; moreover, new containers lack historical data, leading to inaccurate prediction, while existing methods perform poorly due to a lack of generalization ability. To tackle these problems, we propose LLM4Load, an LLM Prompt-Driven Framework for Multi-Scale Workload Prediction. Firstly, we design a structured prompt with dataset introduction, task description, and workload features characterization to extract multi-scale features. Secondly, we introduce a cross-modality alignment mechanism combined with label embedding to further enhance performance. Leveraging the generalization capability of LLM, we also solve the container cold-start problem. The abundant experiments on four real cloud workload datasets validate the effectiveness of LLM4Load. For multi-scale workload prediction, LLM4Load improves up to 42.40%. For container cold-start scenarios, LLM4Load improves up to 93.72%. These results highlight the potential of LLM4Load to drive dynamic and efficient resource management in modern cloud systems.

Retrieving appropriate records from the external knowledge base to generate informative responses is the core capability of end-to-end task-oriented dialogue systems (EToDs). Most of the existing methods additionally train the retrieval model or use the memory network to retrieve the knowledge base, which decouples the knowledge retrieval task from the response generation task, making it difficult to jointly optimize and failing to capture the internal relationship between the two tasks. In this paper, we propose a simple and unified generative model for task-oriented dialogue systems, which recasts the EToDs task as a single sequence generation task and uses maximum likelihood training to train the two tasks in a unified manner. To prevent the generation of non-existent records, we design the prefix trie to constrain the model generation, which ensures consistency between the generated records and the existing records in the knowledge base. Experimental results on three public benchmark datasets demonstrate that our method achieves robust performance on generating system responses and outperforms the baseline systems. To facilitate future research in this area, the code is available at https://github.com/dzy1011/Uni-ToD.

Large language models (LLMs) perform well in many NLP tasks but frequently generate inaccurate information in the biomedical domain, due to hallucination issues. Retrieval-Augmented Generation (RAG) has been introduced to address this issue by integrating external knowledge, enhancing the factual accuracy of outputs. However, naive RAG encounters challenges in effectively utilizing retrieved content, particularly in specialized domains like biomedicine. LLMs often struggle to integrate retrieved content as irrelevant information can interfere with the model’s judgment. Even if relevant documents are retrieved, the model may be unable to accurately comprehend and utilize the domain-specific features due to its inherent knowledge limitations. To overcome these limitations, we propose Document Embeddings Enhanced Biomedical RAG (DEEB-RAG), a framework that incorporates document embeddings along with the original retrieved text. DEEB-RAG uses MedCPT to generate document embeddings and these embeddings are then aligned with the LLM’s semantic space using a two-stage training process on a simple projector. Experimental results on biomedical QA datasets show that DEEB-RAG improves accuracy, with an average performance increase of 2.3% over naive RAG. This demonstrates DEEB-RAG’s ability to mitigate the challenges of utilizing complex biomedical information, thereby enhancing the reliability and effectiveness of LLMs in biomedical domain.

As application migration to the cloud becomes the mainstream way of application deployment, application runtime management presents a significant need for large-scale workload prediction technology. Existing large-scale workload approaches commonly construct training sets based on all the training samples generated from the original data to generate prediction models. However, due to the similar behavior between different container instances of the microservice application, training and modeling in this way results in a huge number of redundant samples, which produces a significant redundant training overhead. Therefore, this paper proposes Less, a largescale workload forecasting model based on multiple sequence compression. First, based on the grouping results of similar containers, a container workload feature recognition algorithm is proposed to determine the common and individual features of container workloads in each prediction period, so as to guide the compression of workload sequences within each group; second, a fitness function that takes into account the common features, individual features, and the number of sequences are designed, and the optimal compressed sequences are solved by the Whale Optimization Algorithm to efficiently reduce the number of redundant training workload sequences, and then the Bidirectional Gated Recurrent Unit model is built and trained based on the compressed sequences, which effectively reduces the model complexity and overhead while ensuring the accuracy. Finally, we validate the comprehensive advantages of Less in terms of accuracy and overhead based on public datasets and verify the effectiveness of each subpart of our model through ablation experiments.

OBJECTIVE: Most existing fine-tuned biomedical large language models (LLMs) focus on enhancing performance in monolingual biomedical question answering and conversation tasks. To investigate the effectiveness of the fine-tuned LLMs on diverse biomedical natural language processing (NLP) tasks in different languages, we present Taiyi, a bilingual fine-tuned LLM for diverse biomedical NLP tasks. MATERIALS AND METHODS: We first curated a comprehensive collection of 140 existing biomedical text mining datasets (102 English and 38 Chinese datasets) across over 10 task types. Subsequently, these corpora were converted to the instruction data used to fine-tune the general LLM. During the supervised fine-tuning phase, a 2-stage strategy is proposed to optimize the model performance across various tasks. RESULTS: Experimental results on 13 test sets, which include named entity recognition, relation extraction, text classification, and question answering tasks, demonstrate that Taiyi achieves superior performance compared to general LLMs. The case study involving additional biomedical NLP tasks further shows Taiyi's considerable potential for bilingual biomedical multitasking. CONCLUSION: Leveraging rich high-quality biomedical corpora and developing effective fine-tuning strategies can significantly improve the performance of LLMs within the biomedical domain. Taiyi shows the bilingual multitasking capability through supervised fine-tuning. However, those tasks such as information extraction that are not generation tasks in nature remain challenging for LLM-based generative approaches, and they still underperform the conventional discriminative approaches using smaller language models.

Automatic extraction of entities and their relations from unstructured literature to form structured triples is essential for biomedical knowledge construction. Although most existing joint methods have effectively addressed some challenging problems in the biomedical corpora, i.e., the prevalent overlapping issue, they still suffer from a lack of consideration for the intrinsic correlations between entities and relations, as well as low computational efficiency. In this paper, we present a joint entity and relation extraction model with unified interaction maps. Specifically, we concatenate all relations in the natural language form with the input text to integrate the semantic information of relations through a deep Transformer-based encoder. In addition, we apply unified interaction maps to capture the correlations, which can naturally handle the overlapping issue. Extensive experiments on the CHEMPROT and DDIExtraction2013 datasets demonstrate the effectiveness of our model, achieving the state-of-the-art performance with higher efficiency.

As application migration to the cloud becomes the mainstream way of application deployment, application runtime management presents a significant need for large-scale workload prediction technology. However, existing large-scale workload forecasting models focus more on improving model accuracy and ignore the models’ storage, training time, and testing time, which leads to colossal overhead. Therefore, this paper proposes a large-scale workload forecasting model for containers. First, based on the workload value features and waveform features, a feature-enhanced workload similarity calculation algorithm is proposed to determine the grouping of containers with similar workload patterns in real time by analyzing the historical similarity and recent similarity of workloads among different containers; second, we employ Transformer as the base model to design position encoding and attention mask based on the real-time workload similarity relationship and achieve forecasting model parallelized training based on the multi-head self-attention mechanism, which balances the workload prediction accuracy and model overhead. Finally, we will validate the comprehensive advantages of our model in terms of accuracy and overhead based on public datasets and verify the effectiveness of each subpart of our model through ablation experiments.

Intent detection and slot filling are two main tasks for building a spoken language understanding (SLU) system. Since the two tasks are closely related, the joint models for the two tasks always outperform the pipeline models in SLU. However, most joint models directly incorporate multiple intent information for each token, which introduces intent noise into the sentence semantics, causing a decrease in the performance of the joint model. In this paper, we propose a Dynamic Graph Model (DGM) for joint multiple intent detection and slot filling, in which we adopt a sentence-level intent-slot interactive graph to model the correlation between the intents and slot. Besides, we design a novel method of constructing the graph, which can dynamically update the interactive graph and further alleviate the error propagation. Experimental results on several multi-intent and single-intent datasets show that our model not only achieves the state-of-the-art (SOTA) performance but also boosts the speed by three to six times over the SOTA model.