Container orchestration platforms have become the backbone of modern private cloud infrastructure, offering flexibility, scalability, and reliability for managing containerized workloads. Among them, Kubernetes has emerged as the de facto standard, widely adopted across the industry, with many organizations operating tens or even hundreds of clusters globally. However, as infrastructure scales to ultra-large clusters (exceeding 10,000 nodes) and faces extreme provisioning scenarios—such as abrupt workload surges, irrespective of available resource headroom—scheduling latency becomes a critical bottleneck. This issue is not adequately addressed by the default Kubernetes scheduler or existing alternative solutions, which struggle to maintain low latency under such conditions.

The event extraction problem involves detecting event trigger words and extracting their corresponding event arguments. In contrast to general event extraction, financial event extraction focuses on financial texts, primarily at the document level. Existing methods mainly rely on a uniform sequence labeling model to identify event triggers and event arguments. However, this approach struggles with identifying nested and long entities and faces difficulties in handling complex event arguments. To address these challenges, we propose MRCFEE, a Chinese financial event extraction model based on machine reading comprehension (MRC). In the multi-turn MRC, the input to the model incorporates external knowledge and historical answers based on question templates and previous extraction results. Subsequently, we utilize a large-scale pre-trained language model in the financial field as the embedding layer and employ BiGRU to extract contextual information further. Then, we employ two binary classifiers to identify the probabilities of the starting and ending positions. Finally, we utilize the dynamic thresholding method to assess the rationality of the boundary of the obtained results. Experimental results on the DuEE-Fin dataset demonstrate that our model outperforms the previous methods, obtaining 84.1% and 75.6% F1 values for event detection and event argument extraction, respectively.

Time series forecasting, with its wide range of practical applications such as power load and weather prediction, has become a pivotal field of research. Over the past few years, neural network models have made remarkable progress in this domain. Many time series forecasting models now employ sequence decomposition techniques to enhance forecasting accuracy, including Autoformer, DLinear, and MICN. These techniques break down the original time series data into two components: trend and seasonal term, to facilitate more accurate predictions. However, a significant limitation of existing models that utilize sequence decomposition is their incomplete exploitation of the trend component. To address this issue, we introduce RDLinear, a structurally simple model designed to fully leverage the unique attributes of sequence decomposition. RDLinear employs distinct forecasting strategies, with a primary focus on utilizing the RevIN method to predict the trend component. In this paper, we present extensive experimental results on multiple real-world datasets. Our findings demonstrate that RDLinear outperforms other time-series forecasting models, particularly in long-term forecasting. Furthermore, ablation experiments confirm the effectiveness of our proposed method.

Document-level relation extraction (DRE) aims to extract relations between entities in unstructured documents. Unlike sentence-level relation extraction, DRE introduces complexities associated with entities that appear in multiple sentences with different mentions, and where the head and tail entities of a given relation triple can be situated in diverse sentences. Consequently, the aggregation of semantic information for entity pairs is a paramount challenge in DRE. To address this challenge, we introduce a novel framework, Entity and Evidence Guided Attention (EEGA). This framework employs a pre-trained language model as an encoder, crafting rich contextual representations for entity pairs through the integration of both relation extraction and evidence retrieval. Initially, we guide the model’s attention towards the contextual information surrounding an entity pair using evidence. We then introduce an attention mechanism that assigns weights to words, guiding the extraction of semantic information at different levels: sentence, document, and evidence. An adaptive fusion module dynamically amalgamates the semantic information of the entity pairs at various granularities to obtain context-aware entity pair representations with rich semantics. Additionally, we propose self-training with relation labels and evidence attention on distantly supervised data to enhance DocRE performance. Experimental results on a benchmark dataset demonstrate the superiority of EEGA over strong baselines.

Over the last few years, at ByteDance, our compute infrastructure scale has been expanding significantly due to expedited business growth. In this journey, to meet hyper-scale growth, some business groups resorted to managing their own compute infrastructure stack running different scheduling systems such as Kubernetes, YARN which created two major pain points: the increasing resource fragmentation across different business groups and the inadequate resource elasticity between workloads of different business priorities. Isolation across different business groups (and their compute infrastructure management) leads to inefficient compute resource utilization and prevents us from serving the business growth needs in the long run.

Emotion-Cause Pair Extraction (ECPE) aims to extract potential emotion-cause pairs from text without emotion labels. It lays an important foundation for downstream research such as causal reasoning, public opinion prediction, and reason detection. However, the ECPE task now faces two dilemmas: 1) insufficient utilization of word sequence information, and 2) inadequate use of position information between clauses. To address the above problems, we proposed an Emotion-Cause Pair Extraction model based on Fusion Word Vectors named FW-ECPE. It is a two-stage model that first extracts emotion clauses and cause clauses respectively then combines them into pairs and filters out the right emotion-cause pairs. The Fusion Word Vector is reflected in two aspects. Firstly, we integrate the clause context vectors and the emotion clauses prediction results with cause context vectors in cause clauses extraction. Secondly, in the emotion-cause pair extraction stage, we fuse the position information between clauses and contextual information. Finally, we extend Easy Data Augmentation, a corpus enhancement algorithm, to enlarge the amount of data and alleviate the risk of overfitting. The experiment results show that our proposed approach outperforms the previous methods on a benchmark dataset.

Image dehazing is a crucial computer vision application with the primary objective of estimating haze-free images from hazy images. Deep neural network architectures have emerged as the dominant approaches and achieved remarkable progress. However, due to the intricacy, existing dehazing methods need help to train large deep learning networks. This work proposes a novel image dehazing network based on Diffusion Model (DehazeDM). Firstly, by segmenting the image into patches during the sampling procedure, we can dehaze images of arbitrary size. Then we compress the image into the latent space via the auto-encoder model and conduct the diffusion operation in the latent space, significantly decreasing the computational complexity associated with the task while exhibiting negligible effects on the perceptual fidelity of the resultant images. Extensive experiments verify the effectiveness and the superior performance of DehazeDM in image dehazing.

Teaching machines to understand human language is one of the most elusive challenges in artificial intelligence. Machine reading comprehension is a crucial task in evaluating how computer systems understand natural language. This study presents a machine reading comprehension model based on external knowledge. We use a framework named K-Adapter to infuse two kinds of external knowledge with two specific adapters. This model can capture richer semantic information, which is more suitable for real application scenarios. The proposed model is evaluated on the COSMOS QA dataset and outperforms the competitive baselines.

LimbMotion is a limb tracking system which enables accurate and real-time tracking with one wearable device on the wrist/ankle of a user. By integrating inertial sensing and acoustic sensing, LimbMotion significantly reduces the search space of a moving limb, and provides accurate limb tracking for further human computer interaction (HCI). Objectives of this demo are to show how LimbMotion works and two HCI applications supported by LimbMotion.