Agentic workflows carry out complex tasks by orchestrating multiple large language models (LLMs) and tools. Serving such workflows at a target throughput with low latency is challenging because they can be defined using arbitrary agentic frameworks and exhibit unpredictable execution times: execution may branch, fan-out, or recur in data-dependent ways. Since LLMs in workflows often outnumber available GPUs, their execution also leads to GPU oversubscription. We describe Scepsy, a new agentic serving system that efficiently schedules arbitrary multi-LLM agentic workflows onto a GPU cluster. Scepsy exploits the insight that, while agentic workflows have unpredictable end-to-end latencies, the shares of each LLM's total execution times are comparatively stable across executions. Scepsy decides on GPU allocations based on these aggregate shares: first, it profiles the LLMs under different parallelism degrees. It then uses these statistics to construct an Aggregate LLM Pipeline, which is a lightweight latency/throughput predictor for allocations. To find a GPU allocation that minimizes latency while achieving a target throughput, Scepsy uses the Aggregate LLM Pipeline to explore a search space over fractional GPU shares, tensor parallelism degrees, and replica counts. It uses a hierarchical heuristic to place the best allocation onto the GPU cluster, minimizing fragmentation, while respecting network topology constraints. Our evaluation on realistic agentic workflows shows that Scepsy achieves up to 2.4x higher throughput and 27x lower latency compared to systems that optimize LLMs independently or rely on user-specified allocations.

Multi-group activity recognition aims to recognize sub-group activities in multi-person scenes. Existing works explore group-level features by simply using graph neural networks for reasoning about the individual interactions and directly aggregating individual features, which cannot fully mine the interactions between people and between sub-groups, resulting in the loss of useful information for group activity recognition. To address this problem, this paper proposes a Spatio-Temporal Interactive Reasoning Model (STIRM) to better exploit potential spatio-temporal interactions for multi-group activity recognition. In particular, we present an interactive feature extraction strategy to explore correlation features between individuals by analyzing the features of their nearest neighbor. We design a new clustering module that combines the action similarity feature and spatio-temporal trajectory feature to divide people into small groups. In addition, to obtain rich and accurate group-level features, a group interaction reasoning module is constructed to explore the interactions between different small groups and among people in the same group and exclude people who have less impact on group activities according to their importance. Extensive experiments on the Social-CAD, PLPS and JRDB-PAR datasets indicate the superiority of the proposed method over state-of-the-art methods.

Parallel applications are typically implemented using multithreading (with shared memory, e.g., OpenMP) or multiprocessing (with message passing, e.g., MPI). While it seems attractive to deploy such applications in cloud virtual machines (VMs), existing cloud schedulers fail to manage such applications efficiently: they cannot scale multi-threaded applications dynamically when more CPU cores in a VM become available, and they cause fragmentation over time due to the static allocation of multi-process applications to VMs. We describe GRANNY, a new distributed runtime that enables the fine-granular management of multi-threaded/process applications in cloud environments. GRANNY supports the vertical scaling of multi-threaded applications within a VM and the horizontal migration of multi-process applications between VMs. GRANNY achieves both through a single WebAssembly-based execution abstraction: Granules can execute application code with thread or process semantics and allow for efficient snapshotting. GRANNY scales up applications by adding more Granules, and de-fragments applications by migrating Granules between VMs. In both cases, it launches new Granules from snapshots efficiently. We evaluate GRANNY with dynamic scheduling policies and show that, compared to current schedulers, it reduces the makespan for OpenMP workloads by up to 60% and the fragmentation for MPI workloads by up to 25%.

Large language models (LLMs) have shown remarkable performance in vision-language tasks, but their application in the medical field remains underexplored, particularly for integrating structured time series data with unstructured clinical notes. In clinical practice, dynamic time series data, such as lab test results, capture critical temporal patterns, while clinical notes provide rich semantic context. Merging these modalities is challenging due to the inherent differences between continuous signals and discrete text. To bridge this gap, we introduce ProMedTS, a novel self-supervised multimodal framework that employs prompt-guided learning to unify these heterogeneous data types. Our approach leverages lightweight anomaly detection to generate anomaly captions that serve as prompts, guiding the encoding of raw time series data into informative prompt embeddings. These prompt embeddings are aligned with textual representations in a shared latent space, preserving fine-grained temporal nuances alongside semantic insights. Furthermore, our framework incorporates tailored self-supervised objectives to enhance both intra- and inter-modal alignment. We evaluate ProMedTS on disease diagnosis tasks using real-world datasets, and the results demonstrate that our method consistently outperforms state-of-the-art approaches.

Remote sensing scene (RSS) image classification plays a vital role in various fields such as urban planning and environmental protection. However, due to higher inter-class similarity and intra-class variability, achieving accurate classification for RSS images poses a considerable challenge for current convolutional neural networks (CNNs)-based and visual transformer (ViT)-based methods. To address these issues, this paper proposes a novel dual-encoding method named master–slave encoding network (MSE-Net) from two perspectives of feature extraction and fusion. The master encoder, based on ViT, extracts higher-level semantic features, while the slave encoder, based on CNN, captures relative lower-level spatial structure information. Secondly, to integrate feature information from the two encoders effectively, this paper further develop two fusion strategies. The first strategy involves the auxiliary enhancement units (AEUs), which eliminates semantic divergence between the two encoders, enhances spatial context awareness of the slave encoder and promotes effective feature learning. The interactive perception unit (IPU), as the second strategy, facilitates interaction and integration of the two encoders’ representations to extract more discriminative feature information. In addition, we conducted comparative experiments on four widely-used RSS datasets, including RSSCN7, SIRI-WHU, the aerial image dataset (AID) and NWPU-RESISC45 (NWPU45), to verify the effectiveness of MSE-Net. The experimental results demonstrate that MSE-Net achieved state-of-the-art (SOTA) performance across all the datasets.

Group activity recognition aims to recognize holistic activity in multi-person scene, necessitating consideration of the interactions between actors and their surroundings. It has various applications, such as public surveillance, video analysis. Nonetheless, existing works merely extract scene features as a supplementary component of activity features, failing to adequately explore the interplay between the scene and actors. To address this limitation, this paper proposes a Local–Global Context-Aware Graph Reasoning Model (LG-CAGR), which leverages and reasons through local and global context features to gain deeper insights into group activity within the scene. In particular, we present an innovative feature extraction strategy to harness local location features and global scene attributes, effectively complementing actor features by capturing spatial group topology and determining relative positions between actors. Subsequently, we delve into these features by devising a local–global group reasoning module that deduces pair-wise interactions between actors and scenes within Graph Convolutional Network, comprehensively elucidating correlations between overall scene and local individuals to construct group-level features. Multi-graphs are constructed considering actor’s features, scene features as nodes, and interactions as edges. A self-attention graph pooling network is introduced to automatically integrate key actor features and form rich group-level features to recognize group activity. The results on Collective Activity Dataset, Collective Activity Extended Dataset, Volleyball Dataset and Public Life in Public Space dataset have reached 94.0%, 97.7%, 92.7% and 56.1%. Compared with existing methods using the same backbone, we exceeded 1%, 2.1%, 0.3%, and 14.9% respectively, affirming the superiority of the proposed method compared with state-of-the-art methods.

Deep learning (DL) jobs use multi-dimensional parallelism, i.e., combining data, model, and pipeline parallelism, to use large GPU clusters efficiently. Long-running jobs may experience changes to their GPU allocation: (i) resource elasticity during training adds or removes GPUs; (ii) hardware maintenance may require redeployment on different GPUs; and (iii) GPU failures force jobs to run with fewer devices. Current DL frameworks tie jobs to a set of GPUs and thus lack support for these scenarios. In particular, they cannot change the multi-dimensional parallelism of an already-running job in an efficient and model-independent way. We describe TENPLEX, a state management library for DL systems that enables jobs to change their parallelism dynamically after the GPU allocation is updated at runtime. TENPLEX achieves this through a new abstraction, a parallelizable tensor collection (PTC), that externalizes the job state during training. After a GPU change, TENPLEX uses the PTC to transform the job state: the PTC repartitions the dataset state under data parallelism and exposes it to GPU workers through a virtual file system; and the PTC obtains the model state as partitioned checkpoints and transforms them to reflect the new parallelization configuration. For efficiency, TENPLEX executes PTC transformations in parallel with minimum data movement between GPU workers. Our experiments show that TENPLEX enables DL jobs to support dynamic parallelization with low overhead.

This report aims to illustrate a highly effective method to obtain a kind of graphene oxide (GO)/reduced graphene oxide (RGO) paper by Ultraviolet (UV) laser direct writing method. Due to removing the oxygen-containing groups and introducing the layered micro-nano structures on one side of the GO film. The as-prepared GO/RGO papers exhibit anisotropically deformation triggered by ethanol vapor and could enable the potential applications for stimuli-response actuators.

Social relationships (SRs) are the basis of human life. Hence, the ability to accurately recognize interpersonal relations in public spaces based on visual observations helps policymakers improve mental health programs and address social challenges. The key to image-based computer-vision research on SR recognition (SRR) is a deep-learning mechanism that can predict SRs based on the contents of visual scenery images. Current methods explore logical constraints using relatively simple scenes with small groups of people. However, this is insufficient when desiring to form relation graphs of multiple groups simultaneously from complex scenes. Generally, complex scenes contain a principal relationship that applies to the largest proportion of people, and secondary relationships apply to smaller proportions. To effectively explore relational situations in complex scenes, we propose a new distributed reasoning strategy that accounts for principal and secondary SRs. First, our novel model enhances principal relation component reasoning, and a new contrastive learning algorithm supplements the principal relationship with secondary types. A shifted-window transformer is applied to extract interactive human relation features and local-global features to support more accurate and comprehensive relation prediction. Extensive experiments demonstrate that each part of the proposed model improves the accuracy of SRR and that the whole model outperforms state-of-the-art methods on public datasets.

Human perception refers to people's psychological feelings about a place. Understanding residents' perception and their activities is important for promoting people-oriented urban construction. Recently, with the development of machine learning, researchers used this technology to study human perception from the open-source street view images. However, the perception measurement is limited, caused by the inadequate feature extraction. Besides, human perceptions and their activities are separately studied, which hinders the process of revealing the relationship between human and environment. Hence, a human perception model was firstly proposed, where a Transformer network was introduced to extract more discriminative semantic features and visual elements were integrated to enhance the feature representations. Experiments showed that the average deviation of perceptual scores was controlled within 1.6 points, and its performance was improved by around 1%–18% in mean square error (MSE), root mean square error (RMSE) and mean absolute error (MAE) compared with the existing best results. Secondly, the collaborative study of environmental perception and residents' activities was carried out in the case area. Specifically, the perceptual measures of environment were implemented based on the street view video data. Meanwhile, the activities of residents were recognized by SlowFast network and quantified by a new informatics-based diversity indicator (Active Index). This study finally obtained their spatial distribution map, and showed that the perceptual dimensions lively, boring, safe, and depressing are correlated with information quantity of activities. The paper provides a novel method to understand better the urban environment and the distribution of residents' activities to facilitate urban planning.

Identifying social relationships from images is a challenging yet promising research area with great potential for improving human health and enhancing our understanding of social networks. However, present endeavors in this field tend to concentrate on leveraging visual features for the exploration of social relationships, while disregarding certain concealed information that lies beneath these features, such as interpersonal similarity. These methodologies may result in inadequate visual data encoding, thereby imposing limitations on the accuracy of social relationship recognition. In light of this, we propose a novel framework that utilizes interpersonal similarities within images to provide more information for identifying social relationships, thereby mitigating the issue of insufficient feature exploration. Furthermore, our proposed framework incorporates an innovative CF-Loss function that effectively incentivizes the identification of accurate social relationships while penalizing incorrect identifications, ultimately bolstering the model's capacity to discriminate between distinct social relationships. Our experimental findings demonstrate the superiority of our proposed framework over state-of-the-art methods on public datasets, confirming its effectiveness and accuracy in identifying social relationships.

In urban public spaces, a social group consists of two or more individuals who share some social relationships and interact based on mutual expectations. However, most existing studies found people’s F-formations on a top view, which is hard to observe their social contexts and the top-view videos are not easily accessible in real urban life. Recently, some researchers turned to urban scenes and analysed front-view human behaviours for social group detection. But these methods still cannot grasp the nature of social groups, i.e., the relationships among individuals. It is the key to finding social groups to judge whether any two individuals belong to the same cluster. Therefore, this paper proposes a new paradigm: relationship existence recognition-based social group detection. Additionally, on top of the paradigm, we designed a new social group detection algorithm incorporated with the visual cue-based and non-visual cue-based components. Specifically, the former exploits the spatial interactions and the temporal information to recognise the existence of social relationships through supervised deep learning. The latter estimates the similarities of trajectory pairs using the unsupervised spatial–temporal position information. Social group detection achieves superior accuracy with the two components’ complementary results. On Social-CAD (Social Collective Activity Dataset) and PLPS (Public Life in Public Space) datasets, extensive experiments demonstrate that our algorithm outperforms the state-of-the-art (SOTA) methods.

Systems for serving inference requests on graph neural networks (GNN) must combine low latency with high throughout, but they face irregular computation due to skew in the number of sampled graph nodes and aggregated GNN features. This makes it challenging to exploit GPUs effectively: using GPUs to sample only a few graph nodes yields lower performance than CPU-based sampling; and aggregating many features exhibits high data movement costs between GPUs and CPUs. Therefore, current GNN serving systems use CPUs for graph sampling and feature aggregation, limiting throughput. We describe Quiver, a distributed GPU-based GNN serving system with low-latency and high-throughput. Quiver's key idea is to exploit workload metrics for predicting the irregular computation of GNN requests, and governing the use of GPUs for graph sampling and feature aggregation: (1) for graph sampling, Quiver calculates the probabilistic sampled graph size, a metric that predicts the degree of parallelism in graph sampling. Quiver uses this metric to assign sampling tasks to GPUs only when the performance gains surpass CPU-based sampling; and (2) for feature aggregation, Quiver relies on the feature access probability to decide which features to partition and replicate across a distributed GPU NUMA topology. We show that Quiver achieves up to 35 times lower latency with an 8 times higher throughput compared to state-of-the-art GNN approaches (DGL and PyG).

The leachate from waste phosphogypsum contains substantial amounts of phosphorus, fluorine, ammonia nitrogen (NH3-N), and magnesium ions, which cause soil and water pollution and have great resource utilization value. In this study, quicklime and alkali were used to treat waste phosphogypsum leachate using step-by-step precipitation with precisely controlled pH values and selective recovery of valuable precipitates. The addition of CaO and NaOH in a specific order to adjust the pH value helped recover NH3-N, phosphate (PO43-), and fluoride (F-) in the leachate with a recovery rate above 99%. After the reaction, the residual ion concentration in the post leachate met the first level standard of the "Integrated Wastewater Discharge Standard " (GB8978-1996) (NH3-N < 5 mg/L, TP < 0.5 mg/L, and F- < 5 mg/L). This study showed that phosphogypsum leachate treatment and standard reuse can be achieved simultaneously. X-ray diffraction (XRD) and scanning electron microscopy (SEM) were used to analyze the composition of the precipitates produced at pH 3, 6, 8.5, and 12 to provide basic data for further experimental research.

Traditional wet and dry Chromite Ore Processing Residue (COPR) detoxification methods consume acid for pH adjustment and produce greenhouse gasses (CO2), yielding potential secondary pollution. It is imperative to develop an environmentally-friendly, efficient, and effective method for COPR treatment. This study reported the one-step natural pyrite-assisted mechanochemical detoxification of Cr(VI) in COPR. Under optimal conditions (i.e., the pyrite/COPR mass ratio of 5% and mechanochemical treatment at 600 rev./min for 2.5 h), Cr(VI) in COPR was efficiently reduced and immobilized, and the leaching concentration of total Cr met the regulatory limit of 5 mg/L. The X-ray absorption near-edge structure (XANES) analysis further verified that the Cr(VI) in treated COPR was 100% reduced to Cr(III). Potential mechanism could be that the enclosed non-exchangeable Cr(VI) in the chromite matrix was exposed to the COPR particle surface under the mechanical action; then, the redox reaction between pyrite and Cr(VI) were induced and enhanced through accumulated crystal defects and generated new surfaces with dangling bonds under mechanical forces. Life cycle assessment (LCA) results highlight that the high energy input is the major contributor eliciting environmental impacts for this current technology. More efficient and cleaner energy supplies should be sought alternatively to overcome the overall sustainability.

Mg/Al-LDH is widely used in the removal of aqueous pollutants due to the large surface and high anion exchange capacity, but the traditional Mg/Al-LDH for arsenic removal has a narrow range of pH application. The present study focused on the modifying of the lamellar Mg/Al-LDH to synthesize flower-like nanoparticles for arsenate removal in a wide pH range. Results showed that Mg/Al/Fe-CLDH (calcined Mg/Al/Fe-LDH) exhibited an extremely high arsenate removal efficiency (e.g., reducing As(V) from 10 mg/L to 1 mu g/L) in a wide pH range (2-12). Furthermore, it can keep excellent removal efficiency for arsenate under a complex anions environment (co-existing with CO32-, SO42-, PO43-, and F-), and maintain 80% removal efficiency for arsenate after 4 regeneration cycles. SEM, XRD, and XPS results unveiled the following enhancing mechanism for arsenate removal by using this modified Mg/Al-LDH: 1) partial content of ferric ions were successfully insert into the layers of Mg/Al-LDH and substituted by aluminium ions, which produce more affiliative sites for arsenate and expand its feasibility in a wide pH ranges; 2) the flower-like structures facilitate to offer more adsorption sites for arsenate due to its larger specific surface area. In brief, this flower-like Mg/Al/Fe-CLDH has great application potential for eco-friendly purity of arsenate from complicated industrial wastewater.

Heart sound diagnosis and classification play an essential role in detecting cardiovascular disorders, especially when the remote diagnosis becomes standard clinical practice. Most of the current work is designed for single category based heard sound classification tasks. To further extend the landscape of the automatic heart sound diagnosis landscape, this work proposes a deep multilabel learning model that can automatically annotate heart sound recordings with labels from different label groups, including murmur's timing, pitch, grading, quality, and shape. Our experiment results show that the proposed method has achieved outstanding performance on the holdout data for the multi-labelling task with sensitivity=0.990, specificity=0.999, F1=0.990 at the segments level, and an overall accuracy=0.969 at the patient's recording level.

Liquid Li–Pb alloy is served as neutron multiplier, tritium breeder, and coolant in Dual Coolant Lithium–Lead (DCLL) breeding blankets. In order to monitor the tritium transport continuously and measure the tritium production rate accurately, the concentration of hydrogen isotopes in liquid breeder is an essential factor for tritium measurement. The hydrogen isotopes permeation-based capsule seems to be the most reliable sensor and is simple from the fabrication point of view. To verify the stability and repeatability of hydrogen isotopes concentration measured by permeation capsule in Li–Pb, a system based on a Nb-capsule and a quadrupole mass spectrometry was devised. The sensor would be applied to determine the variation rate of total pressure while a quadrupole mass spectrometry to measure partial pressure values of different hydrogen isotopes. The response time of the sensor is less than 100 s in liquid Li–Pb alloy and the growth rate of pressure has good repeatability in dynamic mode. Besides, calculations based on numerical simulation of the hydrogen isotopes sensor are presented.

A novel application, including mechanochemical pre-treating and alkali leaching, for arsenic selective leaching from copper smelter flue dusts (CSFUs) was developed to overcome the disadvantages of hydrometallurgical methods. Compared with raw CSFU powders, the mechanical-activated ones showed higher maximum arsenic leaching efficiency (increased by -20%), and lower apparent activation energy (decreased by -7 kJ.mol- 1). Furthermore, this novel process only consumed half of alkali and sulfides and needed one-third of the leaching time to compare with the ones used in the traditional alkali leaching process. The promoting effect of mechanical force on arsenic leaching firstly relied on the physical property changes of CSFU powders, including a decrease of particle sizes and an increase of the specific surface. Secondly, mechanochemical force converted As5+ species into reduced phases (e.g. As2O3, NaAsO2), and thio-arsenates (e.g. AsO2S23- , AsO3S3-), which could spur its leaching due to their stronger mobilities in the alkali solution within sulfides. Finally, mechanochemical activation could be facilitated to separate discrete soluble arsenic species or incorporated ones from sulfate minerals in the CSFUs. This work may have important implications for the development of new eco-friendly technologies for purifying arsenic-bearing materials.

Many multimedia developers are exploring for adopting Deep Reinforcement Learning (DRL) techniques in their applications. They however often find such an adoption challenging. Existing DRL libraries provide poor support for prototyping DRL agents (i.e., models), customising the agents, and comparing the performance of DRL agents. As a result, the developers often report low efficiency in developing DRL agents. In this paper, we introduce RLzoo, a new DRL library that aims to make the development of DRL agents efficient. RLzoo provides developers with (i) high-level yet flexible APIs for prototyping DRL agents, and further customising the agents for best performance, (ii) a model zoo where users can import a wide range of DRL agents and easily compare their performance, and (iii) an algorithm that can automatically construct DRL agents with custom components (which are critical to improve agent's performance in custom applications). Evaluation results show that RLzoo can effectively reduce the development cost of DRL agents, while achieving comparable performance with existing DRL libraries.