arXiv subjects

Weijian Zheng

Publications and source records attributed to Weijian Zheng.

16 recordsLinked to original sources

arXiv2026

Wrong but Useful: Trajectory Value Beyond Answer Correctness in Multi-Agent Messages

Multi-agent reasoning systems often use agreement, confidence, or automated scores to decide which messages should shape a final answer. Such filtering assumes that a message likely to be correct is also worth keeping. Yet a wrong answer can contain a useful decomposition, constraint, or scientific principle. We test this distinction with Diverse Hypothesis Deliberation (DHD), a controlled measurement protocol that caches five independently generated messages and replays the same downstream solver, called the integrator, with each message available or hidden. The replay comparison measures a message's trajectory value: whether making the message available helps or harms subsequent reasoning. Across five mathematics and science benchmarks and two openly available model families, gpt-oss-120b and gemma-4-31B-it, wrong-helpful messages appear in every benchmark-model combination. Among wrong-answer messages that change final correctness, more than four in ten changes are helpful in each model. Controlled repeats show that the number of repeatable message effects is unlikely to arise from replay variation alone (p=0.0002). A focused intervention on repeatable wrong-helpful messages finds that the complete message works best, while retaining its reasoning preserves more success than retaining only its answer; the source of the complete-message advantage remains open. Within the same problem, repeated trajectory-value evidence also identifies a better keep-or-remove choice than answer correctness alone. Answer correctness is therefore informative but does not determine trajectory value. DHD measures this missing property and produces reusable labels for learning when agents should listen.

cs.AI↗

arXiv2026

KV-RM: Regularizing KV-Cache Movement for Static-Graph LLM Serving

Static-graph LLM decoders provide predictable launches, fixed tensor shapes, and low submission overhead, but online decoding exposes highly irregular KV-cache behavior: request lengths differ, EOS events arrive asynchronously, and logical histories fragment over time. Dynamic runtimes recover flexibility through paged KV management and step-level scheduling, while static-graph executors often over-reserve memory and suffer burst-time latency outliers. This paper studies whether much of this variability can be absorbed below a fixed decode interface. We present KV-RM, a runtime design that regularizes KV-cache movement beneath a static-graph LLM decoder. KV-RM decouples logical KV histories from physical storage, tracks active KV state through a block pager, and materializes each decode step through a single committed descriptor. A merge-staged transport path coalesces non-contiguous KV mappings into a small number of large transfer groups before a fixed-shape attention kernel consumes them. Optional bounded far-history summaries can be enabled under the same interface, but the core design does not depend on them. On a 2-GPU NVIDIA A100 node, KV-RM improves mixed-length decoding throughput and tail latency relative to a static-graph baseline, reduces reserved KV memory across workload families, and removes severe burst-time latency spikes under production-trace replay. These results suggest that KV-cache movement, rather than kernel shape, can be an effective boundary for recovering runtime flexibility in static-graph LLM serving.

cs.AR↗

arXiv2026

HeteroViT: A Versatile Single-Layer Vision Transformer Concept, Co-Designed for Distributed Real-Time Data Reduction on Scientific Detectors

Next-generation X-ray detectors generate data faster than any system can affordably store or process. LCLS-II, the upgraded Linac Coherent Light Source at SLAC, produces data on the order of terabytes per second, with raw-data transfer and storage projected to be prohibitively costly, even though much of the data is not scientifically useful. This concept paper focuses on two major points. The first is versatility: a deliberately tiny, single-layer Vision Transformer (ViT) is enough to serve distinct scientific quick-evaluation tasks. We demonstrate this on two very different problems: (a) a supervised hit/miss/maybe classification on the CSPAD dataset, made to resemble ePixUHR-like detector frames, and (b) a self-supervised latent space for rare-event detection in X-ray diffraction spanning two learning paradigms, two output types, and two detector modalities, with one small backbone. The second is hardware co-design: because the ViT's blocks are structurally uniform, the model maps cleanly onto the heterogeneous hardware already present in the LCLS detector pipeline (ASIC -> FPGA -> GPU) under a simple rule one ASIC is one token so the data is reduced progressively at each stage and a keep/discard decision is produced in real time at the edge. The two claims reinforce each other: versatility is precisely what justifies freezing the front-end in silicon, since a reusable front-end is only worth committing to hardware if it serves many tasks. We are explicit that this is a concept supported by early software analysis, not a hardware demonstration. The natural and primary next phase is the hardware implementation of this distributed pipeline. The decisive evidence still owed an end-to-end latency budget, ASIC feasibility of the in-sensor embedding, and the false-negative behavior that matters for a data veto defines that program. HeteroViT is our first step toward it.

physics.ins-det↗

arXiv2026

SDiT: Semantic Region-Adaptive for Diffusion Transformers

Diffusion Transformers (DiTs) achieve state-of-the-art performance in text-to-image synthesis but remain computationally expensive due to the iterative nature of denoising and the quadratic cost of global attention. In this work, we observe that denoising dynamics are spatially non-uniform-background regions converge rapidly while edges and textured areas evolve much more actively. Building on this insight, we propose SDiT, a Semantic Region-Adaptive Diffusion Transformer that allocates computation according to regional complexity. SDiT introduces a training-free framework combining (1) semantic-aware clustering via fast Quickshift-based segmentation, (2) complexity-driven regional scheduling to selectively update informative areas, and (3) boundary-aware refinement to maintain spatial coherence. Without any model retraining or architectural modification, SDiT achieves up to 3.0x acceleration while preserving nearly identical perceptual and semantic quality to full-attention inference.

cs.CV↗

arXiv2025

DABench-LLM: Standardized and In-Depth Benchmarking of Post-Moore Dataflow AI Accelerators for LLMs

The exponential growth of large language models has outpaced the capabilities of traditional CPU and GPU architectures due to the slowdown of Moore's Law. Dataflow AI accelerators present a promising alternative; however, there remains a lack of in-depth performance analysis and standardized benchmarking methodologies for LLM training. We introduce DABench-LLM, the first benchmarking framework designed for evaluating LLM workloads on dataflow-based accelerators. By combining intra-chip performance profiling and inter-chip scalability analysis, DABench-LLM enables comprehensive evaluation across key metrics such as resource allocation, load balance, and resource efficiency. The framework helps researchers rapidly gain insights into underlying hardware and system behaviors, and provides guidance for performance optimizations. We validate DABench-LLM on three commodity dataflow accelerators, Cerebras WSE-2, SambaNova RDU, and Graphcore IPU. Our framework reveals performance bottlenecks and provides specific optimization strategies, demonstrating its generality and effectiveness across a diverse range of dataflow-based AI hardware platforms.

cs.AR↗

arXiv2025

To Stream or Not to Stream: Towards A Quantitative Model for Remote HPC Processing Decisions

Modern scientific instruments generate data at rates that increasingly exceed local compute capabilities and, when paired with the staging and I/O overheads of file-based transfers, also render file-based use of remote HPC resources impractical for time-sensitive analysis and experimental steering. Real-time streaming frameworks promise to reduce latency and improve system efficiency, but lack a principled way to assess their feasibility. In this work, we introduce a quantitative framework and an accompanying Streaming Speed Score to evaluate whether remote high-performance computing (HPC) resources can provide timely data processing compared to local alternatives. Our model incorporates key parameters including data generation rate, transfer efficiency, remote processing power, and file input/output overhead to compute total processing completion time and identify operational regimes where streaming is beneficial. We motivate our methodology with use cases from facilities such as APS, FRIB, LCLS-II, and the LHC, and validate our approach through an illustrative case study based on LCLS-II data. Our measurements show that streaming can achieve up to 97% lower end-to-end completion time than file-based methods under high data rates, while worst-case congestion can increase transfer times by over an order of magnitude, underscoring the importance of tail latency in streaming feasibility decisions.

cs.DC↗

arXiv2025

Globus Service Enhancements for Exascale Applications and Facilities

Many extreme-scale applications require the movement of large quantities of data to, from, and among leadership computing facilities, as well as other scientific facilities and the home institutions of facility users. These applications, particularly when leadership computing facilities are involved, can touch upon edge cases (e.g., terabyte files) that had not been a focus of previous Globus optimization work, which had emphasized rather the movement of many smaller (megabyte to gigabyte) files. We report here on how automated client-driven chunking can be used to accelerate both the movement of large files and the integrity checking operations that have proven to be essential for large data transfers. We present detailed performance studies that provide insights into the benefits of these modifications in a range of file transfer scenarios.

cs.NI↗

arXiv2025

Poly-FEVER: A Multilingual Fact Verification Benchmark for Hallucination Detection in Large Language Models

Hallucinations in generative AI, particularly in Large Language Models (LLMs), pose a significant challenge to the reliability of multilingual applications. Existing benchmarks for hallucination detection focus primarily on English and a few widely spoken languages, lacking the breadth to assess inconsistencies in model performance across diverse linguistic contexts. To address this gap, we introduce Poly-FEVER, a large-scale multilingual fact verification benchmark specifically designed for evaluating hallucination detection in LLMs. Poly-FEVER comprises 77,973 labeled factual claims spanning 11 languages, sourced from FEVER, Climate-FEVER, and SciFact. It provides the first large-scale dataset tailored for analyzing hallucination patterns across languages, enabling systematic evaluation of LLMs such as ChatGPT and the LLaMA series. Our analysis reveals how topic distribution and web resource availability influence hallucination frequency, uncovering language-specific biases that impact model accuracy. By offering a multilingual benchmark for fact verification, Poly-FEVER facilitates cross-linguistic comparisons of hallucination detection and contributes to the development of more reliable, language-inclusive AI systems. The dataset is publicly available to advance research in responsible AI, fact-checking methodologies, and multilingual NLP, promoting greater transparency and robustness in LLM performance. The proposed Poly-FEVER is available at: https://huggingface.co/datasets/HanzhiZhang/Poly-FEVER.

cs.CL↗

arXiv2024

GFormer: Accelerating Large Language Models with Optimized Transformers on Gaudi Processors

Heterogeneous hardware like Gaudi processor has been developed to enhance computations, especially matrix operations for Transformer-based large language models (LLMs) for generative AI tasks. However, our analysis indicates that Transformers are not fully optimized on such emerging hardware, primarily due to inadequate optimizations in non-matrix computational kernels like Softmax and in heterogeneous resource utilization, particularly when processing long sequences. To address these issues, we propose an integrated approach (called GFormer) that merges sparse and linear attention mechanisms. GFormer aims to maximize the computational capabilities of the Gaudi processor's Matrix Multiplication Engine (MME) and Tensor Processing Cores (TPC) without compromising model quality. GFormer includes a windowed self-attention kernel and an efficient outer product kernel for causal linear attention, aiming to optimize LLM inference on Gaudi processors. Evaluation shows that GFormer significantly improves efficiency and model performance across various tasks on the Gaudi processor and outperforms state-of-the-art GPUs.

cs.AR↗

arXiv2024

S3LLM: Large-Scale Scientific Software Understanding with LLMs using Source, Metadata, and Document

The understanding of large-scale scientific software poses significant challenges due to its diverse codebase, extensive code length, and target computing architectures. The emergence of generative AI, specifically large language models (LLMs), provides novel pathways for understanding such complex scientific codes. This paper presents S3LLM, an LLM-based framework designed to enable the examination of source code, code metadata, and summarized information in conjunction with textual technical reports in an interactive, conversational manner through a user-friendly interface. S3LLM leverages open-source LLaMA-2 models to enhance code analysis through the automatic transformation of natural language queries into domain-specific language (DSL) queries. Specifically, it translates these queries into Feature Query Language (FQL), enabling efficient scanning and parsing of entire code repositories. In addition, S3LLM is equipped to handle diverse metadata types, including DOT, SQL, and customized formats. Furthermore, S3LLM incorporates retrieval augmented generation (RAG) and LangChain technologies to directly query extensive documents. S3LLM demonstrates the potential of using locally deployed open-source LLMs for the rapid understanding of large-scale scientific computing software, eliminating the need for extensive coding expertise, and thereby making the process more efficient and effective. S3LLM is available at https://github.com/ResponsibleAILab/s3llm.

cs.SE↗

arXiv2024

Steering a Fleet: Adaptation for Large-Scale, Workflow-Based Experiments

Experimental science is increasingly driven by instruments that produce vast volumes of data and thus a need to manage, compute, describe, and index this data. High performance and distributed computing provide the means of addressing the computing needs; however, in practice, the variety of actions required and the distributed set of resources involved, requires sophisticated "flows" defining the steps to be performed on data. As each scan or measurement is performed by an instrument, a new instance of the flow is initiated resulting in a "fleet" of concurrently running flows, with the overall goal to process all the data collected during a potentially long-running experiment. During the course of the experiment, each flow may need to adapt its execution due to changes in the environment, such as computational or storage resource availability, or based on the progress of the fleet as a whole such as completion or discovery of an intermediate result leading to a change in subsequent flow's behavior. We introduce a cloud-based decision engine, Braid, which flows consult during execution to query their run-time environment and coordinate with other flows within their fleet. Braid accepts streams of measurements taken from the run-time environment or from within flow runs which can then be statistically aggregated and compared to other streams to determine a strategy to guide flow execution. For example, queue lengths in execution environments can be used to direct a flow to run computations in one environment or another, or experiment progress as measured by individual flows can be aggregated to determine the progress and subsequent direction of the flows within a fleet. We describe Braid, its interface, implementation and performance characteristics. We further show through examples and experience modifying an existing scientific flow how Braid is used to make adaptable flows.

cs.DC↗

arXiv2023

Rapid detection of rare events from in situ X-ray diffraction data using machine learning

High-energy X-ray diffraction methods can non-destructively map the 3D microstructure and associated attributes of metallic polycrystalline engineering materials in their bulk form. These methods are often combined with external stimuli such as thermo-mechanical loading to take snapshots over time of the evolving microstructure and attributes. However, the extreme data volumes and the high costs of traditional data acquisition and reduction approaches pose a barrier to quickly extracting actionable insights and improving the temporal resolution of these snapshots. Here we present a fully automated technique capable of rapidly detecting the onset of plasticity in high-energy X-ray microscopy data. Our technique is computationally faster by at least 50 times than the traditional approaches and works for data sets that are up to 9 times sparser than a full data set. This new technique leverages self-supervised image representation learning and clustering to transform massive data into compact, semantic-rich representations of visually salient characteristics (e.g., peak shapes). These characteristics can be a rapid indicator of anomalous events such as changes in diffraction peak shapes. We anticipate that this technique will provide just-in-time actionable information to drive smarter experiments that effectively deploy multi-modal X-ray diffraction methods that span many decades of length scales.

cs.LG↗

arXiv2023

Benchmarking and In-depth Performance Study of Large Language Models on Habana Gaudi Processors

Transformer models have achieved remarkable success in various machine learning tasks but suffer from high computational complexity and resource requirements. The quadratic complexity of the self-attention mechanism further exacerbates these challenges when dealing with long sequences and large datasets. Specialized AI hardware accelerators, such as the Habana GAUDI architecture, offer a promising solution to tackle these issues. GAUDI features a Matrix Multiplication Engine (MME) and a cluster of fully programmable Tensor Processing Cores (TPC). This paper explores the untapped potential of using GAUDI processors to accelerate Transformer-based models, addressing key challenges in the process. Firstly, we provide a comprehensive performance comparison between the MME and TPC components, illuminating their relative strengths and weaknesses. Secondly, we explore strategies to optimize MME and TPC utilization, offering practical insights to enhance computational efficiency. Thirdly, we evaluate the performance of Transformers on GAUDI, particularly in handling long sequences and uncovering performance bottlenecks. Lastly, we evaluate the end-to-end performance of two Transformer-based large language models (LLM) on GAUDI. The contributions of this work encompass practical insights for practitioners and researchers alike. We delve into GAUDI's capabilities for Transformers through systematic profiling, analysis, and optimization exploration. Our study bridges a research gap and offers a roadmap for optimizing Transformer-based model training on the GAUDI architecture.

cs.LG↗

arXiv2021

OpenGraphGym-MG: Using Reinforcement Learning to Solve Large Graph Optimization Problems on MultiGPU Systems

Large scale graph optimization problems arise in many fields. This paper presents an extensible, high performance framework (named OpenGraphGym-MG) that uses deep reinforcement learning and graph embedding to solve large graph optimization problems with multiple GPUs. The paper uses a common RL algorithm (deep Q-learning) and a representative graph embedding (structure2vec) to demonstrate the extensibility of the framework and, most importantly, to illustrate the novel optimization techniques, such as spatial parallelism, graph-level and node-level batched processing, distributed sparse graph storage, efficient parallel RL training and inference algorithms, repeated gradient descent iterations, and adaptive multiple-node selections. This study performs a comprehensive performance analysis on parallel efficiency and memory cost that proves the parallel RL training and inference algorithms are efficient and highly scalable on a number of GPUs. This study also conducts a range of large graph experiments, with both generated graphs (over 30 million edges) and real-world graphs, using a single compute node (with six GPUs) of the Summit supercomputer. Good scalability in both RL training and inference is achieved: as the number of GPUs increases from one to six, the time of a single step of RL training and a single step of RL inference on large graphs with more than 30 million edges, is reduced from 316.4s to 54.5s, and 23.8s to 3.4s, respectively. The research results on a single node lay out a solid foundation for the future work to address graph optimization problems with a large number of GPUs across multiple nodes in the Summit.

cs.DC↗

arXiv2020

Micromobility in Smart Cities: A Closer Look at Shared Dockless E-Scooters via Big Social Data

The micromobility is shaping first- and last-mile travels in urban areas. Recently, shared dockless electric scooters (e-scooters) have emerged as a daily alternative to driving for short-distance commuters in large cities due to the affordability, easy accessibility via an app, and zero emissions. Meanwhile, e-scooters come with challenges in city management, such as traffic rules, public safety, parking regulations, and liability issues. In this paper, we collected and investigated 5.8 million scooter-tagged tweets and 144,197 images, generated by 2.7 million users from October 2018 to March 2020, to take a closer look at shared e-scooters via crowdsourcing data analytics. We profiled e-scooter usages from spatial-temporal perspectives, explored different business roles (i.e., riders, gig workers, and ridesharing companies), examined operation patterns (e.g., injury types, and parking behaviors), and conducted sentiment analysis. To our best knowledge, this paper is the first large-scale systematic study on shared e-scooters using big social data.

cs.SI↗

arXiv2019

FQL: An Extensible Feature Query Language and Toolkit on Searching Software Characteristics for HPC Applications

The amount of large-scale scientific computing software is dramatically increasing. In this work, we designed a new language, named feature query language (FQL), to collect and extract software features from a quick static code analysis. We designed and implemented an FQL toolkit to automatically detect and present the software features using an extensible query repository. Several large-scale, high performance computing (HPC) scientific codes have been used in the paper to demonstrate the HPC-related feature extraction and information collection. Although we emphasized the HPC features in the study, the toolkit can be easily extended to answer general software feature questions, such as coding pattern and hardware dependency.

cs.SE↗