arXiv subjects

Michael Wang

Publications and source records attributed to Michael Wang.

At least 19 recordsLinked to original sources

arXiv2026

MetaKV: Adaptive KV Cache Compression for Constrained LLM Inference

Key--value (KV) cache compression is an effective way to reduce the memory overhead of large language model (LLM) inference, particularly for long-context workloads. However, existing compression methods make different trade-offs among accuracy, inference latency, and peak KV cache memory utilization, making a single fixed configuration unsuitable across different prompts and resource constraints. We introduce MetaKV, an adaptive framework that selects a KV cache compression configuration for each input prompt based on user-specified latency and peak memory budgets. MetaKV uses lightweight prediction models to estimate the end-to-end latency, peak memory, and probability of a correct response for each candidate configuration, and selects the configuration that best satisfies the latency-memory constraints while preserving accuracy. We evaluate MetaKV across ten configurations from three representative KV cache compression methods, KVQuant, H$_2$O, and RocketKV, together with an uncompressed FP16 configuration, on four datasets covering mathematics, science, commonsense reasoning, and reading comprehension. Across a wide range of latency and peak memory constraints, MetaKV consistently outperforms the best static configuration, improving constrained success rate (CSR), the fraction of prompts answered correctly while satisfying both constraints, by approximately 0.07 on average and up to 0.135. These results demonstrate the benefit of adapting KV cache compression to individual prompts and latency-memory constraints. Code is available at https://github.com/MichaelWang0505/MetaKV.git

cs.LG↗

arXiv2026

W-RAG: Source-Aware Retrieval for Enterprise Document Generation from Heterogeneous Knowledge Bases

Retrieval-Augmented Generation (RAG) enables large language models to incorporate external knowledge during generation, improving factual grounding and domain adaptability. However, existing RAG pipelines assume that evidence retrieved from multiple repositories can be ranked globally using a single similarity function. While suitable for open-domain retrieval, this assumption breaks down in enterprise document generation, where heterogeneous knowledge bases (such as policies, regulations, technical documentation, and departmental guidelines) serve distinct roles and must be jointly represented in the generated document. As a result, global ranking often produces unbalanced context dominated by a subset of sources, leading to incomplete enterprise drafts. To address this limitation, we propose W-RAG, a source-aware retrieval framework that performs ontology-guided retrieval, local ranking within each knowledge base, and source-level weighting to regulate evidence composition. We further introduce a new dataset for retrieval-grounded enterprise document generation spanning multiple document types and industry domains. Experiments show that standard RAG pipelines struggle on this task, while W-RAG significantly improves document coverage and generation quality.

cs.SE↗

arXiv2026

Benchmarking LLM Serving Systems for Agentic AI Workloads with XPerf

We present XPerf, a benchmarking framework that load-tests LLM serving systems with diverse agentic AI workloads. It provides detailed profiling of the serving system and hardware, enabling users to identify performance bottlenecks introduced by agentic workloads. Benchmarking LLM serving systems under agentic workloads is challenging - agentic applications rely on nondeterministic LLM outputs to guide their control flow; therefore, workload patterns vary unpredictably from run to run. XPerf minimizes this workload variation with a fine-grained trace replay approach: it enables users to easily collect traces from real agentic applications, synthesize new workloads with various patterns if needed, and reproducibly replay them on different LLM serving systems. XPerf includes eight agentic applications across diverse use cases (e.g., coding, deep research, and Q&A) by default. Our empirical study using these workloads shows that XPerf accurately replays agentic workloads, provides detailed performance breakdowns, scales to larger serving systems, and assists in serving system debugging. We will open-source XPerf on GitHub.

cs.DC↗

arXiv2026

Exploring the Efficiency of 3D-Stacked AI Chip Architecture for LLM Inference with Voxel

To overcome the well-known memory bottleneck of AI chips, 3D stacked architectures that employ advanced packaging technology with high-density through-silicon vias (TSVs) pins have proven to be a promising solution. The 3D-stacked AI chip enables ultra-high memory bandwidth between compute and memory by stacking numerous DRAM banks atop many AI cores in a distributed manner. However, it is not easy to explore the efficiency of the 3D-stacked AI chip, due to its unique distributed nature. And we need to carefully consider multiple intertwined factors that range from upper-level computing paradigm to machine learning (ML) compiler optimizations, and to the underlying hardware architecture. In this paper, we develop Voxel, a fast and compiler-aware end-to-end simulation framework to facilitate exploring the efficiency of 3D-stacked AI chips for large language model (LLM) inference. Voxel enables the software/hardware co-exploration by employing a programming interface that allows ML compilers to customize the model execution plans. After validating the results of Voxel with an emulator on real silicon, we thoroughly examine the impact and correlation of different aspects of 3D-stacked AI chips, including state-of-the-art compute paradigms, tile-to-core mapping, tensor-to-bank mapping, NoC topologies and link bandwidth, DRAM bank bandwidth, per-core SRAM capacity, and energy/thermal constraints. Our findings disclose that the end-to-end efficiency of a 3D stacked AI chip not only is determined by the cooperative function of these factors, but also significantly depends on the mappings from tiles to AI core and DRAM banks. We report our findings throughout the paper, with the expectation that they will shed light on the development of the 3D-stacked AI chip ecosystem. We will open source Voxel and our study results for public research.

cs.AR↗

arXiv2026

A Complexity Agnostic Clustering Engine for Time Projection Chambers and its Implementation in FPGA

A clustering functional block implemented in field-programable-gate-array (FPGA) for time projection chambers (TPC) operating with predictable time regardless the complexity of the event is described in this paper. The clustering functional block reorganizes input data and the hits data belonging to the same clusters are output together for further process in the later stages. The clustering operation consists of two phases, data filling phase and data outputting phase, and the later uses the same number of clock cycles as the data filling phase. The clustering block can accommodate events with arbitrary number of clusters and number of hits per cluster as long as the total number of hits is within a predesigned limit. The operation time is exactly twice of the data filling time with no residual O(n2) term. The clustering block has been implemented with operating frequency of 200 MHz in a low-cost FPGA evaluation module and test results confirm the expected performance.

physics.ins-det↗

arXiv2026

Physics at the Edge: Benchmarking Quantisation Techniques and the Edge TPU for Neutrino Interaction Recognition

This work presents a comprehensive benchmark of different quantisation techniques for convolutional neural networks applied to neutrino interaction recognition. Utilising simulation for a generic liquid argon time-projection chamber, models are quantised and then deployed on the Google Coral Edge TPU. Models are tasked with recognising which neutrino interaction is simulated in the image between neutral current, muon-neutrino charged current, and electron-neutrino charged current. Four Keras models are tested, and accuracy is measured across two different pipelines: using post-training integer quantisation and quantisation-aware training. Inference speed is benchmarked against an AMD EPYC 7763 CPU and NVIDIA A100 GPU. A study of the energy consumption is also presented, with attention to potential costs and environmental issues. Results show that, among the four models tested, accuracy degradation is limited and, in particular, Inception V3 presents almost no accuracy degradation across the two quantisation and deployment pipelines. The speed of the edge TPU is comparable to that of the CPU, and one order of magnitude slower than the GPU. Moreover, the energy consumption of all models deployed on the edge TPU is several orders of magnitude lower than that of the CPU and GPU. In the energy consumption-latency parameter space, CPU, GPU, and edge TPU performances can be clearly separated. This paper explores possible future integrations of edge AI technologies with neutrino physics.

physics.ins-det↗

arXiv2026

Learning Perceptual Representations for Gaming NR-VQA with Multi-Task FR Signals

No-reference video quality assessment (NR-VQA) for gaming videos is challenging due to limited human-rated datasets and unique content characteristics including fast motion, stylized graphics, and compression artifacts. We present MTL-VQA, a multi-task learning framework that uses full-reference (FR) quality metrics as supervisory signals to learn perceptually meaningful features without human labels during pretraining. By jointly optimizing multiple complementary proxy FR objectives with adaptive task weighting, our approach learns shared representations that transfer effectively to downstream NR-VQA. Experiments on gaming video datasets show that MTL-VQA achieves competitive performance against state-of-the-art methods in both mean opinion score-supervised and label-efficient or self-supervised settings.

eess.IV↗

arXiv2026

A New Workflow for Materials Discovery Bridging the Gap Between Experimental Databases and Graph Neural Networks

Incorporating Machine Learning (ML) into material property prediction has become a crucial step in accelerating materials discovery. A key challenge is the severe lack of training data, as many properties are too complicated to calculate with high-throughput first principles techniques. To address this, recent research has created experimental databases from information extracted from scientific literature. However, most existing experimental databases do not provide full atomic coordinate information, which prevents them from supporting advanced ML architectures such as Graph Neural Networks (GNNs). In this work, we propose to bridge this gap through an alignment process between experimental databases and Crystallographic Information Files (CIF) from the Inorganic Crystal Structure Database (ICSD). Our approach enables the creation of a database that can fully leverage state-of-the-art model architectures for material property prediction. It also opens the door to utilizing transfer learning to improve prediction accuracy. To validate our approach, we align NEMAD with the ICSD and compare models trained on the resulting database to those trained on NEMAD originally. We demonstrate significant improvements in both Mean Absolute Error (MAE) and Correct Classification Rate (CCR) in predicting the ordering temperatures and magnetic ground states of magnetic materials, respectively.

cond-mat.mtrl-sci↗

arXiv2025

Deploying UDM Series in Real-Life Stuttered Speech Applications: A Clinical Evaluation Framework

Stuttered and dysfluent speech detection systems have traditionally suffered from the trade-off between accuracy and clinical interpretability. While end-to-end deep learning models achieve high performance, their black-box nature limits clinical adoption. This paper looks at the Unconstrained Dysfluency Modeling (UDM) series-the current state-of-the-art framework developed by Berkeley that combines modular architecture, explicit phoneme alignment, and interpretable outputs for real-world clinical deployment. Through extensive experiments involving patients and certified speech-language pathologists (SLPs), we demonstrate that UDM achieves state-of-the-art performance (F1: 0.89+-0.04) while providing clinically meaningful interpretability scores (4.2/5.0). Our deployment study shows 87% clinician acceptance rate and 34% reduction in diagnostic time. The results provide strong evidence that UDM represents a practical pathway toward AI-assisted speech therapy in clinical environments.

cs.SD↗

arXiv2025

A Comparative Study of Controllability, Explainability, and Performance in Dysfluency Detection Models

Recent advances in dysfluency detection have introduced a variety of modeling paradigms, ranging from lightweight object-detection inspired networks (YOLOStutter) to modular interpretable frameworks (UDM). While performance on benchmark datasets continues to improve, clinical adoption requires more than accuracy: models must be controllable and explainable. In this paper, we present a systematic comparative analysis of four representative approaches--YOLO-Stutter, FluentNet, UDM, and SSDM--along three dimensions: performance, controllability, and explainability. Through comprehensive evaluation on multiple datasets and expert clinician assessment, we find that YOLO-Stutter and FluentNet provide efficiency and simplicity, but with limited transparency; UDM achieves the best balance of accuracy and clinical interpretability; and SSDM, while promising, could not be fully reproduced in our experiments. Our analysis highlights the trade-offs among competing approaches and identifies future directions for clinically viable dysfluency modeling. We also provide detailed implementation insights and practical deployment considerations for each approach.

cs.AI↗

arXiv2025

Self-limiting states of polar misfits: Frustrated assembly of warped-jigsaw particles

We study the ground state thermodynamics of a model class of geometrically frustrated assemblies, known as {\it warped-jigsaw} particles. While it is known that frustration in soft matter assemblies has the ability to propagate up to mesoscopic, multi-particle size scales, notably through the selection of self-limiting domain, little is understood about how the symmetry of shape-misfit at the particle scale influences emergent morphologies at the mesoscale. Here we show that polarity in the shape-misfit of warped-jigsaw puzzles manifests at a larger scale in the morphology and thermodynamics of the ground-state assembly of self-limiting domains. We use a combination of continuum theory and discrete particle simulations to show that the polar misfit gives rise to two mesoscopically distinct polar, self-limiting ribbon domains. Thermodynamic selection between the two ribbon morphologies is controlled by a combination of the binding anisotropy along distinct neighbor directions and the orientation of polar shape-misfit. These predictions are valuable as design features for ongoing efforts to program self-limiting assemblies through the synthesis of intentionally frustrated particles, and further suggests a generic classification of frustrated assembly behavior in terms of the relative symmetries of shape-misfit and the underlying long-range inter-particle order it frustrates.

cond-mat.soft↗

arXiv2025

Evaluation of Rail Decarbonization Alternatives: Framework and Application

The Northwestern University Freight Rail Infrastructure and Energy Network Decarbonization (NUFRIEND) framework is a comprehensive industry-oriented tool for simulating the deployment of new energy technologies including biofuels, e-fuels, battery-electric, and hydrogen locomotives. By classifying fuel types into two categories based on deployment requirements, the associated optimal charging/fueling facility location and sizing problem are solved with a five-step framework. Life cycle analyses (LCA) and techno-economic analyses (TEA) are used to estimate carbon reduction, capital investments, cost of carbon reduction, and operational impacts, enabling sensitivity analysis with operational and technological parameters. The framework is illustrated on lower-carbon drop-in fuels as well as battery-electric technology deployments for US Eastern and Western Class I railroad networks. Drop-in fuel deployments are modeled as admixtures with diesel in existing locomotives, while battery-electric deployments are shown for varying technology penetration levels and locomotive ranges. When mixed in a 50 percent ratio with diesel, results show biodiesel's capacity to reduce emissions at 36 percent with a cost of 0.13 USD per kilogram of CO2 reduced, while e-fuels offer a 50 percent emissions reduction potential at a cost of 0.22 USD per kilogram of CO2 reduced. Battery-electric results for 50 percent deployment over all ton-miles highlight the value of future innovations in battery energy densities as scenarios assuming 800-mile range locomotives show an estimated emissions reduction of 46 percent with a cost of 0.06 USD per kilogram of CO2 reduced, compared to 16 percent emissions reduction at a cost of 0.11 USD per kilogram of CO2 reduced for 400-mile range locomotives.

eess.SY↗

arXiv2024

BetaExplainer: A Probabilistic Method to Explain Graph Neural Networks

Graph neural networks (GNNs) are powerful tools for conducting inference on graph data but are often seen as "black boxes" due to difficulty in extracting meaningful subnetworks driving predictive performance. Many interpretable GNN methods exist, but they cannot quantify uncertainty in edge weights and suffer in predictive accuracy when applied to challenging graph structures. In this work, we proposed BetaExplainer which addresses these issues by using a sparsity-inducing prior to mask unimportant edges during model training. To evaluate our approach, we examine various simulated data sets with diverse real-world characteristics. Not only does this implementation provide a notion of edge importance uncertainty, it also improves upon evaluation metrics for challenging datasets compared to state-of-the art explainer methods.

stat.ML↗

arXiv2024

FogROS2-FT: Fault Tolerant Cloud Robotics

Cloud robotics enables robots to offload complex computational tasks to cloud servers for performance and ease of management. However, cloud compute can be costly, cloud services can suffer occasional downtime, and connectivity between the robot and cloud can be prone to variations in network Quality-of-Service (QoS). We present FogROS2-FT (Fault Tolerant) to mitigate these issues by introducing a multi-cloud extension that automatically replicates independent stateless robotic services, routes requests to these replicas, and directs the first response back. With replication, robots can still benefit from cloud computations even when a cloud service provider is down or there is low QoS. Additionally, many cloud computing providers offer low-cost spot computing instances that may shutdown unpredictably. Normally, these low-cost instances would be inappropriate for cloud robotics, but the fault tolerance nature of FogROS2-FT allows them to be used reliably. We demonstrate FogROS2-FT fault tolerance capabilities in 3 cloud-robotics scenarios in simulation (visual object detection, semantic segmentation, motion planning) and 1 physical robot experiment (scan-pick-and-place). Running on the same hardware specification, FogROS2-FT achieves motion planning with up to 2.2x cost reduction and up to a 5.53x reduction on 99 Percentile (P99) long-tail latency. FogROS2-FT reduces the P99 long-tail latency of object detection and semantic segmentation by 2.0x and 2.1x, respectively, under network slowdown and resource contention.

cs.RO↗

arXiv2024

Machine-Type Communication Waveforms: An Exploration of New Dimensions

This paper derives a generalized class of waveforms with an application to machine-type communication (MTC) while studying its underlying structural characteristics in relation to conventional modulation waveforms. First, a canonical waveform of frequency-error tolerance is identified for a unified preamble and traffic signal design, ideal for MTC use as a composite waveform, commonly known as a transmission burst. It is shown that the most widely used modulation schemes for mIoT traffic signals, e.g., FSK and LoRa modulation, are simply subsets of the canonical waveform. The intrinsic characteristics and degrees of freedom the waveform offers are then explored. Most significantly, a new waveform dimension is uncovered and exploited as additional degrees of freedom for satisfying the MTC requirements, i.e., energy and resource efficiency and robustness. The corresponding benefits are evaluated analytically and numerically in AWGN, frequency-flat, and selective channels. We demonstrate that neither FSK nor LoRa can fully address the mIoT requirements since neither fully exploits the degrees of freedom from the perspective of the generalized waveform class. Finally, a solution is devised to optimize energy and resource efficiency under various deployment environments and practical constraints while maintaining the low-complexity property.

eess.SP↗

arXiv2024

Synthetic Datasets for Program Similarity Research

Program similarity has become an increasingly popular area of research with various security applications such as plagiarism detection, author identification, and malware analysis. However, program similarity research faces a few unique dataset quality problems in evaluating the effectiveness of novel approaches. First, few high-quality datasets for binary program similarity exist and are widely used in this domain. Second, there are potentially many different, disparate definitions of what makes one program similar to another and in many cases there is often a large semantic gap between the labels provided by a dataset and any useful notion of behavioral or semantic similarity. In this paper, we present HELIX - a framework for generating large, synthetic program similarity datasets. We also introduce Blind HELIX, a tool built on top of HELIX for extracting HELIX components from library code automatically using program slicing. We evaluate HELIX and Blind HELIX by comparing the performance of program similarity tools on a HELIX dataset to a hand-crafted dataset built from multiple, disparate notions of program similarity. Using Blind HELIX, we show that HELIX can generate realistic and useful datasets of virtually infinite size for program similarity research with ground truth labels that embody practical notions of program similarity. Finally, we discuss the results and reason about relative tool ranking.

cs.CR↗

arXiv2024

Lifelong LERF: Local 3D Semantic Inventory Monitoring Using FogROS2

Inventory monitoring in homes, factories, and retail stores relies on maintaining data despite objects being swapped, added, removed, or moved. We introduce Lifelong LERF, a method that allows a mobile robot with minimal compute to jointly optimize a dense language and geometric representation of its surroundings. Lifelong LERF maintains this representation over time by detecting semantic changes and selectively updating these regions of the environment, avoiding the need to exhaustively remap. Human users can query inventory by providing natural language queries and receiving a 3D heatmap of potential object locations. To manage the computational load, we use Fog-ROS2, a cloud robotics platform, to offload resource-intensive tasks. Lifelong LERF obtains poses from a monocular RGBD SLAM backend, and uses these poses to progressively optimize a Language Embedded Radiance Field (LERF) for semantic monitoring. Experiments with 3-5 objects arranged on a tabletop and a Turtlebot with a RealSense camera suggest that Lifelong LERF can persistently adapt to changes in objects with up to 91% accuracy.

cs.RO↗

arXiv2023

Thermal stability and secondary aggregation of self-limiting, geometrically-frustrated assemblies: Chain assembly of incommensurate polybricks

In geometrically frustrated assemblies, equilibrium self-limitation manifests in the form of a minimum in the free energy per subunit at a finite, multi-subunit size which results from the competition between the elastic costs of frustration within an assembly and the surface energy at its boundaries. Physical realizations -- from ill-fitting particle assemblies to self-twisting protein superstructures -- are capable of multiple mechanisms of escaping the cumulative costs of frustration, resulting in unlimited equilibrium assembly, including elastic modes of ``shape-flattening'' and the formation of weak, defective bonds that screen intra-assembly stresses. Here we study a model of 1D chain assembly of incommensurate ``polybricks'', and determine its equilibrium assembly as a function of temperature, concentration, degree of shape frustration, elasticity and inter-particle binding, notably focusing on how weakly cohesive, defective bonds give rise to strongly temperature-dependent assembly. Complex assembly behavior derives from the competition between multiple distinct local minima in the free energy landscape, including self-limiting chains, weakly-bound aggregates of self-limiting chains, and strongly-bound, elastically defrustrated assemblies. We show that this scenario, in general, gives rise to anomalous {\it multiple aggregation} behavior, in which disperse subunits first exhibit a primary aggregation transition to self-limiting chains (at intermediate concentration and temperature) which are ultimately unstable to condensation into unlimited assembly of finite-chains through weak binding at a secondary aggregation transition (at low temperature and high concentration). We show that window of stable self-limitation is determined both by the elastic costs of frustration in the assembly as well as energetic and entropic features of inter-subunit binding.

cond-mat.soft↗