Zitong Li

Saibal De, Zitong Li, Hemanth Kolla et al.

The Tucker decomposition, an extension of singular value decomposition for higher-order tensors, is a useful tool in analysis and compression of large-scale scientific data. While it has been studied extensively for static datasets, there are relatively few works addressing the computation of the Tucker factorization of streaming data tensors. In this paper we propose a new streaming Tucker algorithm tailored for scientific data, specifically for the case of a data tensor whose size increases along a single streaming mode that can grow indefinitely, which is typical of time-stepping scientific applications. At any point of this growth, we seek to compute the Tucker decomposition of the data generated thus far, without requiring storing the past tensor slices in memory. Our algorithm accomplishes this by starting with an initial Tucker decomposition and updating its components--the core tensor and factor matrices--with each new tensor slice as it becomes available, while satisfying a user-specified threshold of norm error. We present an implementation within the TuckerMPI software framework, and apply it to synthetic and combustion simulation datasets. By comparing against the standard (batch) decomposition algorithm we show that our streaming algorithm provides significant improvements in memory usage. If the tensor rank stops growing along the streaming mode, the streaming algorithm also incurs less computational time compared to the batch algorithm.

Arthur Feeney, Zitong Li, Ramin Bostanabad et al.

Mosaic Flow is a novel domain decomposition method designed to scale physics-informed neural PDE solvers to large domains. Its unique approach leverages pre-trained networks on small domains to solve partial differential equations on large domains purely through inference, resulting in high reusability. This paper presents an end-to-end parallelization of Mosaic Flow, combining data parallel training and domain parallelism for inference on large-scale problems. By optimizing the network architecture and data parallel training, we significantly reduce the training time for learning the Laplacian operator to minutes on 32 GPUs. Moreover, our distributed domain decomposition algorithm enables scalable inferences for solving the Laplace equation on domains 4096 times larger than the training domain, demonstrating strong scaling while maintaining accuracy on 32 GPUs. The reusability of Mosaic Flow, combined with the improved performance achieved through the distributed-memory algorithms, makes it a promising tool for modeling complex physical phenomena and accelerating scientific discovery.

Zitong Li, Haoyu Wang

Frozen Vision Foundation Models (VFMs) with lightweight classification heads are increasingly used in medical imaging because they offer efficient and reproducible deployment. Yet noisy-label learning methods for this frozen-feature regime remain poorly understood, and most existing methods still rely on a small-loss assumption inherited from end-to-end training. We present a controlled benchmark of eight noisy-label methods across five medical datasets, three backbones, two noise types, and five noise rates (150 conditions, 6,000 training runs), evaluated with balanced accuracy. The benchmark shows that there is no universal winner: Friedman ranking over the 150 conditions yields $χ^2 = 333.2$ ($p = 4.77 \times 10^{-68}$), ELR wins the most conditions (49/150), while CUFIT attains the best mean rank (2.51). The practical cost of method choice grows sharply with noise severity, from 4.5pp on clean data to 18.8pp at asymmetric 40\% noise. To explain these benchmark-level patterns, we revisit the small-loss assumption in a representative high-risk regime. Under frozen DINOv2 features, clean and noisy loss distributions overlap by 53--61\%, and matched-rate clean-sample detection shows that prediction agreement is markedly more stable than loss ranking under asymmetric noise (3pp vs.\ 13pp precision drop). On ISIC2019 with asymmetric 40\% noise, Co-Teaching reaches 68\% overall accuracy while collapsing to 35.1\% balanced accuracy with zero recall on three minority classes. Together, these results recast noisy-label learning for frozen VFMs as a regime-aware method-selection problem rather than a search for a single dominant algorithm. We conclude with evidence-based guidance and a low-regret feature-space selector for practical recommendation.

Sizhe Huang, Zitong Li, Shujie Yang

Self-supervised masked modeling shows promise for encrypted traffic classification by masking and reconstructing raw bytes. Yet recent work reveals these methods fail to reduce reliance on labeled data despite costly pretraining: under frozen encoder evaluation, accuracy drops from greater than 0.9 to less than 0.47. We argue the root cause is inductive bias mismatch: flattening traffic into byte sequences destroys protocol-defined semantics. We identify three specific issues: 1) field unpredictability, random fields like ip.id are unlearnable yet treated as reconstruction targets; 2) embedding confusion, semantically distinct fields collapse into a unified embedding space; 3) metadata loss, capture-time metadata essential for temporal analysis is discarded. To address this, we propose a protocol-native paradigm that treats protocol-defined field semantics as architectural priors, reformulating the task to align with the data's intrinsic tabular modality rather than incrementally adapting sequence-based architectures. Instantiating this paradigm, we introduce FlowSem-MAE, a tabular masked autoencoder built on Flow Semantic Units (FSUs). It features predictability-guided filtering that focuses on learnable FSUs, FSU-specific embeddings to preserve field boundaries, and dual-axis attention to capture intra-packet and temporal patterns. FlowSem-MAE significantly outperforms state-of-the-art across datasets. With only half labeled data, it outperforms most existing methods trained on full data.

Haoyu Wang, Zitong Li

Vision-language models (VLMs) are increasingly used for medical image interpretation, yet they frequently hallucinate, generating clinically plausible but factually incorrect findings that pose direct patient safety risks. We introduce HalluCXR, a benchmark evaluating six architecturally diverse VLMs across 856 stratified MIMIC-CXR chest radiographs and three query types, yielding 15,408 model evaluations. An eight-category hallucination taxonomy with clinical severity ratings and a two-layer detection pipeline are validated against 250 human annotations (auto-detection F1=0.959; LLM judge F1=0.907). We find that 61.9--82.3% of outputs contain hallucinations, with clinically dangerous errors in up to 80.2%. Three key patterns emerge: normal radiographs paradoxically attract the most severe hallucinations, common findings are systematically over-fabricated while rare findings go under-detected, and response length alone predicts hallucination risk (AUC up to 0.908). A six-model ensemble reduces fabrication by up to 84.8% at the cost of increased omission; a three-model subset retains comparable performance at half the cost. These results establish that hallucination auditing, verbosity-based risk monitoring, and ensemble-based safety layers are prerequisites for clinical deployment.

Zemin Chao, Boyu Xiao, Zitong Li et al.

Elastic similarity measures are fundamental to time series similarity search because of their ability to handle temporal misalignments. These measures are inherently computationally expensive, therefore necessitating the use of lower bounds to prune unnecessary comparisons. This paper proposes a new \emph{Bipartite Graph Edge-Cover Paradigm} for deriving lower bounds, which applies to a broad class of elastic similarity measures. This paradigm formulates lower bounding as a vertex-weighting problem on a weighted bipartite graph induced from the input time series. Under this paradigm, most of the existing lower bounds of elastic similarity measures can be viewed as simple instantiations. We further propose \textit{BGLB}, an instantiation of the proposed paradigm that incorporates an additional augmentation term, yielding lower bounds that are provably tighter. Theoretical analysis and extensive experiments on 128 real-world datasets demonstrate that \textit{BGLB} achieves the tightest known lower bounds for six elastic measures (ERP, MSM, TWED, LCSS, EDR, and SWALE). Moreover, \textit{BGLB} remains highly competitive for \textit{DTW} with a favorable trade-off between tightness and computational efficiency. In nearest neighbor search, integrating \textit{BGLB} into filter pipelines consistently outperforms state-of-the-art methods, achieving speedups ranging from $24.6\%$ to $84.9\%$ across various elastic similarity measures. Besides, \textit{BGLB} also delivers a significant acceleration in density-based clustering applications, validating the practical potential of \textit{BGLB} in time series similarity search tasks based on elastic similarity measures.

Zitong Li, Qingqing Ye, Haibo Hu

Machine unlearning is an emerging field that selectively removes specific data samples from a trained model. This capability is crucial for addressing privacy concerns, complying with data protection regulations, and correcting errors or biases introduced by certain data. Unlike traditional machine learning, where models are typically static once trained, machine unlearning facilitates dynamic updates that enable the model to ``forget'' information without requiring complete retraining from scratch. There are various machine unlearning methods, some of which are more time-efficient when data removal requests are fewer. To decrease the execution time of such machine unlearning methods, we aim to reduce the size of data removal requests based on the fundamental assumption that the removal of certain data would not result in a distinguishable retrained model. We first propose the concept of unnecessary unlearning, which indicates that the model would not alter noticeably after removing some data points. Subsequently, we review existing solutions that can be used to solve our problem. We highlight their limitations in adaptability to different unlearning scenarios and their reliance on manually selected parameters. We consequently put forward FUNU, a method to identify data points that lead to unnecessary unlearning. FUNU circumvents the limitations of existing solutions. The idea is to discover data points within the removal requests that have similar neighbors in the remaining dataset. We utilize a reference model to set parameters for finding neighbors, inspired from the area of model memorization. We provide a theoretical analysis of the privacy guarantee offered by FUNU and conduct extensive experiments to validate its efficacy.

Xiaoyu Xu, Minxin Du, Zitong Li et al.

Although machine unlearning is essential for removing private, harmful, or copyrighted content from LLMs, current benchmarks often fail to faithfully represent the true "forgetting scope" learned by the model. We formalize two distinct unlearning granularities, domain-level and instance-level, and propose BiForget, an automated framework for synthesizing high-quality forget sets. Unlike prior work relying on external generators, BiForget exploits the target model per se to elicit data that matches its internal knowledge distribution through seed-guided and adversarial prompting. Our experiments across diverse benchmarks show that it achieves a superior balance of relevance, diversity, and efficiency. Quantitatively, in the Harry Potter domain, it improves relevance by ${\sim}20$ and diversity by ${\sim}$0.05 while halving the total data size compared to SOTAs. Ultimately, it facilitates more robust forgetting and better utility preservation, providing a more rigorous foundation for evaluating LLM unlearning.

Tianhang Cheng, Albert J. Zhai, Evan Z. Chen et al.

Learning 3D parametric shape models of objects has gained popularity in vision and graphics and has showed broad utility in 3D reconstruction, generation, understanding, and simulation. While powerful models exist for humans and animals, equally expressive approaches for modeling plants are lacking. In this work, we present Demeter, a data-driven parametric model that encodes key factors of a plant morphology, including topology, shape, articulation, and deformation into a compact learned representation. Unlike previous parametric models, Demeter handles varying shape topology across various species and models three sources of shape variation: articulation, subcomponent shape variation, and non-rigid deformation. To advance crop plant modeling, we collected a large-scale, ground-truthed dataset from a soybean farm as a testbed. Experiments show that Demeter effectively synthesizes shapes, reconstructs structures, and simulates biophysical processes. Code and data is available at https://tianhang-cheng.github.io/Demeter/.

Zitong Li, Aparna Chandramowlishwaran

Sparse attention is a core building block in many leading neural network models, from graph-structured learning to sparse sequence modeling. It can be decomposed into a sequence of three sparse matrix operations (3S): sampled dense-dense matrix multiplication (SDDMM), softmax normalization, and sparse matrix multiplication (SpMM). Efficiently executing the 3S computational pattern on modern GPUs remains challenging due to (a) the mismatch between unstructured sparsity and tensor cores optimized for dense operations, and (b) the high cost of data movement. Previous works have optimized these sparse operations individually or addressed one of these challenges. This paper introduces Fused3S, the first fused 3S algorithm that jointly maximizes tensor core utilization and minimizes data movement. Across real-world graph datasets, Fused3S achieves $1.6- 16.3\times$ and $1.5-14\times$ speedup over state-of-the-art on H100 and A30 GPUs. Furthermore, integrating Fused3S into Graph Transformer inference accelerates end-to-end performance by $1.05-5.36\times$, consistently outperforming all 3S baselines across diverse datasets (single and batched graphs) and GPU architectures.