arXiv subjects

Di Wang

Publications and source records attributed to Di Wang.

At least 307 records · Page 17Linked to original sources

arXiv2025

Exploiting Student Parallelism for Efficient GPU Inference of BERT-like Models in Online Services

Due to high accuracy, BERT-like models have been widely adopted by text mining and web searching. However, large BERT-like models suffer from inefficient online inference, facing the following two problems on GPUs: (1) their high accuracy relies on the large model depth, which linearly increases the sequential computation on GPUs; (2) stochastic and dynamic online workloads cause extra costs from batching and paddings. Therefore, we present \sys for the real-world setting of GPU inference on online workloads. At its core, \sys adopts stacking distillation and boosting ensemble, distilling the original deep model into a group of shallow but virtually stacked student models running in parallel. This enables \sys to achieve a lower model depth (e.g., two layers) than the others and the lowest inference latency while maintaining accuracy. In addition, adaptive student pruning realizes dynamic student numbers according to changing online workloads. Especially for occasional workload bursts, it can temporarily decrease the student number with minimal accuracy loss to improve system throughput. We conduct comprehensive experiments to verify the effectiveness, whose results show that \sys outperforms the baselines by $4.1\times\sim 1.6\times$ in latency while maintaining accuracy and achieves up to $22.27\times$ higher throughput for workload bursts.

cs.LG↗

arXiv2025

Time-EAPCR-T: A Universal Deep Learning Approach for Anomaly Detection in Industrial Equipment

With the advancement of Industry 4.0, intelligent manufacturing extensively employs sensors for real-time multidimensional data collection, playing a crucial role in equipment monitoring, process optimisation, and efficiency enhancement. Industrial data exhibit characteristics such as multi-source heterogeneity, nonlinearity, strong coupling, and temporal interactions, while also being affected by noise interference. These complexities make it challenging for traditional anomaly detection methods to extract key features, impacting detection accuracy and stability. Traditional machine learning approaches often struggle with such complex data due to limitations in processing capacity and generalisation ability, making them inadequate for practical applications. While deep learning feature extraction modules have demonstrated remarkable performance in image and text processing, they remain ineffective when applied to multi-source heterogeneous industrial data lacking explicit correlations. Moreover, existing multi-source heterogeneous data processing techniques still rely on dimensionality reduction and feature selection, which can lead to information loss and difficulty in capturing high-order interactions. To address these challenges, this study applies the EAPCR and Time-EAPCR models proposed in previous research and introduces a new model, Time-EAPCR-T, where Transformer replaces the LSTM module in the time-series processing component of Time-EAPCR. This modification effectively addresses multi-source data heterogeneity, facilitates efficient multi-source feature fusion, and enhances the temporal feature extraction capabilities of multi-source industrial data.Experimental results demonstrate that the proposed method outperforms existing approaches across four industrial datasets, highlighting its broad application potential.

cs.LG↗

arXiv2025

ZO2: Scalable Zeroth-Order Fine-Tuning for Extremely Large Language Models with Limited GPU Memory

Fine-tuning large pre-trained LLMs generally demands extensive GPU memory. Traditional first-order optimizers like SGD encounter substantial difficulties due to increased memory requirements from storing activations and gradients during both the forward and backward phases as the model size expands. Alternatively, zeroth-order (ZO) techniques can compute gradients using just forward operations, eliminating the need to store activations. Furthermore, by leveraging CPU capabilities, it's feasible to enhance both the memory and processing power available to a single GPU. We propose a novel framework, ZO2 (Zeroth-Order Offloading), for efficient zeroth-order fine-tuning of LLMs with only limited GPU memory. Our framework dynamically shifts model parameters between the CPU and GPU as required, optimizing computation flow and maximizing GPU usage by minimizing downtime. This integration of parameter adjustments with ZO's double forward operations reduces unnecessary data movement, enhancing the fine-tuning efficacy. Additionally, our framework supports an innovative low-bit precision approach in AMP mode to streamline data exchanges between the CPU and GPU. Employing this approach allows us to fine-tune extraordinarily large models, such as the OPT-175B with more than 175 billion parameters, on a mere 18GB GPU--achievements beyond the reach of traditional methods. Moreover, our framework achieves these results with almost no additional time overhead and absolutely no accuracy loss compared to standard zeroth-order methods. ZO2's code has been open-sourced in https://github.com/liangyuwang/zo2.

cs.LG↗

arXiv2025

HunyuanVideo: A Systematic Framework For Large Video Generative Models

Recent advancements in video generation have significantly impacted daily life for both individuals and industries. However, the leading video generation models remain closed-source, resulting in a notable performance gap between industry capabilities and those available to the public. In this report, we introduce HunyuanVideo, an innovative open-source video foundation model that demonstrates performance in video generation comparable to, or even surpassing, that of leading closed-source models. HunyuanVideo encompasses a comprehensive framework that integrates several key elements, including data curation, advanced architectural design, progressive model scaling and training, and an efficient infrastructure tailored for large-scale model training and inference. As a result, we successfully trained a video generative model with over 13 billion parameters, making it the largest among all open-source models. We conducted extensive experiments and implemented a series of targeted designs to ensure high visual quality, motion dynamics, text-video alignment, and advanced filming techniques. According to evaluations by professionals, HunyuanVideo outperforms previous state-of-the-art models, including Runway Gen-3, Luma 1.6, and three top-performing Chinese video generative models. By releasing the code for the foundation model and its applications, we aim to bridge the gap between closed-source and open-source communities. This initiative will empower individuals within the community to experiment with their ideas, fostering a more dynamic and vibrant video generation ecosystem. The code is publicly available at https://github.com/Tencent/HunyuanVideo.

cs.CV↗

arXiv2025

How to Strategize Human Content Creation in the Era of GenAI?

Generative AI (GenAI) will have significant impact on content creation platforms. In this paper, we study the dynamic competition between a GenAI and a human contributor. Unlike the human, the GenAI's content only improves when more contents are created by the human over time; however, GenAI has the advantage of generating content at a lower cost. We study the algorithmic problem in this dynamic competition model about how the human contributor can maximize her utility when competing against the GenAI for content generation over a set of topics. In time-sensitive content domains (e.g., news or pop music creation) where contents' value diminishes over time, we show that there is no polynomial time algorithm for finding the human's optimal (dynamic) strategy, unless the randomized exponential time hypothesis is false. Fortunately, we are able to design a polynomial time algorithm that naturally cycles between myopically optimizing over a short time window and pausing and provably guarantees an approximation ratio of $\frac{1}{2}$. We then turn to time-insensitive content domains where contents do not lose their value (e.g., contents on history facts). Interestingly, we show that this setting permits a polynomial time algorithm that maximizes the human's utility in the long run. Finally, we conduct simulations that demonstrate the advantage of our algorithms in comparison to a collection of baselines.

cs.GT↗

arXiv2025

Online Bidding under RoS Constraints without Knowing the Value

We consider the problem of bidding in online advertising, where an advertiser aims to maximize value while adhering to budget and Return-on-Spend (RoS) constraints. Unlike prior work that assumes knowledge of the value generated by winning each impression ({e.g.,} conversions), we address the more realistic setting where the advertiser must simultaneously learn the optimal bidding strategy and the value of each impression opportunity. This introduces a challenging exploration-exploitation dilemma: the advertiser must balance exploring different bids to estimate impression values with exploiting current knowledge to bid effectively. To address this, we propose a novel Upper Confidence Bound (UCB)-style algorithm that carefully manages this trade-off. Via a rigorous theoretical analysis, we prove that our algorithm achieves $\widetilde{O}(\sqrt{T\log(|\mathcal{B}|T)})$ regret and constraint violation, where $T$ is the number of bidding rounds and $\mathcal{B}$ is the domain of possible bids. This establishes the first optimal regret and constraint violation bounds for bidding in the online setting with unknown impression values. Moreover, our algorithm is computationally efficient and simple to implement. We validate our theoretical findings through experiments on synthetic data, demonstrating that our algorithm exhibits strong empirical performance compared to existing approaches.

cs.LG↗

arXiv2025

GSA-Forecaster: Forecasting Graph-Based Time-Dependent Data with Graph Sequence Attention

Forecasting graph-based, time-dependent data has broad practical applications but presents challenges. Effective models must capture both spatial and temporal dependencies in the data, while also incorporating auxiliary information to enhance prediction accuracy. In this paper, we identify limitations in current state-of-the-art models regarding temporal dependency handling. To overcome this, we introduce GSA-Forecaster, a new deep learning model designed for forecasting in graph-based, time-dependent contexts. GSA-Forecaster utilizes graph sequence attention, a new attention mechanism proposed in this paper, to effectively manage temporal dependencies. GSA-Forecaster integrates the data's graph structure directly into its architecture, addressing spatial dependencies. Additionally, it incorporates auxiliary information to refine its predictions further. We validate its performance using real-world graph-based, time-dependent datasets, where it demonstrates superior effectiveness compared to existing state-of-the-art models.

cs.LG↗

arXiv2025

Automatic Linear Resource Bound Analysis for Rust via Prophecy Potentials

Rust has become a popular system programming language that strikes a balance between memory safety and performance. Rust's type system ensures the safety of low-level memory controls; however, a well-typed Rust program is not guaranteed to enjoy high performance. This article studies static analysis for resource consumption of Rust programs, aiming at understanding the performance of Rust programs. Although there have been tons of studies on static resource analysis, exploiting Rust's memory safety -- especially the borrow mechanisms and their properties -- to aid resource-bound analysis, remains unexplored. This article presents RaRust, a type-based linear resource-bound analysis for well-typed Rust programs. RaRust follows the methodology of automatic amortized resource analysis (AARA) to build a resource-aware type system. To support Rust's borrow mechanisms, including shared and mutable borrows, RaRust introduces shared and novel prophecy potentials to reason about borrows compositionally. To prove the soundness of RaRust, this article proposes Resource-Aware Borrow Calculus (RABC) as a variant of recently proposed Low-Level Borrow Calculus (LLBC). The experimental evaluation of a prototype implementation of RaRust demonstrates that RaRust is capable of inferring symbolic linear resource bounds for Rust programs featuring shared and mutable borrows, reborrows, heap-allocated data structures, loops, and recursion.

cs.PL↗

arXiv2025

Gate-Tunable Spin-to-Charge Conversion in Topological Insulator-Magnetic Insulator Heterostructures at Room Temperature

Over the past decade, topological insulators have received enormous attention for their potential in energy-efficient spin-to-charge conversion, enabled by strong spin-orbit coupling and spin-momentum locked surface states. Despite extensive research, the spin-to-charge conversion efficiency, usually characterized by the spin Hall angle (θSH), remains low at room temperature. In this work, we employed pulsed laser deposition to synthesize high-quality ternary topological insulators (Bi0.1Sb0.9)2Te3 thin films on magnetic insulator Y3Fe5O12. We find that the value of θSH reaches ~0.76 at room temperature and increases to ~0.9 as the Fermi level is tuned to cross topological surface states via electrical gating. Our findings provide an innovative approach to tailoring the spin-to-charge conversion in topological insulators and pave the way for their applications in energy-efficient spintronic devices.

cond-mat.mes-hall↗

arXiv2025

Dual Selective Fusion Transformer Network for Hyperspectral Image Classification

Transformer has achieved satisfactory results in the field of hyperspectral image (HSI) classification. However, existing Transformer models face two key challenges when dealing with HSI scenes characterized by diverse land cover types and rich spectral information: (1) A fixed receptive field overlooks the effective contextual scales required by various HSI objects; (2) invalid self-attention features in context fusion affect model performance. To address these limitations, we propose a novel Dual Selective Fusion Transformer Network (DSFormer) for HSI classification. DSFormer achieves joint spatial and spectral contextual modeling by flexibly selecting and fusing features across different receptive fields, effectively reducing unnecessary information interference by focusing on the most relevant spatial-spectral tokens. Specifically, we design a Kernel Selective Fusion Transformer Block (KSFTB) to learn an optimal receptive field by adaptively fusing spatial and spectral features across different scales, enhancing the model's ability to accurately identify diverse HSI objects. Additionally, we introduce a Token Selective Fusion Transformer Block (TSFTB), which strategically selects and combines essential tokens during the spatial-spectral self-attention fusion process to capture the most crucial contexts. Extensive experiments conducted on four benchmark HSI datasets demonstrate that the proposed DSFormer significantly improves land cover classification accuracy, outperforming existing state-of-the-art methods. Specifically, DSFormer achieves overall accuracies of 96.59%, 97.66%, 95.17%, and 94.59% in the Pavia University, Houston, Indian Pines, and Whu-HongHu datasets, respectively, reflecting improvements of 3.19%, 1.14%, 0.91%, and 2.80% over the previous model. The code will be available online at https://github.com/YichuXu/DSFormer.

cs.CV↗

arXiv2025

BeamVQ: Beam Search with Vector Quantization to Mitigate Data Scarcity in Physical Spatiotemporal Forecasting

In practice, physical spatiotemporal forecasting can suffer from data scarcity, because collecting large-scale data is non-trivial, especially for extreme events. Hence, we propose \method{}, a novel probabilistic framework to realize iterative self-training with new self-ensemble strategies, achieving better physical consistency and generalization on extreme events. Following any base forecasting model, we can encode its deterministic outputs into a latent space and retrieve multiple codebook entries to generate probabilistic outputs. Then BeamVQ extends the beam search from discrete spaces to the continuous state spaces in this field. We can further employ domain-specific metrics (e.g., Critical Success Index for extreme events) to filter out the top-k candidates and develop the new self-ensemble strategy by combining the high-quality candidates. The self-ensemble can not only improve the inference quality and robustness but also iteratively augment the training datasets during continuous self-training. Consequently, BeamVQ realizes the exploration of rare but critical phenomena beyond the original dataset. Comprehensive experiments on different benchmarks and backbones show that BeamVQ consistently reduces forecasting MSE (up to 39%), enhancing extreme events detection and proving its effectiveness in handling data scarcity.

cs.LG↗

arXiv2025

Towards User-level Private Reinforcement Learning with Human Feedback

Reinforcement Learning with Human Feedback (RLHF) has emerged as an influential technique, enabling the alignment of large language models (LLMs) with human preferences. Despite the promising potential of RLHF, how to protect user preference privacy has become a crucial issue. Most previous work has focused on using differential privacy (DP) to protect the privacy of individual data. However, they have concentrated primarily on item-level privacy protection and have unsatisfactory performance for user-level privacy, which is more common in RLHF. This study proposes a novel framework, AUP-RLHF, which integrates user-level label DP into RLHF. We first show that the classical random response algorithm, which achieves an acceptable performance in item-level privacy, leads to suboptimal utility when in the user-level settings. We then establish a lower bound for the user-level label DP-RLHF and develop the AUP-RLHF algorithm, which guarantees $(\varepsilon, δ)$ user-level privacy and achieves an improved estimation error. Experimental results show that AUP-RLHF outperforms existing baseline methods in sentiment generation and summarization tasks, achieving a better privacy-utility trade-off.

cs.LG↗

arXiv2025

Kernel-Based Distributed Q-Learning: A Scalable Reinforcement Learning Approach for Dynamic Treatment Regimes

In recent years, large amounts of electronic health records (EHRs) concerning chronic diseases have been collected to facilitate medical diagnosis. Modeling the dynamic properties of EHRs related to chronic diseases can be efficiently done using dynamic treatment regimes (DTRs). While reinforcement learning (RL) is a widely used method for creating DTRs, there is ongoing research in developing RL algorithms that can effectively handle large amounts of data. In this paper, we present a scalable kernel-based distributed Q-learning algorithm for generating DTRs. We perform both theoretical assessments and numerical analysis for the proposed approach. The results demonstrate that our algorithm significantly reduces the computational complexity associated with the state-of-the-art deep reinforcement learning methods, while maintaining comparable generalization performance in terms of accumulated rewards across stages, such as survival time or cumulative survival probability.

cs.LG↗

arXiv2025

Exploring the Personality Traits of LLMs through Latent Features Steering

Large language models (LLMs) have significantly advanced dialogue systems and role-playing agents through their ability to generate human-like text. While prior studies have shown that LLMs can exhibit distinct and consistent personalities, the mechanisms through which these models encode and express specific personality traits remain poorly understood. To address this, we investigate how various factors, such as cultural norms and environmental stressors, encoded within LLMs, shape their personality traits, guided by the theoretical framework of social determinism. Inspired by related work on LLM interpretability, we propose a training-free approach to modify the model's behavior by extracting and steering latent features corresponding to factors within the model, thereby eliminating the need for retraining. Furthermore, we analyze the implications of these factors for model safety, focusing on their impact through the lens of personality.

cs.CL↗

arXiv2025

Mechanistic Unveiling of Transformer Circuits: Self-Influence as a Key to Model Reasoning

Transformer-based language models have achieved significant success; however, their internal mechanisms remain largely opaque due to the complexity of non-linear interactions and high-dimensional operations. While previous studies have demonstrated that these models implicitly embed reasoning trees, humans typically employ various distinct logical reasoning mechanisms to complete the same task. It is still unclear which multi-step reasoning mechanisms are used by language models to solve such tasks. In this paper, we aim to address this question by investigating the mechanistic interpretability of language models, particularly in the context of multi-step reasoning tasks. Specifically, we employ circuit analysis and self-influence functions to evaluate the changing importance of each token throughout the reasoning process, allowing us to map the reasoning paths adopted by the model. We apply this methodology to the GPT-2 model on a prediction task (IOI) and demonstrate that the underlying circuits reveal a human-interpretable reasoning process used by the model.

cs.AI↗

arXiv2025

EAP-GP: Mitigating Saturation Effect in Gradient-based Automated Circuit Identification

Understanding the internal mechanisms of transformer-based language models remains challenging. Mechanistic interpretability based on circuit discovery aims to reverse engineer neural networks by analyzing their internal processes at the level of computational subgraphs. In this paper, we revisit existing gradient-based circuit identification methods and find that their performance is either affected by the zero-gradient problem or saturation effects, where edge attribution scores become insensitive to input changes, resulting in noisy and unreliable attribution evaluations for circuit components. To address the saturation effect, we propose Edge Attribution Patching with GradPath (EAP-GP), EAP-GP introduces an integration path, starting from the input and adaptively following the direction of the difference between the gradients of corrupted and clean inputs to avoid the saturated region. This approach enhances attribution reliability and improves the faithfulness of circuit identification. We evaluate EAP-GP on 6 datasets using GPT-2 Small, GPT-2 Medium, and GPT-2 XL. Experimental results demonstrate that EAP-GP outperforms existing methods in circuit faithfulness, achieving improvements up to 17.7%. Comparisons with manually annotated ground-truth circuits demonstrate that EAP-GP achieves precision and recall comparable to or better than previous approaches, highlighting its effectiveness in identifying accurate circuits.

cs.LG↗

arXiv2025

Robust spin-orbit coupling in semi-metallic SrIrO3 under hydrostatic pressure

The semi-metallic behavior of the perovskite iridate SrIrO3 shifts the end-member of the strongly spin-orbit (SO) coupled Ruddlesden-Popper series Srn+1IrnO3n+1 away from the Mott insulating regime and the half-filled pseudospin Jeff=1/2 ground state well-established in the layered iridates (n = 1 and 2). To investigate the robustness of the SO coupled ground state of SrIrO3, X-ray absorption spectroscopy was carried out at the Ir L2,3 edges under hydrostatic pressure up to 50 GPa at room temperature. The effective SO coupling was deduced from the branching ratio of the Ir L2 and L3 white lines. With increasing pressure, the branching ratio decreases, and the Ir L2,3 peak positions shift to higher energies. The number of 5d holes remains constant indicating that the spectral weight redistribution and peak shifts arise from orbital mixing between t2g and eg related states. The expectation value of the angular part of the SO operator decreases by about 15% at 50 GPa. This reduction, which is very similar to that observed in the layered iridates, is well explained by an increase of the octahedral crystal field due to the shortening of the Ir-O bond-length under compression. Consistent with theoretical predictions, the orbital mixing and decrease as the crystal field increases. However, the effective SO coupling remains robust against pressure and does not indicate a covalency-driven breakdown within the investigated pressure range.

cond-mat.str-el↗

arXiv2025

Editable Concept Bottleneck Models

Concept Bottleneck Models (CBMs) have garnered much attention for their ability to elucidate the prediction process through a humanunderstandable concept layer. However, most previous studies focused on cases where the data, including concepts, are clean. In many scenarios, we often need to remove/insert some training data or new concepts from trained CBMs for reasons such as privacy concerns, data mislabelling, spurious concepts, and concept annotation errors. Thus, deriving efficient editable CBMs without retraining from scratch remains a challenge, particularly in large-scale applications. To address these challenges, we propose Editable Concept Bottleneck Models (ECBMs). Specifically, ECBMs support three different levels of data removal: concept-label-level, concept-level, and data-level. ECBMs enjoy mathematically rigorous closed-form approximations derived from influence functions that obviate the need for retraining. Experimental results demonstrate the efficiency and adaptability of our ECBMs, affirming their practical value in CBMs.

cs.LG↗