Chang Wang

Haihao Shen, Naveen Mellempudi, Xin He et al.

Recent advances in deep learning methods such as LLMs and Diffusion models have created a need for improved quantization methods that can meet the computational demands of these modern architectures while maintaining accuracy. Towards this goal, we study the advantages of FP8 data formats for post-training quantization across 75 unique network architectures covering a wide range of tasks, including machine translation, language modeling, text generation, image classification, generation, and segmentation. We examine three different FP8 representations (E5M2, E4M3, and E3M4) to study the effects of varying degrees of trade-off between dynamic range and precision on model accuracy. Based on our extensive study, we developed a quantization workflow that generalizes across different network architectures. Our empirical results show that FP8 formats outperform INT8 in multiple aspects, including workload coverage (92.64% vs. 65.87%), model accuracy and suitability for a broader range of operations. Furthermore, our findings suggest that E4M3 is better suited for NLP models, whereas E3M4 performs marginally better than E4M3 on computer vision tasks. The code is publicly available on Intel Neural Compressor: https://github.com/intel/neural-compressor.

Rui Ai, Chang Wang, Chenchen Li et al.

Recently the online advertising market has exhibited a gradual shift from second-price auctions to first-price auctions. Although there has been a line of works concerning online bidding strategies in first-price auctions, it still remains open how to handle budget constraints in the problem. In the present paper, we initiate the study for a buyer with budgets to learn online bidding strategies in repeated first-price auctions. We propose an RL-based bidding algorithm against the optimal non-anticipating strategy under stationary competition. Our algorithm obtains $\widetilde O(\sqrt T)$-regret if the bids are all revealed at the end of each round. With the restriction that the buyer only sees the winning bid after each round, our modified algorithm obtains $\widetilde O(T^{\frac{7}{12}})$-regret by techniques developed from survival analysis. Our analysis extends to the more general scenario where the buyer has any bounded instantaneous utility function with regrets of the same order.

Zhaohua Chen, Chang Wang, Qian Wang et al.

In today's online advertising markets, a crucial requirement for an advertiser is to control her total expenditure within a time horizon under some budget. Among various budget control methods, throttling has emerged as a popular choice, managing an advertiser's total expenditure by selecting only a subset of auctions to participate in. This paper provides a theoretical panorama of a single advertiser's dynamic budget throttling process in repeated second-price auctions. We first establish a lower bound on the regret and an upper bound on the asymptotic competitive ratio for any throttling algorithm, respectively, when the advertiser's values are stochastic and adversarial. Regarding the algorithmic side, we propose the OGD-CB algorithm, which guarantees a near-optimal expected regret with stochastic values. On the other hand, when values are adversarial, we prove that this algorithm also reaches the upper bound on the asymptotic competitive ratio. We further compare throttling with pacing, another widely adopted budget control method, in repeated second-price auctions. In the stochastic case, we demonstrate that pacing is generally superior to throttling for the advertiser, supporting the well-known result that pacing is asymptotically optimal in this scenario. However, in the adversarial case, we give an exciting result indicating that throttling is also an asymptotically optimal dynamic bidding strategy. Our results bridge the gaps in theoretical research of throttling in repeated auctions and comprehensively reveal the ability of this popular budget-smoothing strategy.

Zhaohua Chen, Rui Ai, Mingwei Yang et al.

We study the framework of a dynamic decision-making scenario with resource constraints. In this framework, an agent, whose target is to maximize the total reward under the initial inventory, selects an action in each round upon observing a random request, leading to a reward and resource consumptions that are further associated with an unknown random external factor. While previous research has already established an $\widetilde{O}(\sqrt{T})$ worst-case regret for this problem, this work offers two results that go beyond the worst-case perspective: one for the worst-case gap between benchmarks and another for logarithmic regret rates. We first show that an $Ω(\sqrt{T})$ distance between the commonly used fluid benchmark and the online optimum is unavoidable when the former has a degenerate optimal solution. On the algorithmic side, we merge the re-solving heuristic with distribution estimation skills and propose an algorithm that achieves an $\widetilde{O}(1)$ regret as long as the fluid LP has a unique and non-degenerate solution. Furthermore, we prove that our algorithm maintains a near-optimal $\widetilde{O}(\sqrt{T})$ regret even in the worst cases and extend these results to the setting where the request and external factor are continuous. Regarding information structure, our regret results are obtained under two feedback models, respectively, where the algorithm accesses the external factor at the end of each round and at the end of a round only when a non-null action is executed.

Shira Guskin, Moshe Wasserblat, Chang Wang et al.

Limited computational budgets often prevent transformers from being used in production and from having their high accuracy utilized. A knowledge distillation approach addresses the computational efficiency by self-distilling BERT into a smaller transformer representation having fewer layers and smaller internal embedding. However, the performance of these models drops as we reduce the number of layers, notably in advanced NLP tasks such as span question answering. In addition, a separate model must be trained for each inference scenario with its distinct computational budget. Dynamic-TinyBERT tackles both limitations by partially implementing the Length Adaptive Transformer (LAT) technique onto TinyBERT, achieving x3 speedup over BERT-base with minimal accuracy loss. In this work, we expand the Dynamic-TinyBERT approach to generate a much more highly efficient model. We use MiniLM distillation jointly with the LAT method, and we further enhance the efficiency by applying low-bit quantization. Our quantized length-adaptive MiniLM model (QuaLA-MiniLM) is trained only once, dynamically fits any inference scenario, and achieves an accuracy-efficiency trade-off superior to any other efficient approaches per any computational budget on the SQuAD1.1 dataset (up to x8.8 speedup with <1% accuracy loss). The code to reproduce this work is publicly available on Github.

Haolin Deng, Chang Wang, Xin Li et al.

Enhancing the attribution in large language models (LLMs) is a crucial task. One feasible approach is to enable LLMs to cite external sources that support their generations. However, existing datasets and evaluation methods in this domain still exhibit notable limitations. In this work, we formulate the task of attributed query-focused summarization (AQFS) and present WebCiteS, a Chinese dataset featuring 7k human-annotated summaries with citations. WebCiteS derives from real-world user queries and web search results, offering a valuable resource for model training and evaluation. Prior works in attribution evaluation do not differentiate between groundedness errors and citation errors. They also fall short in automatically verifying sentences that draw partial support from multiple sources. We tackle these issues by developing detailed metrics and enabling the automatic evaluator to decompose the sentences into sub-claims for fine-grained verification. Our comprehensive evaluation of both open-source and proprietary models on WebCiteS highlights the challenge LLMs face in correctly citing sources, underscoring the necessity for further improvement. The dataset and code will be open-sourced to facilitate further research in this crucial field.

Maolin Wang, Dongze Wu, Jianing Zhou et al.

Large language models (LLMs) have become an important semantic infrastructure for modern recommender systems. A prevailing paradigm integrates LLM-derived semantic embeddings with collaborative representations via representation alignment, implicitly assuming that the two views encode a shared latent entity and that stronger alignment yields better results. We formalize this assumption as the global low-complexity alignment hypothesis and argue that it is stronger than necessary and often structurally mismatched with real-world recommendation settings. We propose a complementary perspective in which semantic and collaborative representations are treated as partially shared yet fundamentally heterogeneous views, each containing both shared and view-specific factors. Under this shared-plus-private latent structure, enforcing global geometric alignment may distort local structure, suppress view-specific signals, and reduce informational diversity. To support this perspective, we develop complementarity-aware diagnostics that quantify overlap, unique-hit contribution, and theoretical fusion upper bounds. Empirical analyses on sparse recommendation benchmarks reveal low item-level agreement between semantic and collaborative views and substantial oracle fusion gains, indicating strong complementarity. Furthermore, controlled alignment probes show that low-capacity mappings capture only shared components and fail to recover full collaborative geometry, especially under distribution shift. These findings suggest that alignment should not be treated as the default integration principle. We advocate a shift from alignment-centric modeling to complementarity fusion-centric, complementarity-aware design, where shared factors are selectively integrated while private signals are preserved. This reframing provides a principled foundation for the next generation of LLM-enhanced recommender systems.

Fengyu She, Nan Wang, Hongfei Wu et al.

Scientific literature is growing exponentially, creating a critical bottleneck for researchers to efficiently synthesize knowledge. While general-purpose Large Language Models (LLMs) show potential in text processing, they often fail to capture scientific domain-specific nuances (e.g., technical jargon, methodological rigor) and struggle with complex scientific tasks, limiting their utility for interdisciplinary research. To address these gaps, this paper presents SciGPT, a domain-adapted foundation model for scientific literature understanding and ScienceBench, an open source benchmark tailored to evaluate scientific LLMs. Built on the Qwen3 architecture, SciGPT incorporates three key innovations: (1) low-cost domain distillation via a two-stage pipeline to balance performance and efficiency; (2) a Sparse Mixture-of-Experts (SMoE) attention mechanism that cuts memory consumption by 55\% for 32,000-token long-document reasoning; and (3) knowledge-aware adaptation integrating domain ontologies to bridge interdisciplinary knowledge gaps. Experimental results on ScienceBench show that SciGPT outperforms GPT-4o in core scientific tasks including sequence labeling, generation, and inference. It also exhibits strong robustness in unseen scientific tasks, validating its potential to facilitate AI-augmented scientific discovery.

Naveed Ur Rehman Junejo, Muhammad Wasim Nawaz, Qingsheng Huang et al.

The ability to accurately predict and analyze student performance in online education, both at the outset and throughout the semester, is vital. Most of the published studies focus on binary classification (Fail or Pass) but there is still a significant research gap in predicting students' performance across multiple categories. This study introduces a novel neural network-based approach capable of accurately predicting student performance and identifying vulnerable students at early stages of the online courses. The Open University Learning Analytics (OULA) dataset is employed to develop and test the proposed model, which predicts outcomes in Distinction, Fail, Pass, and Withdrawn categories. The OULA dataset is preprocessed to extract features from demographic data, assessment data, and clickstream interactions within a Virtual Learning Environment (VLE). Comparative simulations indicate that the proposed model significantly outperforms existing baseline models including Artificial Neural Network Long Short Term Memory (ANN-LSTM), Random Forest (RF) 'gini', RF 'entropy' and Deep Feed Forward Neural Network (DFFNN) in terms of accuracy, precision, recall, and F1-score. The results indicate that the prediction accuracy of the proposed method is about 25% more than the existing state-of-the-art. Furthermore, compared to existing methodologies, the model demonstrates superior predictive capability across temporal course progression, achieving superior accuracy even at the initial 20% phase of course completion.

Chang Wang, Ming Zhu, Shahram Latifi et al.

Congenital Heart Disease (CHD) is the most common neonatal anomaly, highlighting the urgent need for early detection to improve outcomes. Yet, fetal ECG (fECG) signals in abdominal ECG (aECG) are often masked by maternal ECG and noise, challenging conventional methods under low signal-to-noise ratio (SNR) conditions. We propose FetalHealthNet (FHNet), a deep learning framework that integrates Graph Neural Networks with a multi-scale enhanced transformer to dynamically model spatiotemporal inter-lead correlations and extract clean fECG signals. On benchmark aECG datasets, FHNet consistently outperforms long short-term memory (LSTM) models, standard transformers, and state-of-the-art models, achieving R2>0.99 and RMSE = 0.015 even under severe noise. Interpretability analyses highlight physiologically meaningful temporal and lead contributions, supporting model transparency and clinical trust. FHNet illustrates the potential of AI-driven modeling to advance fetal monitoring and enable early CHD screening, underscoring the transformative impact of next-generation biomedical signal processing.

Chang Wang, Siyu Yan, Depeng Yuan et al.

The generation of ad headlines plays a vital role in modern advertising, where both quality and diversity are essential to engage a broad range of audience segments. Current approaches primarily optimize language models for headline quality or click-through rates (CTR), often overlooking the need for diversity and resulting in homogeneous outputs. To address this limitation, we propose DIVER, a novel framework based on large language models (LLMs) that are jointly optimized for both diversity and quality. We first design a semantic- and stylistic-aware data generation pipeline that automatically produces high-quality training pairs with ad content and multiple diverse headlines. To achieve the goal of generating high-quality and diversified ad headlines within a single forward pass, we propose a multi-stage multi-objective optimization framework with supervised fine-tuning (SFT) and reinforcement learning (RL). Experiments on real-world industrial datasets demonstrate that DIVER effectively balances quality and diversity. Deployed on a large-scale content-sharing platform serving hundreds of millions of users, our framework improves advertiser value (ADVV) and CTR by 4.0% and 1.4%.

Yueyang Quan, Chang Wang, Shengjie Zhai et al.

Decentralized min-max optimization allows multi-agent systems to collaboratively solve global min-max optimization problems by facilitating the exchange of model updates among neighboring agents, eliminating the need for a central server. However, sharing model updates in such systems carry a risk of exposing sensitive data to inference attacks, raising significant privacy concerns. To mitigate these privacy risks, differential privacy (DP) has become a widely adopted technique for safeguarding individual data. Despite its advantages, implementing DP in decentralized min-max optimization poses challenges, as the added noise can hinder convergence, particularly in non-convex scenarios with complex agent interactions in min-max optimization problems. In this work, we propose an algorithm called DPMixSGD (Differential Private Minmax Hybrid Stochastic Gradient Descent), a novel privacy-preserving algorithm specifically designed for non-convex decentralized min-max optimization. Our method builds on the state-of-the-art STORM-based algorithm, one of the fastest decentralized min-max solutions. We rigorously prove that the noise added to local gradients does not significantly compromise convergence performance, and we provide theoretical bounds to ensure privacy guarantees. To validate our theoretical findings, we conduct extensive experiments across various tasks and models, demonstrating the effectiveness of our approach.

Yubin Zhang, Yanhua Huang, Haiming Xu et al.

Multimodal recommendation has emerged as a critical technique in modern recommender systems, leveraging content representations from advanced multimodal large language models (MLLMs). To ensure these representations are well-adapted, alignment with the recommender system is essential. However, evaluating the alignment of MLLMs for recommendation presents significant challenges due to three key issues: (1) static benchmarks are inaccurate because of the dynamism in real-world applications, (2) evaluations with online system, while accurate, are prohibitively expensive at scale, and (3) conventional metrics fail to provide actionable insights when learned representations underperform. To address these challenges, we propose the Leakage Impact Score (LIS), a novel metric for multimodal recommendation. Rather than directly assessing MLLMs, LIS efficiently measures the upper bound of preference data. We also share practical insights on deploying MLLMs with LIS in real-world scenarios. Online A/B tests on both Content Feed and Display Ads of Xiaohongshu's Explore Feed production demonstrate the effectiveness of our proposed method, showing significant improvements in user spent time and advertiser value.

Mengyun Wang, Bo Wang, Yifeng Niu et al.

Accurate dynamics modeling is essential for quadrotors to achieve precise trajectory tracking in various applications. Traditional physical knowledge-driven modeling methods face substantial limitations in unknown environments characterized by variable payloads, wind disturbances, and external perturbations. On the other hand, data-driven modeling methods suffer from poor generalization when handling out-of-distribution (OoD) data, restricting their effectiveness in unknown scenarios. To address these challenges, we introduce the Physics-Informed Wind-Adaptive Network (PI-WAN), which combines knowledge-driven and data-driven modeling methods by embedding physical constraints directly into the training process for robust quadrotor dynamics learning. Specifically, PI-WAN employs a Temporal Convolutional Network (TCN) architecture that efficiently captures temporal dependencies from historical flight data, while a physics-informed loss function applies physical principles to improve model generalization and robustness across previously unseen conditions. By incorporating real-time prediction results into a model predictive control (MPC) framework, we achieve improvements in closed-loop tracking performance. Comprehensive simulations and real-world flight experiments demonstrate that our approach outperforms baseline methods in terms of prediction accuracy, tracking precision, and robustness to unknown environments.

Chao Yan, Xiaojia Xiang, Chang Wang et al.

Developing the flocking behavior for a dynamic squad of fixed-wing UAVs is still a challenge due to kinematic complexity and environmental uncertainty. In this paper, we deal with the decentralized flocking and collision avoidance problem through deep reinforcement learning (DRL). Specifically, we formulate a decentralized DRL-based decision making framework from the perspective of every follower, where a collision avoidance mechanism is integrated into the flocking controller. Then, we propose a novel reinforcement learning algorithm PS-CACER for training a shared control policy for all the followers. Besides, we design a plug-n-play embedding module based on convolutional neural networks and the attention mechanism. As a result, the variable-length system state can be encoded into a fixed-length embedding vector, which makes the learned DRL policy independent with the number and the order of followers. Finally, numerical simulation results demonstrate the effectiveness of the proposed method, and the learned policies can be directly transferred to semi-physical simulation without any parameter finetuning.

Chang Wang, Jian Liang, Mingkai Huang et al.

We present HDP-VFL, the first hybrid differentially private (DP) framework for vertical federated learning (VFL) to demonstrate that it is possible to jointly learn a generalized linear model (GLM) from vertically partitioned data with only a negligible cost, w.r.t. training time, accuracy, etc., comparing to idealized non-private VFL. Our work builds on the recent advances in VFL-based collaborative training among different organizations which rely on protocols like Homomorphic Encryption (HE) and Secure Multi-Party Computation (MPC) to secure computation and training. In particular, we analyze how VFL's intermediate result (IR) can leak private information of the training data during communication and design a DP-based privacy-preserving algorithm to ensure the data confidentiality of VFL participants. We mathematically prove that our algorithm not only provides utility guarantees for VFL, but also offers multi-level privacy, i.e. DP w.r.t. IR and joint differential privacy (JDP) w.r.t. model weights. Experimental results demonstrate that our work, under adequate privacy budgets, is quantitatively and qualitatively similar to GLMs, learned in idealized non-private VFL setting, rather than the increased cost in memory and processing time in most prior works based on HE or MPC. Our codes will be released if this paper is accepted.

Mingkai Huang, Hao Li, Bing Bai et al.

Privacy-preserving recommendations are recently gaining momentum, since the decentralized user data is increasingly harder to collect, by recommendation service providers, due to the serious concerns over user privacy and data security. This situation is further exacerbated by the strict government regulations such as Europe's General Data Privacy Regulations(GDPR). Federated Learning(FL) is a newly developed privacy-preserving machine learning paradigm to bridge data repositories without compromising data security and privacy. Thus many federated recommendation(FedRec) algorithms have been proposed to realize personalized privacy-preserving recommendations. However, existing FedRec algorithms, mostly extended from traditional collaborative filtering(CF) method, cannot address cold-start problem well. In addition, their performance overhead w.r.t. model accuracy, trained in a federated setting, is often non-negligible comparing to centralized recommendations. This paper studies this issue and presents FL-MV-DSSM, a generic content-based federated multi-view recommendation framework that not only addresses the cold-start problem, but also significantly boosts the recommendation performance by learning a federated model from multiple data source for capturing richer user-level features. The new federated multi-view setting, proposed by FL-MV-DSSM, opens new usage models and brings in new security challenges to FL in recommendation scenarios. We prove the security guarantees of \xxx, and empirical evaluations on FL-MV-DSSM and its variations with public datasets demonstrate its effectiveness. Our codes will be released if this paper is accepted.

Hao Chen, Chang Wang, Jian Huang et al.

In Markov games, playing against non-stationary opponents with learning ability is still challenging for reinforcement learning (RL) agents, because the opponents can evolve their policies concurrently. This increases the complexity of the learning task and slows down the learning speed of the RL agents. This paper proposes efficient use of rough heuristics to speed up policy learning when playing against concurrent learners. Specifically, we propose an algorithm that can efficiently learn explainable and generalized action selection rules by taking advantages of the representation of quantitative heuristics and an opponent model with an eXtended classifier system (XCS) in zero-sum Markov games. A neural network is used to model the opponent from their behaviors and the corresponding policy is inferred for action selection and rule evolution. In cases of multiple heuristic policies, we introduce the concept of Pareto optimality for action selection. Besides, taking advantages of the condition representation and matching mechanism of XCS, the heuristic policies and the opponent model can provide guidance for situations with similar feature representation. Furthermore, we introduce an accuracy-based eligibility trace mechanism to speed up rule evolution, i.e., classifiers that can match the historical traces are reinforced according to their accuracy. We demonstrate the advantages of the proposed algorithm over several benchmark algorithms in a soccer and a thief-and-hunter scenarios.

Tiantian Ji, Yue Wu, Chang Wang et al.

With the success of the Cyber Grand Challenge (CGC) sponsored by DARPA, the topic of Autonomous Cyber Reasoning System (CRS) has recently attracted extensive attention from both industry and academia. Utilizing automated system to detect, exploit and patch software vulnerabilities seems so attractive because of its scalability and cost-efficiency compared with the human expert based solution. In this paper, we give an extensive survey of former representative works related to the underlying technologies of a CRS, including vulnerability detection, exploitation and patching. As an important supplement, we then review several pioneer studies that explore the potential of machine learning technologies in this field, and point out that the future development of Autonomous CRS is inseparable from machine learning.

Haiwen Huang, Chang Wang, Bin Dong

First-order optimization algorithms have been proven prominent in deep learning. In particular, algorithms such as RMSProp and Adam are extremely popular. However, recent works have pointed out the lack of ``long-term memory" in Adam-like algorithms, which could hamper their performance and lead to divergence. In our study, we observe that there are benefits of weighting more of the past gradients when designing the adaptive learning rate. We therefore propose an algorithm called the Nostalgic Adam (NosAdam) with theoretically guaranteed convergence at the best known convergence rate. NosAdam can be regarded as a fix to the non-convergence issue of Adam in alternative to the recent work of [Reddi et al., 2018]. Our preliminary numerical experiments show that NosAdam is a promising alternative algorithm to Adam. The proofs, code and other supplementary materials can be found in an anonymously shared link.

Chang Wang, Liangliang Cao, Bowen Zhou

In this paper, we present a novel approach for medical synonym extraction. We aim to integrate the term embedding with the medical domain knowledge for healthcare applications. One advantage of our method is that it is very scalable. Experiments on a dataset with more than 1M term pairs show that the proposed approach outperforms the baseline approaches by a large margin.

Liangliang Cao, Chang Wang

Synonym extraction is an important task in natural language processing and often used as a submodule in query expansion, question answering and other applications. Automatic synonym extractor is highly preferred for large scale applications. Previous studies in synonym extraction are most limited to small scale datasets. In this paper, we build a large dataset with 3.4 million synonym/non-synonym pairs to capture the challenges in real world scenarios. We proposed (1) a new cost function to accommodate the unbalanced learning problem, and (2) a feature learning based deep neural network to model the complicated relationships in synonym pairs. We compare several different approaches based on SVMs and neural networks, and find out a novel feature learning based neural network outperforms the methods with hand-assigned features. Specifically, the best performance of our model surpasses the SVM baseline with a significant 97\% relative improvement.