arXiv subjects

Wenxuan Zeng

Publications and source records attributed to Wenxuan Zeng.

18 recordsLinked to original sources

arXiv2026

OptiPrime: Optimizing Private Inference through Protocol-Hardware Co-design

Private deep neural network (DNN) inference based on hybrid homomorphic encryption (HE) and multi-party computation (MPC) can protect user data with a formal guarantee, but at the cost of significant latency overhead due to HE. Customized HE accelerators have been proposed and have achieved orders-of-magnitude speedup for individual HE operations. However, when directly applying a commercial HE accelerator to state-of-the-art HE-MPC frameworks, we observe only limited end-to-end performance gain. This is because HE-MPC frameworks often require wireless transmission of input and output ciphertexts for each HE operation, leading to a severe network communication bottleneck. To overcome this challenge, we introduce OptiPrime, a protocol-hardware co-optimization framework for efficient private DNN inference. OptiPrime features a novel HE protocol for convolutions that substantially reduces the number of transmitted output ciphertexts and mitigates the network communication bottleneck. Meanwhile, as the new protocol introduces complex computation for fewer output ciphertext, we observe new memory access challenges due to a high volume of weight plaintexts and intermediate ciphertexts. Hence, we further propose a lightweight compression system for the weight plaintexts, reducing memory traffic by 10 times, as well as a specialized dataflow to maximize on-chip data reuse of intermediate ciphertexts. Extensive experiments show that our framework outperforms the Cheetah baseline by at most 5.7 times on CPUs and 4.2 times with an accelerator.

cs.AR↗

arXiv2026

Attention Sink Forges Native MoE in Attention Layers: Sink-Aware Training to Address Head Collapse

Large Language Models (LLMs) often assign disproportionate attention to the first token, a phenomenon known as the attention sink. Several recent approaches aim to address this issue, including Sink Attention in GPT-OSS and Gated Attention in Qwen3-Next. However, a comprehensive analysis of the relationship among these attention mechanisms is lacking. In this work, we provide both theoretical and empirical evidence demonstrating that the sink in Vanilla Attention and Sink Attention naturally construct a Mixture-of-Experts (MoE) mechanism within attention layers. This insight explains the head collapse phenomenon observed in prior work, where only a fixed subset of attention heads contributes to generation. To mitigate head collapse, we propose a sink-aware training algorithm with an auxiliary load balancing loss designed for attention layers. Extensive experiments show that our method achieves effective head load balancing and improves model performance across Vanilla Attention, Sink Attention, and Gated Attention. We hope this study offers a new perspective on attention mechanisms and encourages further exploration of the inherent MoE structure within attention layers.

cs.CL↗

arXiv2026

UFO: Unlocking Ultra-Efficient Quantized Private Inference with Protocol and Algorithm Co-Optimization

Private convolutional neural network (CNN) inference based on secure two-party computation (2PC) suffers from high communication and latency overhead, especially from convolution layers. In this paper, we propose UFO, a quantized 2PC inference framework that jointly optimizes the 2PC protocols and quantization algorithm. UFO features a novel 2PC protocol that systematically combines the efficient Winograd convolution algorithm with quantization to improve inference efficiency. However, we observe that naively combining quantization and Winograd convolution faces the following challenges: 1) From the inference perspective, Winograd transformations introduce extensive additions and require frequent bit width conversions to avoid inference overflow, leading to non-negligible communication overhead; 2) From the training perspective, Winograd transformations introduce weight outliers that make quantization-aware training (QAT) difficult, resulting in inferior model accuracy. To address these challenges, we co-optimize both protocol and algorithm. 1) At the protocol level, we propose a series of graph-level optimizations for 2PC inference to minimize the communication. 2) At the algorithm level, we develop a mixed-precision QAT algorithm based on layer sensitivity to optimize model accuracy given communication constraints. To accommodate the outliers, we further introduce a 2PC-friendly bit re-weighting algorithm to increase the representation range without explicitly increasing bit widths. With extensive experiments, UFO demonstrates 11.7x, 3.6x, and 6.3x communication reduction with 1.29%, 1.16%, and 1.29% higher accuracy compared to state-of-the-art frameworks SiRNN, COINN, and CoPriv, respectively.

cs.CR↗

arXiv2026

P1-VL: Bridging Visual Perception and Scientific Reasoning in Physics Olympiads

The transition from symbolic manipulation to science-grade reasoning represents a pivotal frontier for Large Language Models (LLMs), with physics serving as the critical test anchor for binding abstract logic to physical reality. Physics demands that a model maintain physical consistency with the laws governing the universe, a task that fundamentally requires multimodal perception to ground abstract logic in reality. At the Olympiad level, diagrams are often constitutive rather than illustrative, containing essential constraints, such as boundary conditions and spatial symmetries, that are absent from the text. To bridge this visual-logical gap, we introduce P1-VL, a family of open-source vision-language models engineered for advanced scientific reasoning. Our method harmonizes Curriculum Reinforcement Learning, which employs progressive difficulty expansion to stabilize post-training, with Agentic Augmentation, enabling iterative self-verification at inference. Evaluated on HiPhO, a rigorous benchmark of 13 exams from 2024-2025, our flagship P1-VL-235B-A22B becomes the first open-source Vision-Language Model (VLM) to secure 12 gold medals and achieves the state-of-the-art performance in the open-source models. Our agent-augmented system achieves the No.2 overall rank globally, trailing only Gemini-3-Pro. Beyond physics, P1-VL demonstrates remarkable scientific reasoning capacity and generalizability, establishing significant leads over base models in STEM benchmarks. By open-sourcing P1-VL, we provide a foundational step toward general-purpose physical intelligence to better align visual perceptions with abstract physical laws for machine scientific discovery.

cs.AI↗

arXiv2025

H2EAL: Hybrid-Bonding Architecture with Hybrid Sparse Attention for Efficient Long-Context LLM Inference

Large language models (LLMs) have demonstrated remarkable proficiency in a wide range of natural language processing applications. However, the high energy and latency overhead induced by the KV cache limits the edge deployment, especially for long contexts. Emerging hybrid bonding (HB) technology has been proposed as a promising alternative to conventional near-memory processing (NMP) architectures, offering improved bandwidth efficiency and lower power consumption while exhibiting characteristics of distributed memory. In this paper, we propose H2EAL, a hybrid bonding-based accelerator with sparse attention algorithm-hardware co-design for efficient LLM inference at the edge. At the algorithm level, we propose a hybrid sparse attention scheme with static and dynamic sparsity for different heads to fully leverage the sparsity with high accuracy. At the hardware level, we co-design the hardware to support hybrid sparse attention and propose memory-compute co-placement to address the distributed memory bottleneck. Since different attention heads exhibit different sparse patterns and the attention structure often mismatches the HB architecture, we further develop a load-balancing scheduler with parallel tiled attention to address workload imbalance and optimize the mapping strategy. Extensive experiments demonstrate H2EAL achieves 5.20~48.21x speedup and 6.22~73.48x energy efficiency improvement over baseline HB implementation, with a negligible average accuracy drop of 0.87% on multiple benchmarks.

cs.PF↗

arXiv2025

Jupiter: Enhancing LLM Data Analysis Capabilities via Notebook and Inference-Time Value-Guided Search

Large language models (LLMs) have shown great promise in automating data science workflows, but existing models still struggle with multi-step reasoning and tool use, which limits their effectiveness on complex data analysis tasks. To address this, we propose a scalable pipeline that extracts high-quality, tool-based data analysis tasks and their executable multi-step solutions from real-world Jupyter notebooks and associated data files. Using this pipeline, we introduce NbQA, a large-scale dataset of standardized task-solution pairs that reflect authentic tool-use patterns in practical data science scenarios. To further enhance multi-step reasoning, we present Jupiter, a framework that formulates data analysis as a search problem and applies Monte Carlo Tree Search (MCTS) to generate diverse solution trajectories for value model learning. During inference, Jupiter combines the value model and node visit counts to efficiently collect executable multi-step plans with minimal search steps. Experimental results show that Qwen2.5-7B and 14B-Instruct models on NbQA solve 77.82% and 86.38% of tasks on InfiAgent-DABench, respectively-matching or surpassing GPT-4o and advanced agent frameworks. Further evaluations demonstrate improved generalization and stronger tool-use reasoning across diverse multi-step reasoning tasks. Code and data are available at https://github.com/microsoft/Jupiter.

cs.AI↗

arXiv2025

MPCache: MPC-Friendly KV Cache Eviction for Efficient Private LLM Inference

Private large language model (LLM) inference based on secure multi-party computation (MPC) achieves formal data privacy protection but suffers from significant latency overhead, especially for long input sequences. While key-value (KV) cache eviction and sparse attention algorithms have been proposed for efficient LLM inference in plaintext, they are not designed for MPC and cannot benefit private LLM inference directly. In this paper, we propose an accurate and MPC-friendly KV cache eviction framework, dubbed MPCache, building on the observation that historical tokens in a long sequence may have different effects on the downstream decoding. Hence, MPCache combines a look-once static eviction algorithm to discard unimportant KV cache and a query-aware dynamic selection algorithm to activate only a small subset of KV cache for attention computation. MPCache further incorporates a series of optimizations for efficient dynamic KV cache selection, including MPC-friendly similarity approximation, hierarchical KV cache clustering, and cross-layer index-sharing strategy. Extensive experiments demonstrate that MPCache consistently outperforms prior-art KV cache eviction baselines across different generation tasks and achieves 1.8 ~ 2.01x and 3.39 ~ 8.37x decoding latency and communication reduction on different sequence lengths, respectively.

cs.CR↗

arXiv2025

Towards Efficient Privacy-Preserving Machine Learning: A Systematic Review from Protocol, Model, and System Perspectives

Privacy-preserving machine learning (PPML) based on cryptographic protocols has emerged as a promising paradigm to protect user data privacy in cloud-based machine learning services. While it achieves formal privacy protection, PPML often incurs significant efficiency and scalability costs due to orders of magnitude overhead compared to the plaintext counterpart. Therefore, there has been a considerable focus on mitigating the efficiency gap for PPML. In this survey, we provide a comprehensive and systematic review of recent PPML studies with a focus on cross-level optimizations. Specifically, we categorize existing papers into protocol level, model level, and system level, and review progress at each level. We also provide qualitative and quantitative comparisons of existing works with technical insights, based on which we discuss future research directions and highlight the necessity of integrating optimizations across protocol, model, and system levels. We hope this survey can provide an overarching understanding of existing approaches and potentially inspire future breakthroughs in the PPML field. As the field is evolving fast, we also provide a public GitHub repository to continuously track the developments, which is available at https://github.com/PKU-SEC-Lab/Awesome-PPML-Papers.

cs.CR↗

arXiv2025

UniCAIM: A Unified CAM/CIM Architecture with Static-Dynamic KV Cache Pruning for Efficient Long-Context LLM Inference

Transformer-based large language models (LLMs) have achieved impressive performance in various natural language processing (NLP) applications. However, the high memory and computation cost induced by the KV cache limits the inference efficiency, especially for long input sequences. Compute-in-memory (CIM)-based accelerators have been proposed for LLM acceleration with KV cache pruning. However, as existing accelerators only support static pruning with a fixed pattern or dynamic pruning with primitive implementations, they suffer from either high accuracy degradation or low efficiency. In this paper, we propose a ferroelectric FET (FeFET)-based unified content addressable memory (CAM) and CIM architecture, dubbed as UniCAIM. UniCAIM features simultaneous support for static and dynamic pruning with 3 computation modes: 1) in the CAM mode, UniCAIM enables approximate similarity measurement in O(1) time for dynamic KV cache pruning with high energy efficiency; 2) in the charge-domain CIM mode, static pruning can be supported based on accumulative similarity score, which is much more flexible compared to fixed patterns; 3) in the current-domain mode, exact attention computation can be conducted with a subset of selected KV cache. We further propose a novel CAM/CIM cell design that leverages the multi-level characteristics of FeFETs for signed multibit storage of the KV cache and in-place attention computation. With extensive experimental results, we demonstrate UniCAIM can reduce the area-energy-delay product (AEDP) by 8.2-831x over the state-ofthe-art CIM-based LLM accelerators at the circuit level, along with high accuracy comparable with dense attention at the application level, showing its great potential for efficient long-context LLM inference.

cs.AR↗

arXiv2025

ASGEA: Exploiting Logic Rules from Align-Subgraphs for Entity Alignment

Entity alignment (EA) aims to identify entities across different knowledge graphs that represent the same real-world objects. Recent embedding-based EA methods have achieved state-of-the-art performance in EA yet faced interpretability challenges as they purely rely on the embedding distance and neglect the logic rules behind a pair of aligned entities. In this paper, we propose the Align-Subgraph Entity Alignment (ASGEA) framework to exploit logic rules from Align-Subgraphs. ASGEA uses anchor links as bridges to construct Align-Subgraphs and spreads along the paths across KGs, which distinguishes it from the embedding-based methods. Furthermore, we design an interpretable Path-based Graph Neural Network, ASGNN, to effectively identify and integrate the logic rules across KGs. We also introduce a node-level multi-modal attention mechanism coupled with multi-modal enriched anchors to augment the Align-Subgraph. Our experimental results demonstrate the superior performance of ASGEA over the existing embedding-based methods in both EA and Multi-Modal EA (MMEA) tasks.

cs.CL↗

arXiv2024

PrivQuant: Communication-Efficient Private Inference with Quantized Network/Protocol Co-Optimization

Private deep neural network (DNN) inference based on secure two-party computation (2PC) enables secure privacy protection for both the server and the client. However, existing secure 2PC frameworks suffer from a high inference latency due to enormous communication. As the communication of both linear and non-linear DNN layers reduces with the bit widths of weight and activation, in this paper, we propose PrivQuant, a framework that jointly optimizes the 2PC-based quantized inference protocols and the network quantization algorithm, enabling communication-efficient private inference. PrivQuant proposes DNN architecture-aware optimizations for the 2PC protocols for communication-intensive quantized operators and conducts graph-level operator fusion for communication reduction. Moreover, PrivQuant also develops a communication-aware mixed precision quantization algorithm to improve inference efficiency while maintaining high accuracy. The network/protocol co-optimization enables PrivQuant to outperform prior-art 2PC frameworks. With extensive experiments, we demonstrate PrivQuant reduces communication by $11\times, 2.5\times \mathrm{and}~ 2.8\times$, which results in $8.7\times, 1.8\times ~ \mathrm{and}~ 2.4\times$ latency reduction compared with SiRNN, COINN, and CoPriv, respectively.

cs.CR↗

arXiv2024

EQO: Exploring Ultra-Efficient Private Inference with Winograd-Based Protocol and Quantization Co-Optimization

Private convolutional neural network (CNN) inference based on secure two-party computation (2PC) suffers from high communication and latency overhead, especially from convolution layers. In this paper, we propose EQO, a quantized 2PC inference framework that jointly optimizes the CNNs and 2PC protocols. EQO features a novel 2PC protocol that combines Winograd transformation with quantization for efficient convolution computation. However, we observe naively combining quantization and Winograd convolution is sub-optimal: Winograd transformations introduce extensive local additions and weight outliers that increase the quantization bit widths and require frequent bit width conversions with non-negligible communication overhead. Therefore, at the protocol level, we propose a series of optimizations for the 2PC inference graph to minimize the communication. At the network level, We develop a sensitivity-based mixed-precision quantization algorithm to optimize network accuracy given communication constraints. We further propose a 2PC-friendly bit re-weighting algorithm to accommodate weight outliers without increasing bit widths. With extensive experiments, EQO demonstrates 11.7x, 3.6x, and 6.3x communication reduction with 1.29%, 1.16%, and 1.29% higher accuracy compared to state-of-the-art frameworks SiRNN, COINN, and CoPriv, respectively.

cs.CR↗

arXiv2024

FaiMA: Feature-aware In-context Learning for Multi-domain Aspect-based Sentiment Analysis

Multi-domain aspect-based sentiment analysis (ABSA) seeks to capture fine-grained sentiment across diverse domains. While existing research narrowly focuses on single-domain applications constrained by methodological limitations and data scarcity, the reality is that sentiment naturally traverses multiple domains. Although large language models (LLMs) offer a promising solution for ABSA, it is difficult to integrate effectively with established techniques, including graph-based models and linguistics, because modifying their internal architecture is not easy. To alleviate this problem, we propose a novel framework, Feature-aware In-context Learning for Multi-domain ABSA (FaiMA). The core insight of FaiMA is to utilize in-context learning (ICL) as a feature-aware mechanism that facilitates adaptive learning in multi-domain ABSA tasks. Specifically, we employ a multi-head graph attention network as a text encoder optimized by heuristic rules for linguistic, domain, and sentiment features. Through contrastive learning, we optimize sentence representations by focusing on these diverse features. Additionally, we construct an efficient indexing mechanism, allowing FaiMA to stably retrieve highly relevant examples across multiple dimensions for any given input. To evaluate the efficacy of FaiMA, we build the first multi-domain ABSA benchmark dataset. Extensive experimental results demonstrate that FaiMA achieves significant performance improvements in multiple domains compared to baselines, increasing F1 by 2.07% on average. Source code and data sets are anonymously available at https://github.com/SupritYoung/FaiMA.

cs.CL↗

arXiv2024

Kuaiji: the First Chinese Accounting Large Language Model

Large Language Models (LLMs) like ChatGPT and GPT-4 have demonstrated impressive proficiency in comprehending and generating natural language. However, they encounter difficulties when tasked with adapting to specialized domains such as accounting. To address this challenge, we introduce Kuaiji, a tailored Accounting Large Language Model. Kuaiji is meticulously fine-tuned using the Baichuan framework, which encompasses continuous pre-training and supervised fine-tuning processes. Supported by CAtAcctQA, a dataset containing large genuine accountant-client dialogues, Kuaiji exhibits exceptional accuracy and response speed. Our contributions encompass the creation of the first Chinese accounting dataset, the establishment of Kuaiji as a leading open-source Chinese accounting LLM, and the validation of its efficacy through real-world accounting scenarios.

cs.CL↗

arXiv2023

BAT: Behavior-Aware Human-Like Trajectory Prediction for Autonomous Driving

The ability to accurately predict the trajectory of surrounding vehicles is a critical hurdle to overcome on the journey to fully autonomous vehicles. To address this challenge, we pioneer a novel behavior-aware trajectory prediction model (BAT) that incorporates insights and findings from traffic psychology, human behavior, and decision-making. Our model consists of behavior-aware, interaction-aware, priority-aware, and position-aware modules that perceive and understand the underlying interactions and account for uncertainty and variability in prediction, enabling higher-level learning and flexibility without rigid categorization of driving behavior. Importantly, this approach eliminates the need for manual labeling in the training process and addresses the challenges of non-continuous behavior labeling and the selection of appropriate time windows. We evaluate BAT's performance across the Next Generation Simulation (NGSIM), Highway Drone (HighD), Roundabout Drone (RounD), and Macao Connected Autonomous Driving (MoCAD) datasets, showcasing its superiority over prevailing state-of-the-art (SOTA) benchmarks in terms of prediction accuracy and efficiency. Remarkably, even when trained on reduced portions of the training data (25%), our model outperforms most of the baselines, demonstrating its robustness and efficiency in predicting vehicle trajectories, and the potential to reduce the amount of data required to train autonomous vehicles, especially in corner cases. In conclusion, the behavior-aware model represents a significant advancement in the development of autonomous vehicles capable of predicting trajectories with the same level of proficiency as human drivers. The project page is available at https://github.com/Petrichor625/BATraj-Behavior-aware-Model.

cs.RO↗

arXiv2023

CoPriv: Network/Protocol Co-Optimization for Communication-Efficient Private Inference

Deep neural network (DNN) inference based on secure 2-party computation (2PC) can offer cryptographically-secure privacy protection but suffers from orders of magnitude latency overhead due to enormous communication. Previous works heavily rely on a proxy metric of ReLU counts to approximate the communication overhead and focus on reducing the ReLUs to improve the communication efficiency. However, we observe these works achieve limited communication reduction for state-of-the-art (SOTA) 2PC protocols due to the ignorance of other linear and non-linear operations, which now contribute to the majority of communication. In this work, we present CoPriv, a framework that jointly optimizes the 2PC inference protocol and the DNN architecture. CoPriv features a new 2PC protocol for convolution based on Winograd transformation and develops DNN-aware optimization to significantly reduce the inference communication. CoPriv further develops a 2PC-aware network optimization algorithm that is compatible with the proposed protocol and simultaneously reduces the communication for all the linear and non-linear operations. We compare CoPriv with the SOTA 2PC protocol, CrypTFlow2, and demonstrate 2.1x communication reduction for both ResNet-18 and ResNet-32 on CIFAR-100. We also compare CoPriv with SOTA network optimization methods, including SNL, MetaPruning, etc. CoPriv achieves 9.98x and 3.88x online and total communication reduction with a higher accuracy compare to SNL, respectively. CoPriv also achieves 3.87x online communication reduction with more than 3% higher accuracy compared to MetaPruning.

cs.CR↗

arXiv2023

MPCViT: Searching for Accurate and Efficient MPC-Friendly Vision Transformer with Heterogeneous Attention

Secure multi-party computation (MPC) enables computation directly on encrypted data and protects both data and model privacy in deep learning inference. However, existing neural network architectures, including Vision Transformers (ViTs), are not designed or optimized for MPC and incur significant latency overhead. We observe Softmax accounts for the major latency bottleneck due to a high communication complexity, but can be selectively replaced or linearized without compromising the model accuracy. Hence, in this paper, we propose an MPC-friendly ViT, dubbed MPCViT, to enable accurate yet efficient ViT inference in MPC. Based on a systematic latency and accuracy evaluation of the Softmax attention and other attention variants, we propose a heterogeneous attention optimization space. We also develop a simple yet effective MPC-aware neural architecture search algorithm for fast Pareto optimization. To further boost the inference efficiency, we propose MPCViT+, to jointly optimize the Softmax attention and other network components, including GeLU, matrix multiplication, etc. With extensive experiments, we demonstrate that MPCViT achieves 1.9%, 1.3% and 3.6% higher accuracy with 6.2x, 2.9x and 1.9x latency reduction compared with baseline ViT, MPCFormer and THE-X on the Tiny-ImageNet dataset, respectively. MPCViT+ further achieves a better Pareto front compared with MPCViT. The code and models for evaluation are available at https://github.com/PKU-SEC-Lab/mpcvit.

cs.CR↗

arXiv2023

Converge to the Truth: Factual Error Correction via Iterative Constrained Editing

Given a possibly false claim sentence, how can we automatically correct it with minimal editing? Existing methods either require a large number of pairs of false and corrected claims for supervised training or do not handle well errors spanning over multiple tokens within an utterance. In this paper, we propose VENCE, a novel method for factual error correction (FEC) with minimal edits. VENCE formulates the FEC problem as iterative sampling editing actions with respect to a target density function. We carefully design the target function with predicted truthfulness scores from an offline trained fact verification model. VENCE samples the most probable editing positions based on back-calculated gradients of the truthfulness score concerning input tokens and the editing actions using a distantly-supervised language model (T5). Experiments on a public dataset show that VENCE improves the well-adopted SARI metric by 5.3 (or a relative improvement of 11.8%) over the previous best distantly-supervised methods.

cs.CL↗