arXiv subjects

Hao Wu

Publications and source records attributed to Hao Wu.

At least 343 records · Page 19Linked to original sources

arXiv2025

An Efficient Alternating Minimization Algorithm for Computing Quantum Rate-Distortion Function

We consider the computation of the entanglement-assisted quantum rate-distortion function, which plays a central role in quantum information theory. We propose an efficient alternating minimization algorithm based on the Lagrangian analysis. Instead of fixing the multiplier corresponding to the distortion constraint, we update the multiplier in each iteration. Hence the algorithm solves the original problem itself, rather than the Lagrangian relaxation of it. Moreover, all the other variables are iterated in closed form without solving multi-dimensional nonlinear equations or multivariate optimization problems. Numerical experiments show the accuracy of our proposed algorithm and its improved efficiency over existing methods.

cs.IT↗

arXiv2025

Step-3 is Large yet Affordable: Model-system Co-design for Cost-effective Decoding

Large language models (LLMs) face low hardware efficiency during decoding, especially for long-context reasoning tasks. This paper introduces Step-3, a 321B-parameter VLM with hardware-aware model-system co-design optimized for minimizing decoding costs. Step-3 innovates in two key dimensions: (1) A novel Multi-Matrix Factorization Attention (MFA) mechanism that significantly reduces both KV cache size and computation while maintaining high attention expressiveness, and (2) Attention-FFN Disaggregation (AFD), a distributed inference system that decouples attention and Feed-Forward Network (FFN) layers into specialized subsystems. This co-design achieves unprecedented cost efficiency: Step-3 significantly reduces theoretical decoding costs compared with models like DeepSeek-V3 and Qwen3 MoE 235B, with the gains widening at longer context. Step-3 achieves low cost while activating 38B parameters per token (more than DeepSeek-V3 and Qwen3 MoE 235B), demonstrating that hardware-aligned attention arithmetic intensity, MoE sparsity, and AFD are critical to cost-effectiveness. We perform a head-to-head comparison with DeepSeek-V3 in its favorable scenarios. Our implementation on Hopper GPUs achieves a decoding throughput of up to 4,039 tokens per second per GPU under 50ms TPOT SLA (4K context, FP8, no MTP). It is higher than DeepSeek-V3's 2,324 in the same setup and sets a new Pareto frontier for LLM decoding.

cs.LG↗

arXiv2025

RDD Function: A Tradeoff Between Rate and Distortion-in-Distortion

In this paper, we propose a novel function named Rate Distortion-in-Distortion (RDD) function as an extension of the classical rate-distortion (RD) function, where the expected distortion constraint is replaced by a Gromov-type distortion. This distortion, integral to the Gromov-Wasserstein (GW) distance, effectively defines the similarity in spaces of possibly different dimensions even without a direct metric between them. While the RDD function qualifies as an informational RD function, encoding theorems substantiate its status as an operational RD function, thereby underscoring its potential applicability in real-world source coding. Due to the high computational complexity associated with Gromov-type distortion, in general, the RDD function cannot be evaluated analytically. Consequently, we develop an alternating mirror descent algorithm that significantly reduces computational complexity by employing decomposition, linearization, and relaxation techniques. Numerical results on classical sources and different grids demonstrate the effectiveness of the developed algorithm. By exploring the relationship between the RDD function and the RD function, we suggest that the RDD function may have potential applications in future scenarios.

cs.IT↗

arXiv2025

When LLMs Copy to Think: Uncovering Copy-Guided Attacks in Reasoning LLMs

Large Language Models (LLMs) have become integral to automated code analysis, enabling tasks such as vulnerability detection and code comprehension. However, their integration introduces novel attack surfaces. In this paper, we identify and investigate a new class of prompt-based attacks, termed Copy-Guided Attacks (CGA), which exploit the inherent copying tendencies of reasoning-capable LLMs. By injecting carefully crafted triggers into external code snippets, adversaries can induce the model to replicate malicious content during inference. This behavior enables two classes of vulnerabilities: inference length manipulation, where the model generates abnormally short or excessively long reasoning traces; and inference result manipulation, where the model produces misleading or incorrect conclusions. We formalize CGA as an optimization problem and propose a gradient-based approach to synthesize effective triggers. Empirical evaluation on state-of-the-art reasoning LLMs shows that CGA reliably induces infinite loops, premature termination, false refusals, and semantic distortions in code analysis tasks. While highly effective in targeted settings, we observe challenges in generalizing CGA across diverse prompts due to computational constraints, posing an open question for future research. Our findings expose a critical yet underexplored vulnerability in LLM-powered development pipelines and call for urgent advances in prompt-level defense mechanisms.

cs.CR↗

arXiv2025

The LDP of McKean-Vlasov stochastic differential equations with Hölder continuous conditions and integrable conditions

In this paper, we first study the large deviation principle (LDP) for non-degenerate McKean-Vlasov stochastic differential equations (MVSDEs) with Hölder continuous drifts by using Zvonkin's transformation. When the drift only satisfies Hölder condition, the skeleton equation may have multiple solutions. Among these solutions, we find one that ensures the MVSDEs satisfy the LDP. Moreover, we introduce a new definition for the rate function that reduces to traditional rate function if the drift satisfies the Lipschitz condition. Secondly, we study the LDP for degenerate MVSDEs with Hölder continuous drifts.

math.PR↗

arXiv2025

Frequency-Aligned Knowledge Distillation for Lightweight Spatiotemporal Forecasting

Spatiotemporal forecasting tasks, such as traffic flow, combustion dynamics, and weather forecasting, often require complex models that suffer from low training efficiency and high memory consumption. This paper proposes a lightweight framework, Spectral Decoupled Knowledge Distillation (termed SDKD), which transfers the multi-scale spatiotemporal representations from a complex teacher model to a more efficient lightweight student network. The teacher model follows an encoder-latent evolution-decoder architecture, where its latent evolution module decouples high-frequency details and low-frequency trends using convolution and Transformer (global low-frequency modeler). However, the multi-layer convolution and deconvolution structures result in slow training and high memory usage. To address these issues, we propose a frequency-aligned knowledge distillation strategy, which extracts multi-scale spectral features from the teacher's latent space, including both high and low frequency components, to guide the lightweight student model in capturing both local fine-grained variations and global evolution patterns. Experimental results show that SDKD significantly improves performance, achieving reductions of up to 81.3% in MSE and in MAE 52.3% on the Navier-Stokes equation dataset. The framework effectively captures both high-frequency variations and long-term trends while reducing computational complexity. Our codes are available at https://github.com/itsnotacie/SDKD

cs.LG↗

arXiv2025

Quantizing Text-attributed Graphs for Semantic-Structural Integration

Text-attributed graphs (TAGs) have emerged as a powerful representation for modeling complex relationships across diverse domains. With the rise of large language models (LLMs), there is growing interest in leveraging their capabilities for graph learning. However, current approaches face significant challenges in embedding structural information into LLM-compatible formats, requiring either computationally expensive alignment mechanisms or manual graph verbalization techniques that often lose critical structural details. Moreover, these methods typically require labeled data from source domains for effective transfer learning, significantly constraining their adaptability. We propose STAG, a novel self-supervised framework that directly quantizes graph structural information into discrete tokens using a frozen codebook. Unlike traditional quantization approaches, our method employs soft assignment and KL divergence guided quantization to address the unique challenges of graph data, which lacks natural tokenization structures. Our framework enables both LLM-based and traditional learning approaches, supporting true zero-shot transfer learning without requiring labeled data even in the source domain. Extensive experiments demonstrate state-of-the-art performance across multiple node classification benchmarks while maintaining compatibility with different LLM architectures, offering an elegant solution to bridging graph learning with LLMs.

cs.LG↗

arXiv2025

Unsupervised super-spatial-resolution Brillouin frequency shift extraction based on physical enhanced spatial resolution neural network

Spatial resolution (SR), a core parameter of Brillouin optical time-domain analysis (BOTDA) sensors, determines the minimum fiber length over which physical perturbations can be accurately detected. However, the phonon lifetime in the fiber imposes an inherent limit on the SR, making sub-meter-level SR challenging in high-SR monitoring scenarios. Conventional SR enhancement approaches, constrained by hardware limitations, often involve complex systems, or increased measurement times. Although traditional deconvolution methods can mitigate hardware constraints, they suffer from distortion due to the nonlinear nature of the BOTDA response. Supervised deep learning approaches have recently emerged as an alternative, offering faster and more accurate post-processing through data-driven models. However, the need for extensive labeled data and the lack of physical priors lead to high computational costs and limited generalization. To overcome these challenges, we propose an unsupervised deep learning deconvolution framework, Physics-enhanced SR deep neural network (PSRN) guided by an approximate convolution model of the Brillouin gain spectrum (BGS).

physics.optics↗

arXiv2025

Performance Bounds of Ranging Precision in SPAD-Based dToF LiDAR

Lidar with direct time-of-flight (dToF) technology based on single-photon avalanche diode detectors (SPAD) has been widely adopted in various applications. However, a comprehensive theoretical understanding of its fundamental ranging performance limits--particularly the degradation caused by pile-up effects due to system dead time and the potential benefits of photon-number-resolving architectures--remains incomplete. In this work, the Cramer-Rao lower bound (CRLB) for dToF systems is theoretically derived accounting for dead time effects, generalized to SPAD detectors with photon-number-resolving capabilities, and are further validated through Monte Carlo simulations and maximum likelihood estimation. Our results reveal that pile-up not only reduces the information contained within individual ToF but also introduces statistical coupling between distance and photon flux rate, further degrading ranging precision. The derived CRLB is used to determine the optimal optical photon flux, laser pulse width, and ToF quantization resolution that yield the best achievable ranging precision. The analysis further quantifies the limited performance improvement enabled by increased photon-number resolution, which exhibits rapidly diminishing returns. These findings provide theoretical guidance for the design of dToF systems and the selection of their optimal operating points.

physics.optics↗

arXiv2025

ILT-Iterative LoRA Training through Focus-Feedback-Fix for Multilingual Speech Recognition

The deep integration of large language models and automatic speech recognition systems has become a promising research direction with high practical value. To address the overfitting issue commonly observed in Low-Rank Adaptation (LoRA) during the supervised fine-tuning (SFT) stage, this work proposes an innovative training paradigm Iterative LoRA Training (ILT) in combination with an Iterative Pseudo Labeling strategy, effectively enhancing the theoretical upper bound of model performance. Based on Whisper-large-v3 and Qwen2-Audio, we conduct systematic experiments using a three-stage training process: Focus Training, Feed Back Training, and Fix Training. Experimental results demonstrate the effectiveness of the proposed method. Furthermore, the MegaAIS research team applied this technique in the Interspeech 2025 Multilingual Conversational Speech Language Modeling Challenge (MLC-SLM), achieving 4th in Track 1 (Multilingual ASR Task) and 1st place in Track 2 (Speech Separation and Recognition Task), showcasing the practical feasibility and strong application potential of our approach.

cs.CL↗

arXiv2025

Mixing and CLT for Hénon-Sibony maps: plurisubharmonic observables

Let $f$ be a complex Hénon map and $μ$ its unique measure of maximal entropy. We prove that $μ$ is exponentially mixing of all orders for all (not necessarily bounded) plurisubharmonic observables, and that all plurisubharmonic functions satisfy the central limit theorem with respect to $μ$. Our results hold more generally for every Hénon-Sibony map on $\mathbb{C}^k$.

math.CV↗

arXiv2025

Exponential mixing of all orders on Kähler manifolds: (quasi-)plurisubharmonic observables

Let $f$ be a holomorphic automorphism of a compact Kähler manifold with simple action on cohomology and $μ$ its unique measure of maximal entropy. We prove that $μ$ is exponentially mixing of all orders for all d.s.h.\ observables, i.e., functions that are locally differences of plurisubharmonic functions. As a consequence, every d.s.h.\ observable satisfies the central limit theorem with respect to $μ$.

math.CV↗

arXiv2025

RPHunter: Unveiling Rug Pull Schemes in Crypto Token via Code-and-Transaction Fusion Analysis

Rug pull scams have emerged as a persistent threat to cryptocurrency, causing significant financial losses. A typical scenario involves scammers deploying honeypot contracts to attract investments, restricting token sales, and draining the funds, which leaves investors with worthless tokens. Current methods either rely on predefined patterns to detect code risks or utilize statistical transaction data to train detection models. However, real-world Rug Pull schemes often involve a complex interplay between malicious code and suspicious transaction behaviors. These methods, which solely focus on one aspect, fall short in detecting such schemes effectively. In this paper, we propose RPHunter, a novel technique that integrates code and transaction for Rug Pull detection. First, RPHunter establishes declarative rules and performs flow analysis to extract code risk information, further constructing a semantic risk code graph (SRCG). Meanwhile, to leverage transaction information, RPHunter formulates dynamic token transaction activities as a token flow behavior graph (TFBG) in which nodes and edges are characterized from network structure and market manipulation perspectives. Finally, RPHunter employs graph neural networks to extract complementary features from SRCG and TFBG, integrating them through an attention fusion model to enhance the detection of Rug Pull. We manually analyzed 645 Rug Pull incidents from code and transaction aspects and constructed a ground-truth dataset. We evaluated RPHunter on our dataset, achieving a precision of 95.3%, a recall of 93.8% and an F1 score of 94.5%, which highlights superior performance compared to existing methods. Furthermore, when applied to the real-world scenarios, RPHunter has identified 4801 Rug Pull tokens, achieving a precision of 90.7%.

cs.SE↗

arXiv2025

From General Relation Patterns to Task-Specific Decision-Making in Continual Multi-Agent Coordination

Continual Multi-Agent Reinforcement Learning (Co-MARL) requires agents to address catastrophic forgetting issues while learning new coordination policies with the dynamics team. In this paper, we delve into the core of Co-MARL, namely Relation Patterns, which refer to agents' general understanding of interactions. In addition to generality, relation patterns exhibit task-specificity when mapped to different action spaces. To this end, we propose a novel method called General Relation Patterns-Guided Task-Specific Decision-Maker (RPG). In RPG, agents extract relation patterns from dynamic observation spaces using a relation capturer. These task-agnostic relation patterns are then mapped to different action spaces via a task-specific decision-maker generated by a conditional hypernetwork. To combat forgetting, we further introduce regularization items on both the relation capturer and the conditional hypernetwork. Results on SMAC and LBF demonstrate that RPG effectively prevents catastrophic forgetting when learning new tasks and achieves zero-shot generalization to unseen tasks.

cs.MA↗

arXiv2025

Zero-Knowledge Verifiable Graph Query Evaluation via Expansion-Centric Operator Decomposition

This paper investigates the feasibility of achieving zero-knowledge verifiability for graph databases, enabling database owners to cryptographically prove the query execution correctness without disclosing the underlying data. Although similar capabilities have been explored for relational databases, their implementation for graph databases presents unique challenges. This is mainly attributed to the relatively large complexity of queries in graph databases. When translating graph queries into arithmetic circuits, the circuit scale can be too large to be practically evaluated. To address this issue, we propose to break down graph queries into more fine-grained, primitive operators, enabling a step-by-step evaluation through smaller-scale circuits. Accordingly, the verification with ZKP circuits of complex graph queries can be decomposed into a series of composable cryptographic primitives, each designed to verify a fundamental structural property such as path ordering or edge directionality. Especially, having noticed that the graph expansion (i.e., traversing from nodes to their neighbors along edges) operation serves as the backbone of graph query evaluation, we design the expansion centric operator decomposition. In addition to constructing circuits for the expansion primitives, we also design specialized ZKP circuits for the various attributes that augment this traversal. The circuits are meticulously designed to take advantage of PLONKish arithmetization. By integrating these optimized circuits, we implement ZKGraph, a system that provides verifiable query processing while preserving data privacy. Performance evaluation indicates that ZKGraph significantly outperforms naive in circuit implementations of graph operators, achieving substantial improvements in both runtime and memory consumption.

cs.DB↗

arXiv2025

GaussMaster: An LLM-based Database Copilot System

In the financial industry, data is the lifeblood of operations, and DBAs shoulder significant responsibilities for SQL tuning, database deployment, diagnosis, and service repair. In recent years, both database vendors and customers have increasingly turned to autonomous database platforms in an effort to alleviate the heavy workload of DBAs. However, existing autonomous database platforms are limited in their capabilities, primarily addressing single-point issues such as NL2SQL, anomaly detection, and SQL tuning. Manual intervention remains a necessity for comprehensive database maintenance. GaussMaster aims to revolutionize this landscape by introducing an LLM-based database copilot system. This innovative solution is designed not only to assist developers in writing efficient SQL queries but also to provide comprehensive care for database services. When database instances exhibit abnormal behavior, GaussMaster is capable of orchestrating the entire maintenance process automatically. It achieves this by analyzing hundreds of metrics and logs, employing a Tree-of-thought approach to identify root causes, and invoking appropriate tools to resolve issues. We have successfully implemented GaussMaster in real-world scenarios, such as the banking industry, where it has achieved zero human intervention for over 34 database maintenance scenarios. In this paper, we present significant improvements in these tasks with code at https://gitcode.com/opengauss/openGauss-GaussMaster.

cs.DB↗

arXiv2025

Super-resolution of two Closely-spaced Electromagnetic Fields via Walsh-Modulated Dynamical Decoupling Spectroscopy

Due to quantum fluctuations, non-orthogonal quantum states cannot be distinguished with complete certainty, making their underlying physical parameters difficult to resolve. Traditionally, it has been believed that the linewidth of a system behaves like these quantum fluctuations to set the ultimate limit on frequency resolution as two oscillating electromagnetic fields are applied. Consequently, the measurement time required to resolve a frequency difference $Δω$ was assumed to diverge as $Δω\rightarrow 0$. Here, we show that linewidth does not play a defining role in resolving two closely spaced frequencies. Instead, the ultimate limit is set by parameter-independent quantum fluctuations, such as shot noise in our case. We propose and experimentally demonstrate the first general broadband protocol for super-resolution spectroscopy. Specifically, our protocol uses a Walsh-modulated dynamical decoupling (WMDD) sequence to encode $Δω$ between two unknown tones into a quantum state. This leverages phase information to suppress parameter-independent shot noise, thereby enhancing the signal-to-noise ratio and enabling super-resolution spectroscopy. With this approach, we resolve two randomly chosen oscillating electric fields of order 100 MHz separated by 5 Hz, with a measured frequency difference of 5.0(1.6) Hz using a measurement time per run of just 1 ms, representing an improvement of 200 beyond the traditional resolution limit. As such, our technique accelerates data acquisition by more than $10^5$ magnitude compared to conventional methods. Crucially, as our protocol is rooted in the motional Raman (quantum vector signal analyzer) framework, it is effective across an arbitrary frequency range and thus promises to enhance broadband sensing of electromagnetic fields and improve spectral efficiency of next-generation communication systems.

physics.atom-ph↗

arXiv2025

Characterization of Rydberg-Atom Signal Reception of Dual-Frequency Signals Coupled with Two Energy Levels

Rydberg atomic sensors have been adopted for novel radio frequency (RF) measurement technique and the sensing capability for signals in multiple frequencies makes it attractive for multi-user communication. However, unlike traditional antennas where the signals in multiple frequencies are orthogonal, the received signals of atomic sensors corresponding to different energy levels will be downconverted to the baseband simultaneously, resulting in multi-user interference. Thus, in this paper, we analyze the mutual interference characteristics of two RF signals with different carrier frequencies coupling different energy levels. We introduce the joint response coefficient based on the receiver characteristics and analyze the interference of one user to another. We analyze the bit-error rate (BER) and symbol-error rate (SER) for two signals coupling two different energy levels. We also conduct experiments to validate the BER and SER results.

eess.SP↗