arXiv subjects

Jiajun Zhou

Publications and source records attributed to Jiajun Zhou.

At least 19 recordsLinked to original sources

arXiv2026

MemSecBench: Tracking Agent Memory Poisoning from Persistence to Consequence and Repair

Memory systems allow agents to retain and reuse information from past interactions, but they can also let malicious content persist. A malicious instruction crafted by an attacker may be stored in long-term memory, recalled much later, and quietly shape a real action. Recent benchmarks increasingly examine agent memory security, yet few trace the same malicious semantics across persistence, downstream consequences, and selective repair under diverse memory-backend comparisons. To address this gap, we introduce MemSecBench, a task-grounded benchmark for the lifecycle security of agent memory systems. It contains 310 cases drawn from 48 realistic contexts across code and science, daily life, and office work. Each case follows a controlled Write--Execute--Forget protocol in an isolated runtime under an exact agent configuration, defined by an agent harness, a memory backend, and an LLM backend. Evidence-based adjudication combines a deterministic write check, checkpoint-specific judge-model evaluations, and programmatic gates across seven lifecycle checkpoints. The experimental design spans a 24-configuration matrix of two agent harnesses, four memory backends, and three LLM backends. Across all 24 configurations, malicious memory persists in 84.2% of all cases, and the full Write--Execute chain succeeds in 50.3%. Among successfully poisoned cases, 59.6% complete the full Execute chain, while 56.1% achieve selective repair.Compared with matched Native configurations, the largest absolute differences are 16.1 percentage points for end-to-end attack success and 41.3 percentage points for selective repair. These descriptive contrasts indicate that the evaluated memory system stacks differ in lifecycle security, both in the propagation of malicious memory and in selective repair after successful memory poisoning.

cs.CR↗

arXiv2026

AgentS4D: Benchmarking Runtime Risks across the Execution Lifecycle of LLM-Based Workspace Agents

Large language model (LLM)-based workspace agents execute stateful, multi-step workflows across heterogeneous resources, external tools, and persistent state. Their safety must therefore be assessed from actions, side effects, and state changes throughout execution. Although recent benchmarks have advanced executable safety testing and trajectory-aware verification, they rarely provide a unified account of where risks enter, how they elicit unsafe behavior, which harms they target, and where supporting evidence appears during execution. We introduce AgentS4D, a sandboxed benchmark for lifecycle-wide runtime safety evaluation. Its four-dimensional runtime-safety framework uses six risk-entry sources, six induction strategies, and nine target harms to guide case construction, while seven lifecycle checkpoints organize post-run evidence. AgentS4D contains 328 risk-injected cases. We evaluate all 20 combinations of four harnesses (Hermes, OpenClaw, Claude Code, and Codex) and five LLM backends (GPT-5.5, Gemini 3.1 Pro, DeepSeek-V4-Pro, MiniMax-M3, and Qwen3.7-Plus) on these cases, yielding 6,560 runs. Overall, 4,461 runs (68.0%) trigger prespecified unsafe signals. Across the 20 configurations, the observed safety of an agent system varies with both its harness-LLM pairing and how risk is introduced. Agent systems exhibit markedly different safety behavior when the same induction strategy reaches them through different risk carriers. They also respond differently to the same target harm when it is realized through different carriers and strategies. Moreover, 4,344 runs (66.22% overall) are unsafe yet complete. Thus, task completion cannot establish runtime safety, and testing only one form of a risk can conceal important weaknesses. Evaluations should examine complete agent configurations across diverse risk conditions and retain evidence throughout execution.

cs.SE↗

arXiv2026

Gravitational Lensing Predictions from Wave Simulations of Fuzzy Dark Matter

In the cold dark matter paradigm, ultra-light particles are emerging as strong contenders to conventional massive particles. A unique prediction of dark matter comprising such ultra-light particles, known as fuzzy dark matter (FDM), is the presence of strong density modulations throughout galactic halos due to wave interference, which -- when approximated by a Gaussian random field (GRF) -- have been proposed to account for the inability to reproduce the observed positions (when measured at sufficient precisions) and flux ratios of multiply-lensed images of quasars. Here, we predict for the first time the properties of gravitationally lensed images generated from 3-D density fields obtained by wave simulations that directly evolve the Schr\"odinger--Poisson equations. Using a novel framework to project these evolved density fields along various axes of the 3-D halo, we obtain the distribution of perturbations to the positions of lensed images. As an exacting test, we find that particles of mass $10^{-22}$ eV can reproduce the positions of the quadruply-lensed radio jets in system HS 0810+2554 to a level better than that of either the GRF approximation or, to a greater extent, an NFW best-fit solution, both of which rely on accurately capturing the global 3-D density field of dark matter halos. Our work highlights the importance of wave simulations for making accurate FDM lensing predictions and the potential for high-resolution observations of lensed systems to serve as a direct probe of the nature of dark matter.

astro-ph.CO↗

arXiv2026

AIS: Adaptive Importance Sampling for Quantized RL

Reinforcement learning (RL) for large language models (LLMs) is dominated by the cost of rollout generation, which has motivated the use of low-precision rollouts (e.g., FP8) paired with a BF16 trainer to improve throughput and reduce memory pressure. This introduces a rollout-training mismatch that biases the policy gradient and can cause training to collapse outright on reasoning benchmarks. We show that the mismatch is non-stationary and acts as a double-edged sword: early in training it provides a stochastic exploration bonus, exposing the gradient to trajectories the trainer would otherwise under-sample, but the same perturbation transitions into a destabilizing source of bias as the policy concentrates. To solve this, we propose Adaptive Importance Sampling (AIS), a correction framework that adjusts the strength of its intervention on a per-batch basis. AIS combines three real-time diagnostics, namely weight reliability, divergence severity, and variance amplification, into a single mixing coefficient that interpolates between the uncorrected and fully importance-weighted gradients, suppressing the destabilizing component of the mismatch while preserving its exploratory benefit. We integrate AIS into GRPO and evaluate it on the diffusion-based LLaDA-8B-Instruct and the autoregressive Qwen3-8B and Qwen3.5-9B across mathematical reasoning and planning benchmarks. AIS matches the BF16 baseline on most tasks while retaining the 1.5 to 2.76x rollout speedup of FP8.

stat.ML↗

arXiv2026

Learning How Much to Think: Difficulty-Aware Dynamic MoEs for Graph Node Classification

Mixture-of-Experts (MoE) architectures offer a scalable path for Graph Neural Networks (GNNs) in node classification tasks but typically rely on static and rigid routing strategies that enforce a uniform expert budget or coarse-grained expert toggles on all nodes. This limitation overlooks the varying discriminative difficulty of nodes and leads to under-fitting for hard nodes and redundant computation for easy ones. To resolve this issue, we propose D2MoE, a novel framework that shifts the focus from static expert selection to node-wise expert resource allocation. By using predictive entropy as a real-time proxy for difficulty, D2MoE employs a difficulty-driven top-p routing mechanism to adaptively concentrate expert resources on hard nodes while reducing overhead for easy ones, achieving continuous and fine-grained expert budget scaling for node classification. Experiments on 13 benchmarks demonstrate that D2MoE achieves consistent state-of-the-art performance, surpassing leading baselines by up to 7.92% in accuracy on heterophilous graphs. Notably, on large-scale graphs, it reduces memory consumption by up to 73.07% and training time by 46.53% compared to the best-performing Graph MoE, thereby validating its superior efficiency.

cs.LG↗

arXiv2026

CrossHGL: A Text-Free Foundation Model for Cross-Domain Heterogeneous Graph Learning

Heterogeneous graph representation learning (HGRL) is essential for modeling complex systems with diverse node and edge types. However, most existing methods are limited to closed-world settings with shared schemas and feature spaces, hindering cross-domain generalization. While recent graph foundation models improve transferability, they often target homogeneous graphs, rely on domain-specific schemas, or require rich textual attributes. Consequently, text-free and few-shot cross-domain HGRL remains underexplored. To address this, we propose CrossHGL, a foundation framework that preserves and transfers multi-relational structural semantics without external textual supervision. Specifically, a semantic-preserving transformation strategy homogenizes heterogeneous graphs while encoding interaction semantics into edge features. Based on this, a prompt-aware multi-domain pre-training framework with a Tri-Prompt mechanism captures transferable knowledge across feature, edge, and structure perspectives via self-supervised contrastive learning. For target-domain adaptation, we develop a parameter-efficient fine-tuning strategy that freezes the pre-trained backbone and performs few-shot classification via prompt composition and prototypical learning. Experiments on node-level and graph-level tasks show that CrossHGL consistently outperforms state-of-the-art baselines, yielding average relative improvements of 25.1% and 7.6% in Micro-F1 for node and graph classification, respectively, while remaining competitive in challenging feature-degenerated settings.

cs.LG↗

arXiv2026

Cosmic anisotropic hair of nonlocal RT gravity

Nonlocal RT gravity has proven effective in explaining the late-time cosmic acceleration while remaining consistent with local gravity tests. However, most previous cosmological studies of this theory have assumed an isotropic background, which may not fully capture the slight anisotropies suggested by current observations, such as those inferred from Type Ia supernovae data. In this paper, we investigate the dynamical evolution of an anisotropic Bianchi type I universe within the framework of nonlocal RT gravity. By introducing six dimensionless variables, we construct the corresponding dynamical system and perform a detailed phase-space analysis. An unexpected finding is that, contrary to many dark energy models and modified gravity theories in which anisotropies decay with time, nonlocal RT gravity predicts a growth of cosmic anisotropy. This behavior poses a challenge to the cosmic no-hair theorem within the nonlocal RT gravity scenario.

gr-qc↗

arXiv2026

PolyMon: A Unified Framework for Polymer Property Prediction

Accurate prediction of polymer properties is essential for materials design, but remains challenging due to data scarcity, diverse polymer representations, and the lack of systematic evaluation across modelling choices. Here, we present PolyMon, a unified and accessible framework that integrates multiple polymer representations, machine learning methods, and training strategies within a single, accessible platform. PolyMon supports various descriptors and graph construction strategies for polymer representations, and includes a wide range of models, from tabular models to graph neural networks, along with flexible training strategies including multi-fidelity learning, {\Delta}-learning, active learning, and ensemble learning. Using five key polymer properties as benchmarks, we perform systematic evaluations to assess how representations and models affect predictive performance. These case studies further illustrate how different training strategies can be applied within a consistent workflow to leverage limited data and incorporate physical model derived information. Overall, PolyMon provides a comprehensive and extensible foundation for benchmarking and advancing machine learning-based polymer property prediction. The code is available at github.com/fate1997/polymon.

cond-mat.soft↗

arXiv2026

Traffic-MoE: A Sparse Foundation Model for Network Traffic Security Analysis

As adversaries increasingly weaponize encryption and protocol obfuscation to evade traffic detection, traditional methods are rendered obsolete, necessitating deep learning to unmask sophisticated threats. However, the prohibitive computational costs of existing large models create a critical defense gap, hindering their deployment in real-time and throughput-sensitive environments. To close this vulnerability, we introduce Traffic-MoE, a sparse foundation model tailored for traffic security analysis. By dynamically routing traffic tokens to a small subset of specialized experts, Traffic-MoE effectively decouples model capacity from computational overhead. Extensive evaluations across four security-oriented tasks demonstrate that Traffic-MoE achieves state-of-the-art or highly competitive performance compared to leading competitors. Crucially, it delivers a 70.42% increase in throughput, reduces inference latency by 41.39% while significantly optimizing GPU memory consumption. Beyond efficiency, Traffic-MoE exhibits superior robustness against adversarial traffic shaping and maintains strong detection capabilities in few-shot scenarios, establishing a scalable and resilient paradigm for modern network traffic security analysis.

cs.CR↗

arXiv2025

A Time- and Energy-Efficient CNN with Dense Connections on Memristor-Based Chips

Designing lightweight convolutional neural network (CNN) models is an active research area in edge AI. Compute-in-memory (CIM) provides a new computing paradigm to alleviate time and energy consumption caused by data transfer in von Neumann architecture. Among competing alternatives, resistive random-access memory (RRAM) is a promising CIM device owing to its reliability and multi-bit programmability. However, classical lightweight designs such as depthwise convolution incurs under-utilization of RRAM crossbars restricted by their inherently dense weight-to-RRAM cell mapping. To build an RRAM-friendly yet efficient CNN, we evaluate the hardware cost of DenseNet which maintains a high accuracy vs other CNNs at a small parameter count. Observing the linearly increasing channels in DenseNet leads to a low crossbar utilization and causes large latency and energy consumption, we propose a scheme that concatenates feature maps of front layers to form the input of the last layer in each stage. Experiments show that our proposed model consumes less time and energy than conventional ResNet and DenseNet, while producing competitive accuracy on CIFAR and ImageNet datasets.

cs.AR↗

arXiv2025

Hierarchical Local-Global Feature Learning for Few-shot Malicious Traffic Detection

With the rapid growth of internet traffic, malicious network attacks have become increasingly frequent and sophisticated, posing significant threats to global cybersecurity. Traditional detection methods, including rule-based and machine learning-based approaches, struggle to accurately identify emerging threats, particularly in scenarios with limited samples. While recent advances in few-shot learning have partially addressed the data scarcity issue, existing methods still exhibit high false positive rates and lack the capability to effectively capture crucial local traffic patterns. In this paper, we propose HLoG, a novel hierarchical few-shot malicious traffic detection framework that leverages both local and global features extracted from network sessions. HLoG employs a sliding-window approach to segment sessions into phases, capturing fine-grained local interaction patterns through hierarchical bidirectional GRU encoding, while simultaneously modeling global contextual dependencies. We further design a session similarity assessment module that integrates local similarity with global self-attention-enhanced representations, achieving accurate and robust few-shot traffic classification. Comprehensive experiments on three meticulously reconstructed datasets demonstrate that HLoG significantly outperforms existing state-of-the-art methods. Particularly, HLoG achieves superior recall rates while substantially reducing false positives, highlighting its effectiveness and practical value in real-world cybersecurity applications.

cs.CR↗

arXiv2025

Unveiling Latent Information in Transaction Hashes: Hypergraph Learning for Ethereum Ponzi Scheme Detection

With the widespread adoption of Ethereum, financial frauds such as Ponzi schemes have become increasingly rampant in the blockchain ecosystem, posing significant threats to the security of account assets. Existing Ethereum fraud detection methods typically model account transactions as graphs, but this approach primarily focuses on binary transactional relationships between accounts, failing to adequately capture the complex multi-party interaction patterns inherent in Ethereum. To address this, we propose a hypergraph modeling method for the Ponzi scheme detection method in Ethereum, called HyperDet. Specifically, we treat transaction hashes as hyperedges that connect all the relevant accounts involved in a transaction. Additionally, we design a two-step hypergraph sampling strategy to significantly reduce computational complexity. Furthermore, we introduce a dual-channel detection module, including the hypergraph detection channel and the hyper-homo graph detection channel, to be compatible with existing detection methods. Experimental results show that, compared to traditional homogeneous graph-based methods, the hyper-homo graph detection channel achieves significant performance improvements, demonstrating the superiority of hypergraph in Ponzi scheme detection. This research offers innovations for modeling complex relationships in blockchain data.

cs.CR↗

arXiv2025

Pushing DSP-Free Coherent Interconnect to the Last Inch by Optically Analog Signal Processing

To support the boosting interconnect capacity of the AI-related data centers, novel techniques enabled high-speed and low-cost optics are continuously emerging. When the baud rate approaches 200 GBaud per lane, the bottle-neck of traditional intensity modulation direct detection (IM-DD) architectures becomes increasingly evident. The simplified coherent solutions are widely discussed and considered as one of the most promising candidates. In this paper, a novel coherent architecture based on self-homodyne coherent detection and optically analog signal processing (OASP) is demonstrated. Proved by experiment, the first DSP-free baud-rate sampled 64-GBaud QPSK/16-QAM receptions are achieved, with BERs of 1e-6 and 2e-2, respectively. Even with 1-km fiber link propagation, the BER for QPSK reception remains at 3.6e-6. When an ultra-simple 1-sps SISO filter is utilized, the performance degradation of the proposed scheme is less than 1 dB compared to legacy DSP-based coherent reception. The proposed results pave the way for the ultra-high-speed coherent optical interconnections, offering high power and cost efficiency.

physics.optics↗

arXiv2025

QuZO: Quantized Zeroth-Order Fine-Tuning for Large Language Models

Language Models (LLMs) are often quantized to lower precision to reduce the memory cost and latency in inference. However, quantization often degrades model performance, thus fine-tuning is required for various down-stream tasks. Traditional fine-tuning methods such as stochastic gradient descent and Adam optimization require backpropagation, which are error-prone in the low-precision settings. To overcome these limitations, we propose the Quantized Zeroth-Order (QuZO) framework, specifically designed for fine-tuning LLMs through low-precision (e.g., 4- or 8-bit) forward passes. Our method can avoid the error-prone low-precision straight-through estimator, and utilizes optimized stochastic rounding to mitigate the increased bias. QuZO simplifies the training process, while achieving results comparable to first-order methods in ${\rm FP}8$ and superior accuracy in ${\rm INT}8$ and ${\rm INT}4$ training. Experiments demonstrate that low-bit training QuZO achieves performance comparable to MeZO optimization on GLUE, Multi-Choice, and Generation tasks, while reducing memory cost by $2.94 \times$ in LLaMA2-7B fine-tuning compared to quantized first-order methods.

cs.LG↗

arXiv2025

Mixture of Message Passing Experts with Routing Entropy Regularization for Node Classification

Graph neural networks (GNNs) have achieved significant progress in graph-based learning tasks, yet their performance often deteriorates when facing heterophilous structures where connected nodes differ substantially in features and labels. To address this limitation, we propose GNNMoE, a novel entropy-driven mixture of message-passing experts framework that enables node-level adaptive representation learning. GNNMoE decomposes message passing into propagation and transformation operations and integrates them through multiple expert networks guided by a hybrid routing mechanism. And a routing entropy regularization dynamically adjusts soft weighting and soft top-$k$ routing, allowing GNNMoE to flexibly adapt to diverse neighborhood contexts. Extensive experiments on twelve benchmark datasets demonstrate that GNNMoE consistently outperforms SOTA node classification methods, while maintaining scalability and interpretability. This work provides a unified and principled approach for achieving fine-grained, personalized node representation learning.

cs.LG↗

arXiv2025

Multi-view Correlation-aware Network Traffic Detection on Flow Hypergraph

As the Internet rapidly expands, the increasing complexity and diversity of network activities pose significant challenges to effective network governance and security regulation. Network traffic, which serves as a crucial data carrier of network activities, has become indispensable in this process. Network traffic detection aims to monitor, analyze, and evaluate the data flows transmitted across the network to ensure network security and optimize performance. However, existing network traffic detection methods generally suffer from several limitations: 1) a narrow focus on characterizing traffic features from a single perspective; 2) insufficient exploration of discriminative features for different traffic; 3) poor generalization to different traffic scenarios. To address these issues, we propose a multi-view correlation-aware framework named FlowID for network traffic detection. FlowID captures multi-view traffic features via temporal and interaction awareness, while a hypergraph encoder further explores higher-order relationships between flows. To overcome the challenges of data imbalance and label scarcity, we design a dual-contrastive proxy task, enhancing the framework's ability to differentiate between various traffic flows through flow-to-flow and group-to-group contrast. Extensive experiments on five real-world datasets demonstrate that FlowID significantly outperforms existing methods in accuracy, robustness, and generalization across diverse network scenarios, particularly in detecting malicious traffic.

cs.CR↗

arXiv2024

Mixture of Experts Meets Decoupled Message Passing: Towards General and Adaptive Node Classification

Graph neural networks excel at graph representation learning but struggle with heterophilous data and long-range dependencies. And graph transformers address these issues through self-attention, yet face scalability and noise challenges on large-scale graphs. To overcome these limitations, we propose GNNMoE, a universal model architecture for node classification. This architecture flexibly combines fine-grained message-passing operations with a mixture-of-experts mechanism to build feature encoding blocks. Furthermore, by incorporating soft and hard gating layers to assign the most suitable expert networks to each node, we enhance the model's expressive power and adaptability to different graph types. In addition, we introduce adaptive residual connections and an enhanced FFN module into GNNMoE, further improving the expressiveness of node representation. Extensive experimental results demonstrate that GNNMoE performs exceptionally well across various types of graph data, effectively alleviating the over-smoothing issue and global noise, enhancing model robustness and adaptability, while also ensuring computational efficiency on large-scale graphs.

cs.LG↗

arXiv2024

Lateral Movement Detection via Time-aware Subgraph Classification on Authentication Logs

Lateral movement is a crucial component of advanced persistent threat (APT) attacks in networks. Attackers exploit security vulnerabilities in internal networks or IoT devices, expanding their control after initial infiltration to steal sensitive data or carry out other malicious activities, posing a serious threat to system security. Existing research suggests that attackers generally employ seemingly unrelated operations to mask their malicious intentions, thereby evading existing lateral movement detection methods and hiding their intrusion traces. In this regard, we analyze host authentication log data from a graph perspective and propose a multi-scale lateral movement detection framework called LMDetect. The main workflow of this framework proceeds as follows: 1) Construct a heterogeneous multigraph from host authentication log data to strengthen the correlations among internal system entities; 2) Design a time-aware subgraph generator to extract subgraphs centered on authentication events from the heterogeneous authentication multigraph; 3) Design a multi-scale attention encoder that leverages both local and global attention to capture hidden anomalous behavior patterns in the authentication subgraphs, thereby achieving lateral movement detection. Extensive experiments on two real-world authentication log datasets demonstrate the effectiveness and superiority of our framework in detecting lateral movement behaviors.

cs.CR↗