arXiv subjects

Qingshuai Feng

Publications and source records attributed to Qingshuai Feng.

6 recordsLinked to original sources

arXiv2026

Evolving Skill-Structured Attack Memory Enhances LLM Jailbreaking

Jailbreak attacks on large language models (LLMs) aim to induce LLMs to produce content that they are expected to refuse. Automated black-box jailbreak generation is important for safety evaluation, where the attacker observes only model outputs and needs to search for effective adversarial prompts. Existing black-box jailbreak methods either depend on sample-wise heuristic search or leverage attack experience through accumulating strategy pools or method libraries, lacking a systematic organization and management of attack experience. To mitigate these drawbacks, we propose MemoAttack, a memory-driven black-box jailbreak framework with comprehensive attack memory modeling, evolution, and selection. Specifically, MemoAttack comprises three key designs: (1) Skill-Structured Memory Modeling, which abstracts accumulated attack experience into reusable skill-structured attack memory whose units pair attack skills with templates, evidence, and lifecycle state; (2) Lifecycle-Driven Memory Evolution, which evolves the memory through evidence-based probation, promotion, retirement, reactivation, elimination, and storage cleanup; and (3) Posterior-Guided Contextual Memory Selection, which balances reliable memory reuse with uncertainty-driven exploration via contextual Thompson sampling. Across three target models on AdvBench, MemoAttack achieves attack success rates of 93.33-96.67%, exceeding the strongest baseline on each target by 10.00-12.00 percentage points while reducing mean expansion cost on its own successful goals by 20.2-51.6%. In a sequential 400-goal evaluation on Qwen3.5, the trailing 50-goal mean expansion-attempt count decreases overall from 19.74 to 10.92 as memory accumulates.

cs.CR↗

arXiv2026

Efficient $(α,β)$-core Computation and On-the-fly Query at Billion Scale with GPUs

In bipartite graphs, $(α,β)$-core is a widely used model for cohesive subgraph mining. Specifically, an $(α,β)$-core is a maximal subgraph in which each vertex in the upper layer has degree at least $α$, and each vertex in the lower layer has degree at least $β$. The state-of-the-art CPU-based solutions incur extensive costs to construct an index structure for all $α$ and $β$ combinations, leading to scalability challenges on large bipartite graphs. Moreover, on-the-fly queries, which aim to determine whether an edge update belongs to a target $(α,β)$-core, are essential for real-time applications such as fraud monitoring and recommendation systems. However, existing index-based methods struggle to support such queries at scale due to their high maintenance overhead. In this paper, we investigate how to leverage GPU architectures to enable efficient $(α,β)$-core computation and support on-the-fly queries. While GPUs are widely used to accelerate graph processing, their limited memory capacity makes it impractical to store large index structures. To address this issue, we propose GCC, an index-free GPU-based peeling algorithm that accelerates $(α,β)$-core computation via warp-centric processing. To further improve efficiency, we develop GCC+, which leverages the nested property of $(α,β)$-core with a core-based early pruning strategy. For handling on-the-fly queries, we propose GFQ, a connectivity-aware algorithm that significantly narrows the computation scope by leveraging connected component information, thereby avoiding full-graph peeling. Extensive experiments on 11 datasets demonstrate that our proposed techniques outperform existing CPU-based solutions in terms of both space and time efficiency.

cs.DB↗

arXiv2025

Task-Aware Retrieval Augmentation for Dynamic Recommendation

Dynamic recommendation systems aim to provide personalized suggestions by modeling temporal user-item interactions across time-series behavioral data. Recent studies have leveraged pre-trained dynamic graph neural networks (GNNs) to learn user-item representations over temporal snapshot graphs. However, fine-tuning GNNs on these graphs often results in generalization issues due to temporal discrepancies between pre-training and fine-tuning stages, limiting the model's ability to capture evolving user preferences. To address this, we propose TarDGR, a task-aware retrieval-augmented framework designed to enhance generalization capability by incorporating task-aware model and retrieval-augmentation. Specifically, TarDGR introduces a Task-Aware Evaluation Mechanism to identify semantically relevant historical subgraphs, enabling the construction of task-specific datasets without manual labeling. It also presents a Graph Transformer-based Task-Aware Model that integrates semantic and structural encodings to assess subgraph relevance. During inference, TarDGR retrieves and fuses task-aware subgraphs with the query subgraph, enriching its representation and mitigating temporal generalization issues. Experiments on multiple large-scale dynamic graph datasets demonstrate that TarDGR consistently outperforms state-of-the-art methods, with extensive empirical evidence underscoring its superior accuracy and generalization capabilities.

cs.IR↗

arXiv2025

CS-Agent: LLM-based Community Search via Dual-agent Collaboration

Large Language Models (LLMs) have demonstrated remarkable capabilities in natural language processing tasks, yet their application to graph structure analysis, particularly in community search, remains underexplored. Community search, a fundamental task in graph analysis, aims to identify groups of nodes with dense interconnections, which is crucial for understanding the macroscopic structure of graphs. In this paper, we propose GraphCS, a comprehensive benchmark designed to evaluate the performance of LLMs in community search tasks. Our experiments reveal that while LLMs exhibit preliminary potential, they frequently fail to return meaningful results and suffer from output bias. To address these limitations, we introduce CS-Agent, a dual-agent collaborative framework to enhance LLM-based community search. CS-Agent leverages the complementary strengths of two LLMs acting as Solver and Validator. Through iterative feedback and refinement, CS-Agent dynamically refines initial results without fine-tuning or additional training. After the multi-round dialogue, Decider module selects the optimal community. Extensive experiments demonstrate that CS-Agent significantly improves the quality and stability of identified communities compared to baseline methods. To our knowledge, this is the first work to apply LLMs to community search, bridging the gap between LLMs and graph analysis while providing a robust and adaptive solution for real-world applications.

cs.SI↗

arXiv2023

DSPC: Efficiently Answering Shortest Path Counting on Dynamic Graphs

The widespread use of graph data in various applications and the highly dynamic nature of today's networks have made it imperative to analyze structural trends in dynamic graphs on a continual basis. The shortest path is a fundamental concept in graph analysis and recent research shows that counting the number of shortest paths between two vertices is crucial in applications like potential friend recommendation and betweenness analysis. However, current studies that use hub labeling techniques for real-time shortest path counting are limited by their reliance on a pre-computed index, which cannot tackle frequent updates over dynamic graphs. To address this, we propose a novel approach for maintaining the index in response to changes in the graph structure and develop incremental (IncSPC) and decremental (DecSPC) update algorithms for inserting and deleting vertices/edges, respectively. The main idea of these two algorithms is that we only locate the affected vertices to update the index. Our experiments demonstrate that our dynamic algorithms are up to four orders of magnitude faster processing for incremental updates and up to three orders of magnitude faster processing for hybrid updates than reconstruction.

cs.DB↗

arXiv2022

Towards Real-Time Counting Shortest Cycles on Dynamic Graphs: A Hub Labeling Approach

With the ever-increasing prevalence of graph data in a wide spectrum of applications, it becomes essential to analyze structural trends in dynamic graphs on a continual basis. The shortest cycle is a fundamental pattern in graph analytics. In this paper, we investigate the problem of shortest cycle counting for a given vertex in dynamic graphs in light of its applicability to problems such as fraud detection. To address such queries efficiently, we propose a 2-hop labeling based algorithm called Counting Shortest Cycle (CSC for short). Additionally, techniques for dynamically updating the CSC index are explored. Comprehensive experiments are conducted to demonstrate the efficiency and effectiveness of our method. In particular, CSC enables query evaluation in a few hundreds of microseconds for graphs with millions of edges, and improves query efficiency by two orders of magnitude when compared to the baseline solutions. Also, the update algorithm could efficiently cope with edge insertions (deletions).

cs.SI↗