Selected work

Representative Papers

AgenTRIM: Tool Risk Mitigation for Agentic AI

Jan 18, 2026

This work addresses the security risks—such as indirect prompt injection—and performance degradation arising from improper tool permission configurations in AI agents, which often manifest as overuse or underuse of tools. The authors propose AgenTRIM, a novel framework that formally characterizes the problem of tool-induced capability imbalance in agents for the first time. Without modifying the agent’s internal logic, AgenTRIM enables runtime risk detection and mitigation through offline interface reconstruction verification and online dynamic filtering based on the principle of least privilege. By integrating code and execution trace analysis, state-aware validation, and adaptive call filtering, the approach significantly reduces attack success rates on the AgentDojo benchmark while maintaining high task completion rates, demonstrating strong robustness against both descriptive attacks and explicit security policies.

2 citationsRead paper

CaRGo-T: Causal Reasoning Graph-of-Thought improves Multimodal Humor Comprehension

Aug 24, 2026

本文提出CaRGo-T方法，通过构建因果关系图来改进多模态幽默理解与检测问题，实验显示该方法在多个数据集上优于现有基线。

0 citationsRead paper

RIFT-Bench: Dynamic Red-teaming For Agentic AI Systems

Jun 22, 2026

Existing security evaluation methods struggle to conduct unified and dynamic red-teaming assessments of heterogeneous AI agent systems due to the absence of a cross-architecture general framework. This work proposes RIFT-Bench, a dynamic red-teaming approach based on hierarchical graph representations, which enables unified modeling, multi-target adversarial attacks, and comprehensive evaluation through a two-stage pipeline comprising automated structural discovery and adaptive adversarial probing. RIFT-Bench introduces, for the first time, a graph-based representation tailored for AI agents alongside a dynamic probing mechanism, facilitating cross-architectural evaluation and validation of mitigation strategies. Experiments across 45 heterogeneous agent systems demonstrate that RIFT-Bench achieves strong generalization and scalability, establishing a robust foundation for the security assessment of AI agent systems.

0 citationsRead paper

PLGC: Pseudo-Labeled Graph Condensation

Jan 15, 2026

Training graph neural networks on large-scale graphs is computationally expensive, and existing graph condensation methods rely heavily on clean labels, leading to significant performance degradation under label scarcity, noise, or distribution shifts. This work proposes a self-supervised graph condensation framework that generates latent pseudo-labels from node embeddings without requiring ground-truth labels. By jointly optimizing prototypes and node assignments, the method constructs a compact synthetic graph whose structural and feature statistics closely match those of the original graph. Theoretical analysis demonstrates that the approach effectively preserves the original graph structure and ensures embedding alignment. Experiments show that the method matches state-of-the-art supervised approaches on clean data and substantially outperforms all baselines under label noise, exhibiting remarkable robustness in both node classification and link prediction tasks.

0 citationsRead paper

Quantum Generative Adversarial Autoencoders: Learning latent representations for quantum data generation

Sep 19, 2025

Addressing the challenges of representation learning and generative modeling for quantum data, this paper introduces the Quantum Generative Adversarial Autoencoder (QGAA), the first framework to integrate generative adversarial principles into a quantum autoencoder architecture. QGAA jointly employs a Quantum Autoencoder (QAE) for quantum state compression and a Quantum Generative Adversarial Network (QGAN) to model the latent-space distribution, enabling end-to-end training via variational quantum circuits and gradient-based optimization. The model learns interpretable, low-dimensional latent representations and accurately synthesizes target quantum states: it reconstructs pure entangled states on a 6-qubit system and generates ground states of H₂ and LiH molecules with mean energy estimation errors of only 0.02 Ha and 0.06 Ha, respectively. This work significantly extends the applicability of quantum generative models to quantum chemistry simulation and near-term noisy intermediate-scale quantum (NISQ) hardware.

0 citationsRead paper

Recent publications

Latest Papers

CaRGo-T: Causal Reasoning Graph-of-Thought improves Multimodal Humor Comprehension

Aug 24, 2026

本文提出CaRGo-T方法，通过构建因果关系图来改进多模态幽默理解与检测问题，实验显示该方法在多个数据集上优于现有基线。

0 citationsRead paper

RIFT-Bench: Dynamic Red-teaming For Agentic AI Systems

Jun 22, 2026

Existing security evaluation methods struggle to conduct unified and dynamic red-teaming assessments of heterogeneous AI agent systems due to the absence of a cross-architecture general framework. This work proposes RIFT-Bench, a dynamic red-teaming approach based on hierarchical graph representations, which enables unified modeling, multi-target adversarial attacks, and comprehensive evaluation through a two-stage pipeline comprising automated structural discovery and adaptive adversarial probing. RIFT-Bench introduces, for the first time, a graph-based representation tailored for AI agents alongside a dynamic probing mechanism, facilitating cross-architectural evaluation and validation of mitigation strategies. Experiments across 45 heterogeneous agent systems demonstrate that RIFT-Bench achieves strong generalization and scalability, establishing a robust foundation for the security assessment of AI agent systems.

0 citationsRead paper

AgenTRIM: Tool Risk Mitigation for Agentic AI

Jan 18, 2026

This work addresses the security risks—such as indirect prompt injection—and performance degradation arising from improper tool permission configurations in AI agents, which often manifest as overuse or underuse of tools. The authors propose AgenTRIM, a novel framework that formally characterizes the problem of tool-induced capability imbalance in agents for the first time. Without modifying the agent’s internal logic, AgenTRIM enables runtime risk detection and mitigation through offline interface reconstruction verification and online dynamic filtering based on the principle of least privilege. By integrating code and execution trace analysis, state-aware validation, and adaptive call filtering, the approach significantly reduces attack success rates on the AgentDojo benchmark while maintaining high task completion rates, demonstrating strong robustness against both descriptive attacks and explicit security policies.

2 citationsRead paper

PLGC: Pseudo-Labeled Graph Condensation

Jan 15, 2026

Training graph neural networks on large-scale graphs is computationally expensive, and existing graph condensation methods rely heavily on clean labels, leading to significant performance degradation under label scarcity, noise, or distribution shifts. This work proposes a self-supervised graph condensation framework that generates latent pseudo-labels from node embeddings without requiring ground-truth labels. By jointly optimizing prototypes and node assignments, the method constructs a compact synthetic graph whose structural and feature statistics closely match those of the original graph. Theoretical analysis demonstrates that the approach effectively preserves the original graph structure and ensures embedding alignment. Experiments show that the method matches state-of-the-art supervised approaches on clean data and substantially outperforms all baselines under label noise, exhibiting remarkable robustness in both node classification and link prediction tasks.

0 citationsRead paper

Quantum Generative Adversarial Autoencoders: Learning latent representations for quantum data generation

Sep 19, 2025

Addressing the challenges of representation learning and generative modeling for quantum data, this paper introduces the Quantum Generative Adversarial Autoencoder (QGAA), the first framework to integrate generative adversarial principles into a quantum autoencoder architecture. QGAA jointly employs a Quantum Autoencoder (QAE) for quantum state compression and a Quantum Generative Adversarial Network (QGAN) to model the latent-space distribution, enabling end-to-end training via variational quantum circuits and gradient-based optimization. The model learns interpretable, low-dimensional latent representations and accurately synthesizes target quantum states: it reconstructs pure entangled states on a 6-qubit system and generates ground states of H₂ and LiH molecules with mean energy estimation errors of only 0.02 Ha and 0.06 Ha, respectively. This work significantly extends the applicability of quantum generative models to quantum chemistry simulation and near-term noisy intermediate-scale quantum (NISQ) hardware.

0 citationsRead paper