Influence Is Not Authority: When Causal Guardrail Signals Make Legitimate Tool Use Look Like an Attack in Tool-Using LLM Agents
研究解决了当前基于影响的防护机制无法可靠区分合法与恶意行为的问题，通过授权等效审计方法揭示了信号误判的原因及其对系统安全性和实用性的负面影响。
0 citationsRead paper
Selected work
Representative Papers
GuidedFlow: An Attention-Guided Framework for Anomaly Detection in Additive Manufacturing
为解决3D打印中的微小缺陷检测问题，提出GuidedFlow模型，结合注意力机制与归一化流方法提高异常检测准确性。
0 citationsRead paper
Spiking Neural Networks for Continuous Control: Neuromorphic Reinforcement Learning in Conventional Computing
研究通过提出一种基于脉冲神经网络的软演员评论家算法（SANSAC），在连续控制任务中验证了其与传统方法相近的性能，探索了脉冲神经网络在强化学习中的应用潜力。
0 citationsRead paper
Density-Driven Optimal Control: Convergence Guarantees for Stochastic LTI Multi-Agent Systems
This work addresses the decentralized non-uniform coverage problem for multi-agent systems under resource constraints and high spatial priority tasks by proposing a Stochastic Density-Driven Optimal Control (D²OC) approach. The method formulates a Lagrangian framework under stochastic linear time-invariant (LTI) dynamics, minimizing the Wasserstein distance as the running cost to drive the empirical distribution of agents toward a nonparametric target density. It provides the first formal convergence guarantee for stochastic LTI multi-agent systems and integrates reachability analysis to ensure bounded tracking errors in the presence of process and measurement noise. Numerical experiments demonstrate that the proposed method significantly outperforms existing heuristic strategies in both coverage optimality and consensus, achieving robust decentralized coverage.
0 citationsRead paper
Computationally Efficient Density-Driven Optimal Control via Analytical KKT Reduction and Contractive MPC
This work addresses the high computational complexity inherent in achieving efficient spatial density distributions in multi-agent systems, where conventional density-driven optimal control strategies are often intractable for online implementation. To overcome this challenge, the authors propose a dimensionality reduction approach based on analytical Karush–Kuhn–Tucker (KKT) conditions, which reformulates the multi-step predictive control problem into a quadratic program with linear time complexity O(T). This method is integrated within a contractive model predictive control (MPC) framework, incorporating Lyapunov-based contraction constraints to guarantee input-to-state stability of the closed-loop system. The resulting algorithm substantially reduces computational overhead, enabling real-time density regulation for large-scale multi-agent systems over extended prediction horizons. Numerical simulations demonstrate its superior performance in rapid spatial coverage and computational efficiency.
0 citationsRead paper
Recent publications
Latest Papers
Influence Is Not Authority: When Causal Guardrail Signals Make Legitimate Tool Use Look Like an Attack in Tool-Using LLM Agents
研究解决了当前基于影响的防护机制无法可靠区分合法与恶意行为的问题，通过授权等效审计方法揭示了信号误判的原因及其对系统安全性和实用性的负面影响。
0 citationsRead paper
GuidedFlow: An Attention-Guided Framework for Anomaly Detection in Additive Manufacturing
为解决3D打印中的微小缺陷检测问题，提出GuidedFlow模型，结合注意力机制与归一化流方法提高异常检测准确性。
0 citationsRead paper
Spiking Neural Networks for Continuous Control: Neuromorphic Reinforcement Learning in Conventional Computing
研究通过提出一种基于脉冲神经网络的软演员评论家算法（SANSAC），在连续控制任务中验证了其与传统方法相近的性能，探索了脉冲神经网络在强化学习中的应用潜力。
0 citationsRead paper
Density-Driven Optimal Control: Convergence Guarantees for Stochastic LTI Multi-Agent Systems
This work addresses the decentralized non-uniform coverage problem for multi-agent systems under resource constraints and high spatial priority tasks by proposing a Stochastic Density-Driven Optimal Control (D²OC) approach. The method formulates a Lagrangian framework under stochastic linear time-invariant (LTI) dynamics, minimizing the Wasserstein distance as the running cost to drive the empirical distribution of agents toward a nonparametric target density. It provides the first formal convergence guarantee for stochastic LTI multi-agent systems and integrates reachability analysis to ensure bounded tracking errors in the presence of process and measurement noise. Numerical experiments demonstrate that the proposed method significantly outperforms existing heuristic strategies in both coverage optimality and consensus, achieving robust decentralized coverage.
0 citationsRead paper
Computationally Efficient Density-Driven Optimal Control via Analytical KKT Reduction and Contractive MPC
This work addresses the high computational complexity inherent in achieving efficient spatial density distributions in multi-agent systems, where conventional density-driven optimal control strategies are often intractable for online implementation. To overcome this challenge, the authors propose a dimensionality reduction approach based on analytical Karush–Kuhn–Tucker (KKT) conditions, which reformulates the multi-step predictive control problem into a quadratic program with linear time complexity O(T). This method is integrated within a contractive model predictive control (MPC) framework, incorporating Lyapunov-based contraction constraints to guarantee input-to-state stability of the closed-loop system. The resulting algorithm substantially reduces computational overhead, enabling real-time density regulation for large-scale multi-agent systems over extended prediction horizons. Numerical simulations demonstrate its superior performance in rapid spatial coverage and computational efficiency.
0 citationsRead paper