From Prior-Guided Heuristics to Deployable Agents: Accelerating Demonstration-Driven Reinforcement Learning for Deadline-Constrained Network Control
本文针对网络控制中延迟敏感信息的及时传递问题，提出了一种结合有效拥塞度量与多智能体深度强化学习的新框架，并引入模型引导退火强化学习协议以提高训练效率。
0 citationsRead paper
Selected work
Representative Papers
Vaulted Passkeys: A Device-Bound Proposal for Authenticated Credential Export and Import
This study addresses the lack of secure portability and disaster recovery mechanisms for device-bound credentials by proposing Vaulted Passkeys. The architecture prevents plaintext private key exposure through decoupled registration and export workflows, a role separation model, and dedicated key derivation. Furthermore, it enables secure credential migration by integrating HKDF-based key separation, AEAD configuration, and device-bound encrypted envelopes. A prototype implementation validates the feasibility of this approach, accompanied by a comprehensive threat analysis and a falsifiable evaluation plan. Collectively, this work effectively bridges the critical gap in secure recovery mechanisms for hardware authenticators, offering a robust solution for managing device-bound credentials without compromising their underlying security guarantees during transfer or restoration processes.
0 citationsRead paper
An open-source implementation and validation of 5G NR Configured Grant for URLLC in ns-3 5G LENA: a scheduling case study in Industry 4.0 scenarios
This study addresses the challenge of excessive uplink latency in existing 5G dynamic scheduling—caused by signaling overhead—which hinders support for ultra-reliable low-latency communication (URLLC) requirements in Industry 4.0. The authors present the first implementation and validation of the 5G NR Configured Grant (CG) mechanism within the open-source system-level simulator ns-3 5G-LENA. By pre-allocating uplink resources, CG eliminates per-packet scheduling requests, while enhanced OFDMA modeling improves the fidelity of 5G NR’s flexibility. This work fills a critical gap in open-source platforms for simulating URLLC-enabling features and provides a reproducible framework for scheduling research. Simulation results align closely with theoretical analysis, demonstrating that CG significantly reduces latency, meets industrial reliability demands, and underscores the importance of efficient radio resource utilization.
0 citationsRead paper
Toward Experimentation-as-a-Service in 5G/6G: The Plaza6G Prototype for AI-Assisted Trials
This work addresses the high complexity, poor reproducibility, and lack of a unified Experimentation-as-a-Service (ExaS) environment in existing 5G/6G wireless testbeds by introducing Plaza6G—the first integrated ExaS platform combining open-source and commercial 5G core networks, a programmable radio access network (RAN), and an AI-driven natural language interface. Plaza6G integrates GPU-accelerated clusters, multiple core networks (Free5GC and Cumucore), and sub-6GHz/mmWave dual-site over-the-air (OTA) testing capabilities. It features an intelligent experimentation assistant built upon LLM, RAG, and LoRA techniques to support machine-readable experiment specifications. The platform enables fully automated CI/CD deployment within ten minutes, facilitating low-barrier, reproducible, and interactive wireless experimentation, thereby laying the groundwork for future federated testbeds and policy-aware orchestration.
0 citationsRead paper
Degradation of Feature Space in Continual Learning
This study investigates whether enforcing isotropy in the feature space mitigates catastrophic forgetting and improves representation quality in continual learning. By comparing the geometric structure of feature spaces under centralized versus continual learning settings, the work reveals fundamental differences between the two and challenges the applicability of isotropy as a universal inductive bias. Experimental results on CIFAR-10 and CIFAR-100 demonstrate that isotropic regularization not only fails to enhance performance but actually degrades accuracy, suggesting that isotropy is ill-suited as an inductive bias in non-stationary learning scenarios. These findings offer a new perspective on the geometry of representations in continual learning and question the uncritical adoption of isotropy-promoting objectives from static learning contexts.
0 citationsRead paper
Recent publications
Latest Papers
From Prior-Guided Heuristics to Deployable Agents: Accelerating Demonstration-Driven Reinforcement Learning for Deadline-Constrained Network Control
本文针对网络控制中延迟敏感信息的及时传递问题，提出了一种结合有效拥塞度量与多智能体深度强化学习的新框架，并引入模型引导退火强化学习协议以提高训练效率。
0 citationsRead paper
Vaulted Passkeys: A Device-Bound Proposal for Authenticated Credential Export and Import
This study addresses the lack of secure portability and disaster recovery mechanisms for device-bound credentials by proposing Vaulted Passkeys. The architecture prevents plaintext private key exposure through decoupled registration and export workflows, a role separation model, and dedicated key derivation. Furthermore, it enables secure credential migration by integrating HKDF-based key separation, AEAD configuration, and device-bound encrypted envelopes. A prototype implementation validates the feasibility of this approach, accompanied by a comprehensive threat analysis and a falsifiable evaluation plan. Collectively, this work effectively bridges the critical gap in secure recovery mechanisms for hardware authenticators, offering a robust solution for managing device-bound credentials without compromising their underlying security guarantees during transfer or restoration processes.
0 citationsRead paper
An open-source implementation and validation of 5G NR Configured Grant for URLLC in ns-3 5G LENA: a scheduling case study in Industry 4.0 scenarios
This study addresses the challenge of excessive uplink latency in existing 5G dynamic scheduling—caused by signaling overhead—which hinders support for ultra-reliable low-latency communication (URLLC) requirements in Industry 4.0. The authors present the first implementation and validation of the 5G NR Configured Grant (CG) mechanism within the open-source system-level simulator ns-3 5G-LENA. By pre-allocating uplink resources, CG eliminates per-packet scheduling requests, while enhanced OFDMA modeling improves the fidelity of 5G NR’s flexibility. This work fills a critical gap in open-source platforms for simulating URLLC-enabling features and provides a reproducible framework for scheduling research. Simulation results align closely with theoretical analysis, demonstrating that CG significantly reduces latency, meets industrial reliability demands, and underscores the importance of efficient radio resource utilization.
0 citationsRead paper
Toward Experimentation-as-a-Service in 5G/6G: The Plaza6G Prototype for AI-Assisted Trials
This work addresses the high complexity, poor reproducibility, and lack of a unified Experimentation-as-a-Service (ExaS) environment in existing 5G/6G wireless testbeds by introducing Plaza6G—the first integrated ExaS platform combining open-source and commercial 5G core networks, a programmable radio access network (RAN), and an AI-driven natural language interface. Plaza6G integrates GPU-accelerated clusters, multiple core networks (Free5GC and Cumucore), and sub-6GHz/mmWave dual-site over-the-air (OTA) testing capabilities. It features an intelligent experimentation assistant built upon LLM, RAG, and LoRA techniques to support machine-readable experiment specifications. The platform enables fully automated CI/CD deployment within ten minutes, facilitating low-barrier, reproducible, and interactive wireless experimentation, thereby laying the groundwork for future federated testbeds and policy-aware orchestration.
0 citationsRead paper
Degradation of Feature Space in Continual Learning
This study investigates whether enforcing isotropy in the feature space mitigates catastrophic forgetting and improves representation quality in continual learning. By comparing the geometric structure of feature spaces under centralized versus continual learning settings, the work reveals fundamental differences between the two and challenges the applicability of isotropy as a universal inductive bias. Experimental results on CIFAR-10 and CIFAR-100 demonstrate that isotropic regularization not only fails to enhance performance but actually degrades accuracy, suggesting that isotropy is ill-suited as an inductive bias in non-stationary learning scenarios. These findings offer a new perspective on the geometry of representations in continual learning and question the uncritical adoption of isotropy-promoting objectives from static learning contexts.
0 citationsRead paper