arXiv subjects

Zihao Feng

Publications and source records attributed to Zihao Feng.

11 recordsLinked to original sources

arXiv2026

MANA: Towards Efficient Mobile Ad Detection via Multimodal Agentic UI Navigation

Mobile advertising dominates app monetization but introduces risks ranging from intrusive user experience to malware delivery. Existing detection methods rely either on static analysis, which misses runtime behaviors, or on heuristic UI exploration, which struggles with sparse and obfuscated ads. In this paper, we present MANA, the first agentic multimodal reasoning framework for mobile ad detection. MANA integrates static, visual, temporal, and experiential signals into a reasoning-guided navigation strategy that determines not only how to traverse interfaces but also where to focus, enabling efficient and robust exploration. We implement and evaluate MANA on commercial smartphones over 200 apps, achieving state-of-the-art accuracy and efficiency. Compared to baselines, it improves detection accuracy by 30.5%-56.3% and reduces exploration steps by 29.7%-63.3%. Case studies further demonstrate its ability to uncover obfuscated and malicious ads, underscoring its practicality for mobile ad auditing and its potential for broader runtime UI analysis (e.g., permission abuse). Code and dataset are available at https://github.com/MANA-2026/MANA.

cs.CR↗

arXiv2025

Abnormal Polarity Effect on the DC Breakdown Voltage in Short SF6 Gap

In this Letter, through the comparison between experiment and numerical simulation, we reveal the dynamic mechanism underlying the abnormal polarity effect in SF6 short-gap DC breakdown, as well as a novel criterion for predicting negative breakdown voltage. Using the traditional single-streamer breakdown criterion, the simulated positive breakdown voltage agrees well with experimental measurements, whereas the simulated negative breakdown voltage deviates markedly from the experiments, so the single-streamer breakdown criterion fails to reproduce the abnormal polarity effect observed experimentally. In addressing this, we propose an ion-ion plasma breakdown criterion for negative breakdown voltage. When this novel criterion is applied, the simulated negative breakdown voltage agrees with the experiments and reflects the abnormal polarity effect. Analysis of the spatiotemporal evolution of key physical parameters reveals that, the dynamic mechanism for ion-ion plasma breakdown for negative polarity can be divided into four stages: primary streamer stage, ion accumulation stage, reconstructive ionization stage, and ion-ion plasma propagation stage. Notably, the ion-ion plasma propagation stage is dominated by photoionization-driven negative-ion accumulation rather than conventional impact ionization.

physics.plasm-ph↗

arXiv2025

Experimental and numerical study on Following Streamer mechanism for SF6 breakdown induced by floating linear metal particles

A recently proposed Following Streamer mechanism (Feng et al. 2025 Phys. Rev. Applied 23 064039) seeks to explain how floating metal particles induce SF6 streamer breakdown in the combined gap. This mechanism is derived from a 2D axisymmetric fluid model, which has limitations in describing multiple streamer events in real-world 3D scenarios. To validate the Following Streamer mechanism, we experimentally investigate the discharge morphology of SF6 streamers induced by a floating linear metal particle under negative pulsed voltage. The results are then compared with those from 2D axisymmetric fluid simulations. The comparison reveals both consistencies and discrepancies. Regarding consistencies, experimentally observed features-such as streamer inception at both ends of the metal particle and the formation of subsequent following streamers-support the general idea of the Following Streamer mechanism. Regarding discrepancies, the experiments show a larger number of following streamers and off-axis propagation paths, which cannot be described in the 2D simulation. For scientific rigor, an extended physical model was proposed to improve the description of the Following Streamer mechanism.

physics.plasm-ph↗

arXiv2025

ToolSample: Dual Dynamic Sampling Methods with Curriculum Learning for RL-based Tool Learning

While reinforcement learning (RL) is increasingly used for LLM-based tool learning, its efficiency is often hampered by an overabundance of simple samples that provide diminishing learning value as training progresses. Existing dynamic sampling techniques are ill-suited for the multi-task structure and fine-grained reward mechanisms inherent to tool learning. This paper introduces Dynamic Sampling with Curriculum Learning (DSCL), a framework specifically designed to address this challenge by targeting the unique characteristics of tool learning: its multiple interdependent sub-tasks and multi-valued reward functions. DSCL features two core components: Reward-Based Dynamic Sampling, which uses multi-dimensional reward statistics (mean and variance) to prioritize valuable data, and Task-Based Dynamic Curriculum Learning, which adaptively focuses training on less-mastered sub-tasks. Through extensive experiments, we demonstrate that DSCL significantly improves training efficiency and model performance over strong baselines, achieving a 3.29\% improvement on the BFCLv3 benchmark. Our method provides a tailored solution that effectively leverages the complex reward signals and sub-task dynamics within tool learning to achieve superior results.

cs.LG↗

arXiv2025

Empowering LLMs in Task-Oriented Dialogues: A Domain-Independent Multi-Agent Framework and Fine-Tuning Strategy

Task-oriented dialogue systems based on Large Language Models (LLMs) have gained increasing attention across various industries and achieved significant results. Current approaches condense complex procedural workflows into a single agent to achieve satisfactory performance on large-scale LLMs. However, these approaches face challenges to achieve comparable performance on fine-tuned lightweight LLMs, due to their limited capabilities in handling multiple complex logic. In this work, we design a Domain-Independent Multi-Agent Framework (DIMF), which contains Intent Classification Agent, Slot Filling Agent and Response Agent. This approach simplifies the learning complexity and enhances the generalization ability by separating the tasks into domain-independent components. In this framework, we enhance the capabilities in contextual understanding using the Direct Preference Optimisation (DPO) method, and propose a simple and effective Data Distribution Adaptation (DDA) method to mitigate degradation issues during DPO training. Experiments conducted on the MultiWOZ datasets show that our proposed method achieves a better average performance among all the baselines. Extensive analysis also demonstrates that our proposed framework exhibits excellent generalizability and zero-shot capability.

cs.MA↗

arXiv2025

Computational Model for Photoionization in Pure SF6 Streamer at 1-15 atm

Photoionization plays a crucial role in achieving accurate quantitative predictions in SF6 streamer simulations, but accurate models for SF6 photoionization remains limited, motivating this paper. First, we develop a computational model for SF6 photoionization and provide the detailed theoretical modeling process, as well as comparison between experiment and simulation. A concise summary of model parameters within the comprehensive pressure range of 1 - 15 atm is provided for direct reference. Then, we perform comparative studies against simplified approaches. The results demonstrate that the proposed model effectively captures the non-local effects of SF6 photoionization, enhancing both the spatial numerical convergence and the accuracy of the streamer structure. Finally, we perform comparative studies by artificially increasing the photoionization intensity through multiplying the photoionization source term Sph by a factor of 50 (50*Sph) relative to the baseline intensity. Regarding breakdown voltage prediction, 50*Sph leads to a significant underestimation of the breakdown voltage for positive streamers, introducing errors greater than 0.5 kV, while exerting a small impact on negative streamers. Regarding streamer propagation dynamics, the radius of the positive streamer head exhibits pronounced shrinking, and 50*Sph reduces this shrinking and significantly lowers the head field by more than 700 Td. In contrast, 50*Sph has little impact on the morphology of the negative streamers and slightly enhances the head field by less than 30 Td.

physics.plasm-ph↗

arXiv2025

Improving Generalization in Intent Detection: GRPO with Reward-Based Curriculum Sampling

Intent detection, a critical component in task-oriented dialogue (TOD) systems, faces significant challenges in adapting to the rapid influx of integrable tools with complex interrelationships. Existing approaches, such as zero-shot reformulations and LLM-based dynamic recognition, struggle with performance degradation when encountering unseen intents, leading to erroneous task routing. To enhance the model's generalization performance on unseen tasks, we employ Reinforcement Learning (RL) combined with a Reward-based Curriculum Sampling (RCS) during Group Relative Policy Optimization (GRPO) training in intent detection tasks. Experiments demonstrate that RL-trained models substantially outperform supervised fine-tuning (SFT) baselines in generalization. Besides, the introduction of the RCS, significantly bolsters the effectiveness of RL in intent detection by focusing the model on challenging cases during training. Moreover, incorporating Chain-of-Thought (COT) processes in RL notably improves generalization in complex intent detection tasks, underscoring the importance of thought in challenging scenarios. This work advances the generalization of intent detection tasks, offering practical insights for deploying adaptable dialogue systems.

cs.CL↗

arXiv2025

SF6 streamer breakdown induced by floating linear metal particles: Following streamers and side streamers

The electrical breakdown of SF6 in the presence of floating metal particles is facilitated by two key factors: the role of floating metal particles and the nonlinear breakdown behavior of high-pressure SF6. However, the microscopic transient processes remain unclear, motivating this paper. Using 2D fluid models, we investigate SF6 streamer breakdown induced by a floating linear metal particle under negative applied voltage. First, We identify a characteristic double-end streamer inception in the combined gap. Then, we propose the following streamer (FS) mechanism to explain the metal particle's role. Two following streamers, FS1 and FS2, arise from the interaction between space charge and metal particle. FS1 facilitates breakdown via the negative space charge field generated by its head. FS2 facilitates breakdown by merging with FS1, accelerating its propagation and enhancing the electric field at the primary streamer head. Finally, we propose the side streamer (SS) mechanism to explain the nonlinear breakdown behavior of high-pressure SF6. The SS is identified as a new forward ionization wave that develops along the sides of the primary streamer, due to photoionization-driven negative ion accumulation. SS facilitates breakdown by merging with the primary streamer, increasing negative space charge and leading to three distinct propagation modes. Higher pressure increases the production rate of negative ions along the streamer sides, making SS more likely to form. Under overvoltage, the facilitating effect of SS diminishes as the background field (E/N)b strengthens, disappearing when (E/N)b exceeds 245 Td. This study provides new insights into the SF6 streamer breakdown mechanisms induced by floating metal particles and offers theoretical references for further investigation on the quantitative characterization.

physics.plasm-ph↗

arXiv2024

Radio Frequency Ray Tracing with Neural Object Representation

Radio frequency (RF) propagation modeling poses unique electromagnetic simulation challenges. While recent neural representations have shown success in visible spectrum rendering, the fundamentally different scales and physics of RF signals require novel modeling paradigms. In this paper, we introduce RFScape, a novel framework that bridges the gap between neural scene representation and RF propagation modeling. Our key insight is that complex RF-object interactions can be captured through object-centric neural representations while preserving the composability of traditional ray tracing. Unlike previous approaches that either rely on crude geometric approximations or require dense spatial sampling of entire scenes, RFScape learns per-object electromagnetic properties and enables flexible scene composition. Through extensive evaluation on real-world RF testbeds, we demonstrate that our approach achieves 13 dB improvement over conventional ray tracing and 5 dB over state-of-the-art neural baselines in modeling accuracy while requiring only sparse training samples.

eess.SP↗

arXiv2024

Spatial Fluctuation of the Electric Field within SF6 Streamer Channel in Highly Non-Uniform Fields: Phenomenon, Validation, and Mechanism

The electric field within the streamer channel is a critical parameter in the calculation model for the nonlinear breakdown voltage of SF6, motivating the research presented in this paper. By using a 2D fluid model, we investigate the microscopic characteristics of the SF6 streamer channel in highly non-uniform fields and uncover a previously unexplained coherent structure: the spatial fluctuation of the electric field (SFEF). We validate the physical validity of SFEF by modifying model parameters that could potentially introduce non-physical effects. Further comparative analysis reveals that SFEF is driven by an ion-conducting channel formed due to the strong electronegativity of SF6. This ion-conducting channel exhibits local characteristics, which fundamentally arise from the slow response of charged species to local charge relaxation. We identify that some charge separation originates from the accumulation of negative ions at the rear edge of the streamer head due to strong electric field shielding in this region. As the streamer propagates, charge separation is continuously generated and passively carried into the streamer channel, ultimately forming the SFEF. Finally, we confirm that SFEF does not occur in uniform fields, indicating that it is a phenomenon exclusive to highly non-uniform fields. These findings provide a deep insight into the electric field within the SF6 streamer channel and offer a potential avenue for further investigation into the mechanisms of SF6 nonlinear breakdown voltage.

physics.plasm-ph↗

arXiv2024

Microscopic characteristics of SF6 partial discharge induced by a floating linear metal particle

Direct current (DC) gas insulated transmission lines (GILs) have been widely used in power transmission, but might be threatened by partial discharge due to the presence of floating impurities (e.g., dust and metal particles) inside the sealed chamber. In this letter, by using a 2D fluid model we characterize the microscopic properties of the partial discharge induced by a floating linear metal particle in SF6 (both the discharge propagation and interaction between space charge and metal particle) under negative high voltage direct current (HVDC) conditions. Due to the strong electronegativity of SF6, the spatiotemporal distributions of the charged species (electrons, positive and negative ions), space charge, and reduced electric field are rather different from those in air. Notably, a negative ion region is observed around the top tip of the metal particle, and it plays an important role in the generation and propagation of primary and secondary streamers in SF6, which may lead to severe motion characteristics of the particle and aliasing of partial discharge signals. Additionally, we analyze the charging process and electric force reversal phenomenon, which may provide a more precise understanding of the underlying mechanisms of the firefly motion previously reported for DC GILs.

physics.plasm-ph↗