Recent advancements in improving the reasoning capabilities of Large Language Models have underscored the efficacy of Process Reward Models (PRMs) in addressing intermediate errors through structured feedback mechanisms. This study analyzes PRMs from multiple perspectives, including training methodologies, scalability, and generalization capabilities. We investigate the interplay between pre-training and reward model training FLOPs to assess their influence on PRM efficiency and accuracy in complex reasoning tasks. Our analysis reveals a pattern of diminishing returns in performance with increasing PRM scale, highlighting the importance of balancing model size and computational cost. Furthermore, the diversity of training datasets significantly impacts PRM performance, emphasizing the importance of diverse data to enhance both accuracy and efficiency. We further examine test-time scaling strategies, identifying Monte Carlo Tree Search as the most effective method when computational resources are abundant, while Best-of-N Sampling serves as a practical alternative under resource-limited conditions. Notably, our findings indicate that PRMs trained on mathematical datasets exhibit performance comparable to those tailored for code generation, suggesting robust cross-domain generalization. Employing a gradient-based metric, we observe that PRMs exhibit a preference for selecting responses with similar underlying patterns, further informing their optimization.

Vision-Language-Action (VLA) models revolutionize robotic systems by enabling end-to-end perception-to-action pipelines that integrate multiple sensory modalities, such as visual signals processed by cameras and auditory signals captured by microphones. This multi-modality integration allows VLA models to interpret complex, real-world environments using diverse sensor data streams. Given the fact that VLA-based systems heavily rely on the sensory input, the security of VLA models against physical-world sensor attacks remains critically underexplored. To address this gap, we present the first systematic study of physical sensor attacks against VLAs, quantifying the influence of sensor attacks and investigating the defenses for VLA models. We introduce a novel ``Real-Sim-Real" framework that automatically simulates physics-based sensor attack vectors, including six attacks targeting cameras and two targeting microphones, and validates them on real robotic systems. Through large-scale evaluations across various VLA architectures and tasks under varying attack parameters, we demonstrate significant vulnerabilities, with susceptibility patterns that reveal critical dependencies on task types and model designs. We further develop an adversarial-training-based defense that enhances VLA robustness against out-of-distribution physical perturbations caused by sensor attacks while preserving model performance. Our findings expose an urgent need for standardized robustness benchmarks and mitigation strategies to secure VLA deployments in safety-critical environments.

The rapid development of large language models (LLMs) has led to their increasing deployment on edge devices for applications such as autonomous driving. These edge deployments enable efficient localized processing while enhancing user privacy. However, unintentional leaks of model families and hyperparameters during LLM execution serve as critical attack vectors, facilitating high-level threats like model extraction and membership inference. In this paper, we propose ArchSnoop, which is a non intrusive edge LLM architecture eavesdropping attack based on electromagnetic (EM) leakage. Our key insight is that the EM signals emitted by GPU and memory activities during inference reflect hierarchical architectural features ranging from token generation to individual linear projections. We designed a two-stage hierarchical reconstruction model to recover fine-grained architectural information from these EM signals. Our evaluation on the NVIDIA Jetson Orin Nano platform demonstrates that ArchSnoop achieves high accuracy in architecture reconstruction, including 99.12% in model family classification and 97.11% for hyperparameter estimation. We reveal the mapping between EM signals and LLM architectures on edge devices for the first time. By proposing potential countermeasures such as physical EM shielding and software-level perturbations, this work provides a new dimension to secure edge computing platforms against physical threats.

This paper analyzes the security of Internet of Things (IoT) devices from the perspective of sensing, actuating, and communicating. Particularly, we discover a vulnerability in power supply modules and propose VoltSiren attacks. To launch a VoltSiren attack, attackers may compromise the power source and inject malicious signals through the power supply module, which is indispensable in most devices. Consequently, VoltSiren attacks can cause sensor measurements irrelevant to reality, maneuver actuators in a way disregarding the desired command, or disrupt communications. To understand VoltSiren, we systematically analyze the underlying principle of power supply signals affecting the electronic components, which are building blocks to constitute the sensors, actuators, or communication modules. Based on these findings, we implement and validate VoltSiren on off-the-shelf products: six sensors, three actuators, and two communication modules, which are used in applications ranging from automobile braking systems, industrial process control to robotic arms. The root cause of this vulnerability lies in the common belief that noises from the power line are unintentional, and our work aims to call for attention to enhancing the security of power supply modules and adding countermeasures to mitigate the attacks.

Large Vision Language Models (VLMs) have long struggled with spatial reasoning tasks. Surprisingly, even simple spatial reasoning tasks, such as recognizing “under” or “behind” relationships between only two objects, pose significant challenges for current VLMs. We believe it is crucial to use the lens of mechanism interpretability, opening up the model and diving into model’s internal states to examine the interactions between image and text tokens during spatial reasoning. Our analysis of attention behaviors reveals significant differences in how VLMs allocate attention to image versus text. By tracing the areas of images that receive the highest attention scores throughout intermediate layers, we observe a notable pattern: errors often coincide with attention being misdirected towards irrelevant objects within the image. Moreover, such attention patterns exhibit substantial differences between familiar (e.g., “on the left side of ”) and unfamiliar (e.g.,“in front of ”) spatial relationships. Motivated by these findings, we propose ADAPTVIS based on inference-time confidence scores to sharpen the attention on highly relevant regions when the model exhibits high confidence, while smoothing and broadening the attention window to consider a wider context when confidence is lower. This training-free decoding method shows significant improvement (e.g., up to a 50 absolute point improvement) on spatial reasoning benchmarks such as WhatsUp and VSR with negligible additional cost.

Unauthorized covert voice recording presents a significant threat to privacy-sensitive scenarios, such as confidential meetings and private conversations. Due to their miniaturization and disguise characteristics, hidden voice recorders are difficult to notice. In this article, we present DeHiREC , the first proof-of-concept system capable of detecting offline hidden voice recorders from their electromagnetic radiations (EMR). We first characterize the unique patterns of the emanated EMR signals and then locate the EMR source, i.e., the analog-to-digital converter module embedded in the mixed signal system-on-chips. Since these unintentional EMR signals can be extremely noisy and weak, accurately detecting them can be challenging. To address this challenge, we design an EMR Catalyzing method to actively stimulate the EMR signals and then employ an adaptive-folding algorithm to improve the signal-to-noise ratio of the sensed EMRs. We evaluate the performance of DeHiREC on 18 commercial voice recorders under various impacts, including interference from other devices. Experimental results reveal that DeHiREC is effective in detecting all 18 voice recorders and achieves an overall success rate of 94.72% and a recall rate of 92.03% at a distance of 0.2 m.

Unauthorized covert voice recording brings a remarkable threat to privacy-sensitive scenarios, such as confidential meetings and private conversations. Due to the miniaturization and disguise characteristics, hidden voice recorders are difficult to be noticed in their surroundings. In this paper, we present DeHiREC, the first proof-of-concept system that can detect offline hidden voice recorders from their electromagnetic radiations (EMR). We first characterize the unique patterns of the emanated EMR signals and then locate the EMR source, i.e., the analog-to-digital converter (ADC) module embedded in the mixed signal system-on-chips (MSoCs). Since these unintentional EMR signals can be extremely noisy and weak, accurately detecting them can be challenging. To address this challenge, we first design an EMR Catalyzing method to stimulate the EMR signals actively and then employ an adaptive-folding algorithm to improve the signal-to-noise ratio (SNR) of the sensed EMRs. Once the sensed EMR variation corresponds to our active stimulation, we can determine that there exists a hidden voice recorder. We evaluate the performance of DeHiREC on 13 commercial voice recorders under various impacts, including interference from other devices. Experimental results reveal that DeHiREC is effective in detecting all 13 voice recorders and achieves an overall success rate of 92.17% and a recall rate of 86.14% at a distance of 0.2 m.