Vision-Language Models (VLMs) extend Large Language Models (LLMs) with visual perception capabilities, unlocking broad applications across many domains. However, ensuring their safety remains a critical challenge, as adversarial visual inputs can easily bypass built-in safeguards and elicit harmful content. In this paper, we uncover a phenomenon we call delayed safety awareness, where a jailbroken VLM initially produces harmful content but ultimately recognizes the harmfulness at the end of the generation process. We attribute this phenomenon to the fact that the model's safety awareness against jailbreaks cannot be effectively transferred to the intermediate stages of text generation. Motivated by this insight, we introduce SafetyReminder, a simple yet effective defense that optimizes a learnable soft prompt using our proposed Safety-Activation Prompt Tuning (SAPT). This soft prompt is inserted into the generated text to activate the safety awareness of the model, steering it toward refusal when harmful content arises while preserving helpfulness in benign scenarios. We evaluate our method on three established harmful benchmarks and across three types of adversarial attacks. Experimental results demonstrate that our method achieves state-of-the-art defense performance with strong generalization, offering a practical and lightweight solution for safe deployment of VLMs.

As deep neural networks (DNNs) continue to grow in scale and complexity, GPU memory limitations have become a significant challenge for DNN model training, especially on resource-constrained commercial GPUs. While model quantization facilitates memory-efficient training, it often necessitates a tradeoff between quantization granularity and model accuracy. And quantization imposes additional computational overhead, which adversely affects the training throughput and apportions out the performance gains it brings. In this article, we propose FDSR, an adaptive tensor quantization method that leverages frequency domain division and similarity-based data reuse to break the memory bottleneck in visual model training. FDSR leverages the frequency-domain characteristics of tensors in terms of memory consumption and model accuracy, and proposes a fine-grained tensor quantization with different quantization bit-widths. It adaptively optimizes the quantization parameters according to model accuracy during training while employing sparsification according to data frequency-domain features, minimizing memory consumption and accuracy loss. To counteract the computational cost, FDSR incorporates a novel similarity-based reuse strategy that avoids redundant quantization/dequantization computations, further enhanced by a tailored Locality-Sensitive Hashing (LSH) mechanism and optimized kernels. Experimental results demonstrate that FDSR achieves an average of 10.20× activation memory compression with only 1.10% average accuracy loss across various models on the commercial GPU. Compared to the state-of-the-art quantization methods, FDSR improves memory optimization by up to 68.6% and increases throughput by up to 25.55%, with consistent performance improvements on different GPU architectures.

Testing in the closed field is one of the important means to verify the features and performance of automated vehicles. Due to the complicated and changeable traffic conditions, how to design the testing scenarios with relatively few number but significant value is a problem that deserves research. We analyze the features of the Level 2 automatic driving production models in the market. Based on the applicable scenarios of the main functions such as adaptive cruise control, active lane changing control and active lane keeping control for automated vehicles, the relative location among the ego-vehicle and its surrounding obstacle vehicles are permutated and combined to form a vehicle combination groups, and then for each vehicle combination, the directions of motion of the ego-vehicle and obstacle vehicles are arranged and combined to obtain a possible test scenarios groups. Based up the test scenario generated by the combination of the ego-vehicle and the obstacle vehicles, the valuable and simple primary test scenario is generated through the analysis of the scenario importance, and then an obstacle vehicle is added to form the new scenario of higher level. The valuable test scenarios with various levels are screened out by analyzing the new impacts of detection and response of the ego-vehicle exerted by the movement of the obstacle vehicles and evaluating the importance of the scenarios. Finally, total test scenario groups with test significance is obtained. The validity of this method is verified, and a test case for parameter design is given in the paper.

Nanonetworks composed by communicating nano-devices enable new applications in the consumer, biomedical, and environmental fields. Three main characteristics introduce strict requirements for routing protocols design for nanonetworks, namely, short transmission range at Terahertz (THz) frequency (0.1-10 THz), fluctuations in the energy of nano-nodes due to the energy harvesting processes and very limited memory/buffer size of nano-nodes. In this paper, a multi-hop deflection routing algorithm based on Q-learning for energy-harvesting nanonetworks (MDRQEN) is proposed to guarantee the network energy efficiency, while ensuring a low packet loss probability. First, a deflection table is introduced to deflect the packets when the next hop nano-nodes are unavailable due to energy or memory/buffer constraints. Then, a Q-learning scheme is proposed to update the routing table and deflection table by utilizing the reward information contained in the forwarded packet from the previous nano-node. In the Q-learning update scheme, packet deflection ratio, packet loss ratio, packet hop count and node energy status of nano-nodes are taken into consideration. As numerically shown through extensive simulations in Network Simulator 3 (NS-3), the proposed MDRQEN algorithm can achieve a better packet delivery ratio and energy efficiency than random routing algorithm, flooding routing algorithm and the MDRQEN algorithm without the Q-learning update scheme.

For a set of real-time tasks that have mixed preference of being executed at early or late times before their deadlines, we have recently studied both earliest-deadline based and fixed-priority preference-oriented (PO) scheduling algorithms for uniprocessor systems. In this work, focusing on multiprocessor real-time systems, we study the foundational guidelines to design partition-based PO scheduling algorithms for tasks with mixed preference requirements. In particular, through a concrete example, we illustrate that the harmonicity of tasks' periods should be incorporated when making scheduling decisions in addition to their execution preferences to obtain favorable schedules that better fulfill tasks' preference requirements. Based on such guidelines, we design a period-aware preference-oriented (PAPO) partitioned scheduling algorithm and discuss several variations by considering harmonicity as well as utilization of tasks.