Ubiquitous Energy (United States)

Vision-Language Models (VLMs) have achieved impressive performance across various tasks, but often struggle to apply newly introduced visual concepts during inference. A common failure pattern is what we call Mixing Things Up: VLMs frequently confuse concept names, resulting in vague descriptions and failure to ground the concept correctly. Existing approaches mainly address person-related concepts through text prompts or tokenizer modifications. However, VLMs still miss or misinterpret untrained visual concepts, underscoring the need to learn new concepts directly from visual input, without relying on prior textual injection. To overcome these limitations, we propose BISCUIT (Basis-aligned Inference through Structured Concept Unification and Identification-aware Tuning), a two-step training method. Step I proposes a dual-stream structure-aware vision encoder that fuses RGB and edge-based embeddings within a shared basis space to enhance concept recognition. Step II enhances generation quality through identification-aware tuning, which encourages alignment between the generated text and the newly introduced visual concepts. Existing methods mainly focus on person concepts and lack comprehensive evaluation across diverse visual categories. We further propose a benchmark BiscuitVQA to evaluate VLMs performance on recognizing and applying novel image-introduced concepts across diverse concept types and task types, including real people, cartoons, animals, and symbolic content. We apply BISCUIT to LLaVA-1.5 and Qwen2.5-VL, achieving competitive results among open-source models and narrowing the gap to Gemini-2.5 and GPT-4o. Interestingly, our BISCUIT maintains strong generalization, showing minimal degradation on other downstream tasks.

In software engineering, empirical studies on automated fault localization (FL) methods mainly focus on general software, and substantial progress has been made. However, the applicability and efficacy of these methods in specialized, domain-specific software like industrial software remains under-explored. Such specialized software is usually characterized by complex inputs and iterative computing paradigms, which could significantly influence the effectiveness of existing FL methods. To address this gap, this study takes a typical categorical of industrial software (i.e., computer-aided engineering (CAE) projects) as a case study, to investigate the feasibility and effectiveness of state-of-the-art FL methods within CAE projects. Through the reproduction of 76 real-world bugs from three widely used CAE projects (i.e., FDS, deal.II, and MFEM), we find that even the most precise FL methods require developers to examine on average 467.18 statements before finding bugs and can take 208.13 hours to execute. The complex inputs and long-term computation characteristics of CAE projects further increase the difficulty of FL. Moreover, FL on CAE also faces challenges, such as insufficient differentiation of coverage information and missing CAE-specific FL features. Based on our findings, we improve FL on CAE projects by proposing a set of CAE main module-based features, which improve the best-performed FL method in this study (i.e., DeepFL) by 35.93% and 45%, in terms of MAR and MFR , respectively.

As demand grows for multigigabit data links, high-resolution radar, and precision sensing, operating frequencies are being driven deep into the sub-terahertz (sub-THz) spectrum, where the <italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">D</i>- and <italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">G</i>-bands (110–220 GHz) offer immense bandwidth (BW). However, operating at these frequencies with CMOS silicon technologies reveals several fundamental physical limitations, including limited breakdown voltage, low intrinsic gain, dominant parasitics, and significant passive losses. This article brings together different design methodologies that push these limits and expand the achievable power, gain, and BW of CMOS power amplifiers (PAs) operating near <inline-formula xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink"><tex-math notation="LaTeX">${f}_{\max}$</tex-math></inline-formula> The first approach revisits the concept of transistor stacking from a large-signal modeling perspective. By analytically accounting for all parasitic elements and the intrinsic delay of each device, a systematic framework is developed to determine the optimum input, output, and gate impedances for every transistor in the stack. A 45-nm SOI prototype that combines multiple stacked stages through a low-loss 4:1 combiner demonstrates 18.7 dBm saturated output power and 4.8% power-added efficiency (PAE) at 200 GHz, an example of how careful modeling can turn CMOS voltage limits into a scalable power generation strategy. The second methodology focuses on gain enhancement through embedding. At sub-THz frequencies, where matching-network losses can rival transistor gain, a passive feedback network is introduced and tuned on a “gain plane” that maps equi-gain and equi-load-conductance contours. This graphical design tool allows the simultaneous optimization of gain and output power. Implemented in a 65-nm CMOS 2 × 8 cell amplifier, the embedded approach achieves 19.5 dB gain and 9.4 dBm saturated power at 200 GHz, showing how modest feedback can recover otherwise-lost gain and extend usable output levels. The third technique, reverse-feedback amplification, breaks the BW barrier. By introducing a controlled amount of negative resistance into each amplifier cell, a distributed line is formed that supports constructive wave propagation and continuous power addition without interstage matching. Two 45-nm CMOS <italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">G</i>-band amplifiers fabricated using this principle deliver up to 20.4 dB gain and 3.8% PAE across a 156- to 194-GHz 3-dB BW. Together, these methodologies, large-signal stacking, gain plane embedding, and distributed reverse feedback, demonstrate how thoughtful circuit architectures can overcome fundamental device limits to achieve high-power, high-efficiency, and broadband amplification in standard CMOS.

To ensure safe driving in dynamic environments, autonomous vehicles should possess the capability to accurately predict lane change intentions of surrounding vehicles in advance and forecast their future trajectories. Existing motion prediction approaches have ample room for improvement, particularly in terms of long-term prediction accuracy and interpretability. In this study, we address these challenges by proposing a Lane Change-Large Language Model (LC-LLM), an explainable lane change prediction model that leverages the strong reasoning capabilities and self explanation abilities of Large Language Models (LLMs). Essentially, we reformulate the lane change prediction task as a language modeling problem, processing heterogeneous driving scenario information as natural language prompts for LLMs and employing supervised fine-tuning to tailor LLMs specifically for lane change prediction task. Additionally, we finetune the Chain-of-Thought (CoT) reasoning to improve prediction transparency and reliability, and include explanatory requirements in the prompts during the inference stage. Therefore, our LC-LLM not only predicts lane change intentions and trajectories but also provides CoT reasoning and explanations for its predictions, enhancing its interpretability. Extensive experiments based on the large-scale highD dataset demonstrate the superior performance and interpretability of our LC-LLM in lane change prediction task. To the best of our knowledge, this is the first attempt to utilize LLMs for predicting lane change behavior. Our study shows that LLMs can effectively encode comprehensive interaction information for understanding driving behavior.

Deep learning (DL)-based channel estimation paradigm fundamentally requires balancing two objectives: environment-specific reconstruction and cross-scenario generalization. This paper proposes a novel DL architecture for efficient channel estimation, which features dedicated neural networks for channel reconstruction and distribution generalization. We first propose the dual-attention-aided super-resolution neural network (DA-SRNN) for efficient time-frequency channel reconstruction from pilot measurements. Specifically, the channel-spatial attention module is designed to sequentially generate dimension-specific attention maps that capture the underlying channel correlations. Then, we incorporate continual learning (CL) strategies to maintain model plasticity across varying channel environments. Specifically, elastic weight consolidation (EWC) regularizes the reconstruction loss function, which imposes Fisher-information-weighted constraints on critical network parameters to stabilize knowledge retention during channel distribution shifts. Meanwhile, the corresponding training process is provided in detail. Evaluations under 3rd Generation Partnership Project (3GPP) channel models demonstrate the proposed scheme's superiority, which achieves substantial gains in both reconstruction accuracy and distribution generalization over state-of-the-art baselines.

In recent years, with the widespread application of robotics, designing efficient and generalizable robot control algorithms has become a critical research topic. Meta reinforcement learning as a method that enhances an agent's adaptability through cross-task learning, has demonstrated significant potential in robotic tasks. However, traditional meta reinforcement learning algorithms often rely on manually designed network architectures, which limit performance improvement and task adaptability. To address this issue, this paper proposes a meta reinforcement learning algorithm based on Neural Architecture Search. This method leverages robot simulation environments to automatically explore the optimal neural network architectures for meta reinforcement learning algorithms within the search space. By incorporating neural architecture search, the proposed algorithm aims to enhance their performance and efficiency. This study focuses on the design of the search space and search strategies tailored to meta reinforcement learning algorithms. Experimental results demonstrate that the proposed algorithm significantly improves task adaptability and training efficiency while reducing the labor costs associated with network architecture design, further demonstrating that this approach offers a new perspective for addressing robot control problems. It also establishes a solid foundation for future applications in real-world robotic systems.

This paper proposes a generative artificial intelligence-driven cross-modal semantic communication framework (GC-SemCom) designed for efficient image transmission. Specifically, at the transmitter, a multi-modal large language model is utilized to extract compact semantic information, i.e. image difference captions between the original image and retrieved background knowledge. This semantic information, expressed in textual instruction, is then transmitted based on joint source-channel codec. At the receiver, the diffusion model, guided by both received semantic information and synchronized background knowledge, iteratively generates the recovered image through the reverse denoising process. Furthermore, the semantic knowledge base provides background knowledge for both transceivers, defined as the semantically similar image retrieved. Experimental results demonstrate that the proposed framework reduces the overall data size to just 0.06% of the original data, while maintaining high-quality transmission even under substantial noise conditions.