MOTIVATION: MicroRNAs (miRNAs) are key post-transcriptional regulators involved in diverse biological processes, and their dysregulation is closely associated with the onset and progression of many diseases. Accurate prediction of miRNA-disease association types is therefore essential for understanding disease mechanisms and advancing precision medicine. Although computational methods provide efficient alternatives to wet-lab experiments, existing approaches often focus on binary association prediction, inadequately integrate local semantic dependencies and global topological structures, and suffer from class imbalance. RESULTS: To address these limitations, we propose MRGBMDAT, a multi-relational graph encoder network with bilinear fusion for miRNA-disease association type prediction. Specifically, a multi-relational graph convolution module with bidirectional cross-attention captures global topological structures, while a local subgraph sampling module extracts local semantic dependencies. A bilinear fusion decoder with element-wise attention jointly models their linear and nonlinear interactions. In addition, an iterative feature similarity-based negative sample selection strategy is introduced to alleviate class imbalance. Experimental results on the HMDD v3.2 dataset demonstrate that MRGBMDAT significantly outperforms five state-of-the-art methods across multiple evaluation metrics, exhibiting strong discriminative power and generalization capability. AVAILABILITY AND IMPLEMENTATION: The source code is publicly available at https://github.com/CDMBlab/MRGBMDAT.

Abstract Partial neural distinguishers limit the available ciphertext bit combinations in differential neural cryptanalysis. When the training data size and the number of bits are not appropriately selected, label collisions can occur, which adversely affects key recovery efficiency. This paper conducts an analysis to investigate the correlation between the number of bits and the data size, aiming to address the aforementioned issue. It develops a strategy to control collisions and mitigate the impact of these collisions on model performance. A Collision-Aware Key Recovery (CAKR) framework is proposed tailored for high-collision data based on this strategy. This framework leverages the distribution characteristics of labels, eliminating the need for training neural distinguishers and significantly reducing both time and resource consumption. Experimental results show that the CAKR framework reduces the key recovery time by 96.8%, 95.5%, and 91.0% for the Speck32/64, Speck64/96, and Speck96/128, respectively. Additionally, a bit search algorithm is proposed that incorporates a differential evolution strategy and uses the non-uniformity of the ciphertext difference distribution among positive samples as the fitness criterion. Frequent calls to the neural distinguisher are avoided by our method, reducing the search time from 3.286 h to 7.464 s for 8-bit combinations in Speck32/64. The CAKR framework also offers a quantum version that theoretically further reduces time complexity.

Current long context large language models (LLMs) can process inputs up to 100,000 tokens, yet struggle to generate outputs exceeding even a modest length of 2,000 words. Through controlled experiments, we find that the model's effective generation length is inherently bounded by the sample it has seen during supervised fine-tuning (SFT). In other words, their output limitation is due to the scarcity of long-output examples in existing SFT datasets. To address this, we introduce AgentWrite, an agent-based pipeline that decomposes ultra-long generation tasks into subtasks, enabling off-the-shelf LLMs to generate coherent outputs exceeding 20,000 words. Leveraging AgentWrite, we construct LongWriter-6k, a dataset containing 6,000 SFT data with output lengths ranging from 2k to 32k words. By incorporating this dataset into model training, we successfully scale the output length of existing models to over 10,000 words while maintaining output quality. We also develop LongBench-Write, a comprehensive benchmark for evaluating ultra-long generation capabilities. Our 9B parameter model, further improved through DPO, achieves state-of-the-art performance on this benchmark, surpassing even much larger proprietary models. In general, our work demonstrates that existing long context LLM already possesses the potential for a larger output window--all you need is data with extended output during model alignment to unlock this capability.

The Fiber Bragg grating (FBG) three-axial force sensor provides force feedback for an endoscopic surgical robot, reducing operational difficulty and risks. However, the packaging method of the optical fiber sensor demonstrates limited adaptability to both high-temperature sterilization environments and the wet operative areas encountered during surgery. Based on this, this paper presents step-reduced FBG and enclosed three-axial force sensor. The sensor adopts an integrated design with a maximum outer diameter of 4.5 mm and can be seamlessly integrated into the end of the flexible endoscopic surgical robot. At the same time, the hydrofluoric acid corrosion process is introduced to obtain the twin reflection spectrum and realize the decoupling of three-axial forces and temperature. Static calibration demonstrates etched grating sensitivities of 227.78 pm/N (Fx), 242.63 pm/N (Fy), and 233.50 pm/N (Fz) via least-squares fitting. Force-temperature coupling experiment confirms maximum full-scale force errors remain below 5% under temperature perturbation, verifying the reliability of the sensor. Finally, the high-temperature sterilization experiment at 180°C was conducted, demonstrating the designed sensor’s thermal stability under medical device sterilization protocols.

Test-time reasoning algorithms such as chain-of-thought, self-consistency, and MCTS enhance LLM problem-solving but can wastefully generate many tokens without improving accuracy. At the same time, we observe that these algorithms exhibit answer stabilization: their intermediate solutions often cease to change after a certain point, and further investment of compute does not change their final answer. To quantify this phenomenon, we introduce Certaindex, an algorithm-agnostic metric measuring this evolving stability, signaling when further computation is unlikely to alter the final result. Certaindex is lightweight, can accelerate reasoning program inference via early exit, and further enables dynamic token allocation, gang scheduling, and many opportunities when integrated with real-world LLM serving systems. To quantify real-world benefits, we built Certaindex as a scheduler into Dynasor, our reasoning-aware LLM serving system, and demonstrate up to 50\% compute savings and 3.3$\times$ higher throughput in real workloads with no accuracy drop. Our code is available at https://github.com/hao-ai-lab/Dynasor.git

In Large Language Model (LLM) inference, the output length of an LLM request is typically regarded as not known a priori. Consequently, most LLM serving systems employ a simple First-come-first-serve (FCFS) scheduling strategy, leading to Head-Of-Line (HOL) blocking and reduced throughput and service quality. In this paper, we reexamine this assumption -- we show that, although predicting the exact generation length of each request is infeasible, it is possible to predict the relative ranks of output lengths in a batch of requests, using learning to rank. The ranking information offers valuable guidance for scheduling requests. Building on this insight, we develop a novel scheduler for LLM inference and serving that can approximate the shortest-job-first (SJF) schedule better than existing approaches. We integrate this scheduler with the state-of-the-art LLM serving system and show significant performance improvement in several important applications: 2.8x lower latency in chatbot serving and 6.5x higher throughput in synthetic data generation. Our code is available at https://github.com/hao-ai-lab/vllm-ltr.git

Incompleteness of domain ontology and unavailability of some values are two inevitable problems of dialogue state tracking (DST).Existing approaches generally fall into two extremes: choosing models without ontology or embedding ontology in models leading to over-dependence.In this paper, we propose a new architecture to cleverly exploit ontology, which consists of Slot Attention (SA) and Value Normalization (VN), referred to as SAVN.Moreover, we supplement the annotation of supporting span for MultiWOZ 2.1, which is the shortest span in utterances to support the labeled value.SA shares knowledge between slots and utterances and only needs a simple structure to predict the supporting span.VN is designed specifically for the use of ontology, which can convert supporting spans to the values.Empirical results demonstrate that SAVN achieves the state-of-the-art joint accuracy of 54.52% on MultiWOZ 2.0 and 54.86% on MultiWOZ 2.1.Besides, we evaluate VN with incomplete ontology.The results show that even if only 30% ontology is used, VN can also contribute to our model.

An ultra-wideband low phase variation variable gain amplifier (VGA) in 0.18/am BiCMOS process with high linearity is presented in this paper. The proposed VGA uses novel current steering gain blocks to achieve gain steps with low phase variation. The VGA has a measured gain range of 6.3-8.1 dB over the entire frequency range of 12-40 GHz and shows a phase variation of 0.2-0.78 °/dB. The VGA achieves a simulated input P1dB of 0 dBm at 26 GHz while consuming only 20.5 mW from 1.5 V power supply and occupies an active area of just 0.05 mm2. The VGA shows 7.6 times better Figure of Merit as compared to current state-of-the-art VGAs.