arXiv subjects

Kui Zhang

Publications and source records attributed to Kui Zhang.

At least 19 recordsLinked to original sources

arXiv2026

TPCD: Tone-Pressure Contrastive Decoding and the Label-Free Gating Bottleneck in Vision-Language Models

High-pressure prompts can push vision-language models (VLMs) into unsupported commitments, such as reading illegible text, reporting indeterminate times, or affirming absent objects. This paper asks whether the pressure-induced distribution itself can serve as a contrastive-decoding negative branch. Tone-pressure contrastive decoding (TPCD) subtracts logits produced under a high-pressure instruction from logits produced under a safe neutral instruction. On the 800-example tone-matters benchmark, LLaVA-1.5-7B under pressure reaches 66.75% attack success rate (ASR); safe neutralization reduces ASR to 9.88%; full TPCD reaches 0.50% but collapses positives to 15.56%. A benchmark-specific task-prior/disagreement gate preserves measured positive accuracy (54.44%) while lowering ASR to 1.63% on LLaVA. Treating this LLaVA analysis as the design split, full $n=800$ negative and $n=780$ matched-positive held-out runs on GLM-4.6V and Llama-3.2-Vision show that simple gates can improve over safe neutralization, with sensitivity analyses bounding the weak time-positive subtask. A category-prior-free answer-disagreement router reduces held-out aggregate ASR to 6.93%, improving over both safe neutralization (10.98%) and branch disagreement (9.67%) while matching branch disagreement's 79.94% positive accuracy, although it remains post-hoc and surface-form based. We conclude that pressure is a useful probe of commitment bias and a viable mitigation signal, but the current gates are not yet independently validated grounding-aware detectors.

cs.CV↗

arXiv2026

Artificial Intelligence Across the Cardiac Amyloidosis Diagnostic Pathway: From Single-Modality Detection to Multimodal Clinical Integration

Cardiac amyloidosis (CA) is increasingly recognized but remains substantially underdiagnosed, because its clinical and imaging phenotype overlaps with more common cardiomyopathies. Definitive subtype assignment and management further require integration of multimodal evidence to distinguish transthyretin from light chain disease. Machine learning and deep learning have been applied across the diagnostic and management pathway. These applications span ECG, echocardiography, and health record-based case finding, as well as CMR and nuclear interpretation, including SPECT/CT biomarker quantification, prognostic modeling, and treatment response assessment. This narrative review synthesizes these studies by clinical tasks, namely screening, detection, quantification, prognosis, and treatment response monitoring, rather than by input modality. This task-based organization clarifies why apparently similar AI models require different cohorts, reference standards, evaluation metrics, and implementation thresholds. The evidence reveals a maturity gradient. Binary detection and AI assisted quantification on bone scintigraphy and SPECT/CT are closest to clinical translation. Detection is supported by large externally validated cohorts, and quantification by interpretable, outcome linked measurement of myocardial tracer burden. By contrast, subtype aware classification, prognostic risk stratification, and treatment response monitoring remain at an early stage. These tasks are limited by small cohorts, enriched retrospective designs, heterogeneous labels, incomplete external validation, and uncertain calibration in realistic prevalence settings. Across tasks, high discrimination alone is insufficient.

physics.med-ph↗

arXiv2026

When Good Enough Is Optimal: Multiplication-Only Matrix Inversion Approximation for Quantized Gated DeltaNet

Matrix inversion in chunk-wise parallel linear attention is a major bottleneck for long-context modeling, particularly on NPUs, where forward-substitution-based methods exhibit limited parallelism and poor hardware utilization. We propose a fast, Matrix Multiplication (MatMul)-based algorithm tailored for strictly lower-triangular matrices arising in chunk-wise linear attention. Motivated by the rapid growth of Neumann-series terms and the diagonal concentration of the inverse matrix, we employ a truncated Neumann expansion with structural masking and parallel residual correction to eliminate sequential dependencies. We further extend our method to low-bits INT by mitigating the dynamic range expansion arising from repeated matrix power operations, and adapt the approximation order and residual step to the chunk size to minimize computational cost while preserving the model's accuracy. Experiments on Qwen3.5-family models demonstrate up to 5$\times$ kernel-level speedup and a 20% reduction in decode-layer overhead, while preserving accuracy under both floating-point and low-precision inference. Our method offers an efficient and hardware-friendly solution for scalable linear attention.

cs.LG↗

arXiv2026

DXA-Derived Skeletal Phenotypes and Hip Fracture Risk: A Backdoor-Adjusted Causal Analysis

Purpose: To compare dual-energy X-ray absorptiometry (DXA)-derived hip skeletal phenotypes in relation to hip fracture risk using prespecified confounder adjustment and to assess whether phenotypes ranked by their backdoor-adjusted average treatment effects (ATEs) improve risk stratification. Methods: We analyzed 21,098 UK Biobank participants with linked health records, hip DXA-derived skeletal measures, and prespecified covariates. Sixteen phenotypes spanning bone mineral content (BMC), bone mineral density (BMD), and T-score across hip-related regions were evaluated. Confounder selection was guided by a prespecified directed acyclic graph (DAG). Backdoor-adjusted ATEs were estimated on the absolute risk-difference scale per standard deviation (SD) increase. Effect heterogeneity was evaluated for total femur BMD, and downstream prediction was assessed using clinical variables combined with phenotypes ranked by ATE magnitude. Results: Among 21,098 participants, 115 had hip fractures. All 16 phenotypes showed negative backdoor-adjusted ATEs per SD increase. The largest ATEs were observed for total femur BMC and total femur BMD, each with a risk difference of -0.0047, corresponding to approximately 4.7 fewer hip fractures per 1,000 participants per SD higher phenotype value. Conditional effects of total femur BMD were stronger among older participants and those with lower BMI. In prediction, clinical variables plus the top 11 ATE-ranked phenotypes achieved higher AUC than FRAX with femoral neck BMD (0.842 vs. 0.709), with higher sensitivity (0.748 vs. 0.443) and similar specificity (0.793 vs. 0.777). Conclusion: DXA-derived hip skeletal phenotypes differed in their backdoor-adjusted ATEs. Phenotype-level causal evaluation may help identify informative DXA measures for risk stratification.

q-bio.QM↗

arXiv2026

Learning to Adapt SFT Data for Better Reasoning Generalization

Large language models (LLMs) have achieved remarkable progress, with post-training playing a crucial role in enhancing their reasoning capabilities. Among post-training paradigms, supervised fine-tuning (SFT) is widely used: it leverages external data to provide dense supervision and enables efficient training. However, directly fine-tuning on expert data can hurt generalization when the data distribution is mismatched with the target model's own distribution. In this work, we propose Data Adaptation for Reasoning Tuning (DART), which formulates the use of a fixed, potentially distributionally misaligned SFT dataset as an optimization problem over demonstration transformations. DART trains a mapper model with reinforcement learning to convert original SFT data into model-adapted supervision that better matches the target model's distribution and learning preferences. The transformed data are then used for SFT, allowing the target model to better exploit external supervision. Experiments across multiple models and datasets show that DART improves generalization, achieves higher training efficiency than direct RL, and helps models surpass standard SFT. Our code is available at https://anonymous.4open.science/r/DART525E50D.

cs.CL↗

arXiv2026

Reflect to Inform: Boosting Multimodal Reasoning via Information-Gain-Driven Verification

Multimodal Large Language Models (MLLMs) achieve strong multimodal reasoning performance, yet we identify a recurring failure mode in long-form generation: as outputs grow longer, models progressively drift away from image evidence and fall back on textual priors, resulting in ungrounded reasoning and hallucinations. Interestingly, Based on attention analysis, we find that MLLMs have a latent capability for late-stage visual verification that is present but not consistently activated. Motivated by this observation, we propose Visual Re-Examination (VRE), a self-evolving training framework that enables MLLMs to autonomously perform visual introspection during reasoning without additional visual inputs. Rather than distilling visual capabilities from a stronger teacher, VRE promotes iterative self-improvement by leveraging the model itself to generate reflection traces, making visual information actionable through information gain. Extensive experiments across diverse multimodal benchmarks demonstrate that VRE consistently improves reasoning accuracy and perceptual reliability, while substantially reducing hallucinations, especially in long-chain settings. Code is available at https://github.com/Xiaobu-USTC/VRE.

cs.CV↗

arXiv2026

TAMTRL: Teacher-Aligned Reward Reshaping for Multi-Turn Reinforcement Learning in Long-Context Compression

The rapid progress of large language models (LLMs) has led to remarkable performance gains across a wide range of tasks. However, when handling long documents that exceed the model's context window limit, the entire context cannot be processed in a single pass, making chunk-wise processing necessary. This requires multiple turns to read different chunks and update memory. However, supervision is typically provided only by the final outcome, which makes it difficult to evaluate the quality of memory updates at each turn in the multi-turn training setting. This introduces a temporal credit assignment challenge. Existing approaches, such as LLM-as-a-judge or process reward models, incur substantial computational overhead and suffer from estimation noise. To better address the credit assignment problem in multi-turn memory training, we propose Teacher-Aligned Reward Reshaping for Multi-Turn Reinforcement Learning (TAMTRL). TAMTRL leverages relevant documents as teacher signals by aligning them with each turn of model input and assigns rewards through normalized probabilities in a self-supervised manner. This provides fine-grained learning signals for each memory update and improves long-context processing. Experiments with multiple models of varying scales across seven long-context benchmarks show that TAMTRL consistently outperforms strong baselines, demonstrating its effectiveness. Our code is available at https://anonymous.4open.science/r/TAMTRL-F1F8.

cs.CL↗

arXiv2026

Improving Generalizability of Hip Fracture Risk Prediction via Domain Adaptation Across Multiple Cohorts

Clinical risk prediction models often fail to be generalized across cohorts because underlying data distributions differ by clinical site, region, demographics, and measurement protocols. This limitation is particularly pronounced in hip fracture risk prediction, where the performance of models trained on one cohort (the source cohort) can degrade substantially when deployed in other cohorts (target cohorts). We used a shared set of clinical and DXA-derived features across three large cohorts - the Study of Osteoporotic Fractures (SOF), the Osteoporotic Fractures in Men Study (MrOS), and the UK Biobank (UKB), to systematically evaluate the performance of three domain adaptation methods - Maximum Mean Discrepancy (MMD), Correlation Alignment (CORAL), and Domain - Adversarial Neural Networks (DANN) and their combinations. For a source cohort with males only and a source cohort with females only, domain-adaptation methods consistently showed improved performance than the no-adaptation baseline (source-only training), and the use of combinations of multiple domain adaptation methods delivered the largest and most stable gains. The method that combines MMD, CORAL, and DANN achieved the highest discrimination with the area under curve (AUC) of 0.88 for a source cohort with males only and 0.95 for a source cohort with females only), demonstrating that integrating multiple domain adaptation methods could produce feature representations that are less sensitive to dataset differences. Unlike existing methods that rely heavily on supervised tuning or assume known outcomes of samples in target cohorts, our outcome-free approaches enable the model selection under realistic deployment conditions and improve generalization of models in hip fracture risk prediction.

cs.LG↗

arXiv2025

An Advanced Two-Stage Model with High Sensitivity and Generalizability for Prediction of Hip Fracture Risk Using Multiple Datasets

Hip fractures are a major cause of disability, mortality, and healthcare burden in older adults, underscoring the need for early risk assessment. However, commonly used tools such as the DXA T-score and FRAX often lack sensitivity and miss individuals at high risk, particularly those without prior fractures or with osteopenia. To address this limitation, we propose a sequential two-stage model that integrates clinical and imaging information to improve prediction accuracy. Using data from the Osteoporotic Fractures in Men Study (MrOS), the Study of Osteoporotic Fractures (SOF), and the UK Biobank, Stage 1 (Screening) employs clinical, demographic, and functional variables to estimate baseline risk, while Stage 2 (Imaging) incorporates DXA-derived features for refinement. The model was rigorously validated through internal and external testing, showing consistent performance and adaptability across cohorts. Compared to T-score and FRAX, the two-stage framework achieved higher sensitivity and reduced missed cases, offering a cost-effective and personalized approach for early hip fracture risk assessment. Keywords: Hip Fracture, Two-Stage Model, Risk Prediction, Sensitivity, DXA, FRAX

cs.LG↗

arXiv2025

Decoupling Understanding from Reasoning via Problem Space Mapping for Small-Scale Model Reasoning

Despite recent advances in the reasoning capabilities of Large Language Models (LLMs), improving the reasoning ability of Small Language Models (SLMs, e.g., up to 1.5B parameters) remains challenging. A key obstacle lies in the complexity and variability of natural language: essentially equivalent problems often appear in diverse surface forms, often obscured by redundant or distracting details. This imposes a dual burden on SLMs: they must first extract the core problem from complex linguistic input, and then perform reasoning based on that understanding. The resulting vast and noisy problem space hinders optimization, particularly for models with limited capacity. To address this, we propose a new framework that decouples understanding from reasoning by mapping natural language problems into a canonical problem space-a semantically simplified yet expressive domain. This enables SLMs to focus on reasoning over standardized inputs, free from linguistic variability. Within this framework, we introduce DURIT (Decoupled Understanding from Reasoning via Iterative Training), a three-step algorithm that iteratively: (1) mapping natural language problems via reinforcement learning, (2) aligns reasoning trajectories through self-distillation, and (3) trains reasoning policies in the problem space. The mapper and reasoner are co-trained in an alternating loop throughout this process. Experiments show that DURIT substantially improves SLMs' performance on both in-domain and out-of-domain mathematical and logical reasoning tasks. Beyond improving reasoning capabilities, DURIT also improves the robustness of reasoning, validating decoupling understanding from reasoning as an effective strategy for strengthening SLMs.

cs.CL↗

arXiv2025

CoE-Ops: Collaboration of LLM-based Experts for AIOps Question-Answering

With the rapid evolution of artificial intelligence, AIOps has emerged as a prominent paradigm in DevOps. Lots of work has been proposed to improve the performance of different AIOps phases. However, constrained by domain-specific knowledge, a single model can only handle the operation requirement of a specific task,such as log parser,root cause analysis. Meanwhile, combining multiple models can achieve more efficient results, which have been proved in both previous ensemble learning and the recent LLM training domain. Inspired by these works,to address the similar challenges in AIOPS, this paper first proposes a collaboration-of-expert framework(CoE-Ops) incorporating a general-purpose large language model task classifier. A retrieval-augmented generation mechanism is introduced to improve the framework's capability in handling both Question-Answering tasks with high-level(Code,build,Test,etc.) and low-level(fault analysis,anomaly detection,etc.). Finally, the proposed method is implemented in the AIOps domain, and extensive experiments are conducted on the DevOps-EVAL dataset. Experimental results demonstrate that CoE-Ops achieves a 72% improvement in routing accuracy for high-level AIOps tasks compared to existing CoE methods, delivers up to 8% accuracy enhancement over single AIOps models in DevOps problem resolution, and outperforms larger-scale Mixture-of-Experts (MoE) models by up to 14% in accuracy.

cs.CL↗

arXiv2025

SyncLoop: A Multimodal Dual-Loop Framework for Self-Improving Mathematical Reasoning

Recent advances in multimodal large language models (MLLMs) have shown impressive reasoning capabilities. However, further enhancing existing MLLMs necessitates high-quality vision-language datasets with carefully curated task complexities, which are both costly and challenging to scale. Although recent self-improving models that iteratively refine themselves offer a feasible solution, they still suffer from two core challenges: (i) most existing methods augment visual or textual data separately, resulting in discrepancies in data complexity (e.g., over-simplified diagrams paired with redundant textual descriptions); and (ii) the evolution of data and models is also separated, leading to scenarios where models are exposed to tasks with mismatched difficulty levels. To address these issues, we propose C2-Evo, an automatic, closed-loop self-improving framework that jointly evolves both training data and model capabilities. Specifically, given a base dataset and a base model, C2-Evo enhances them by a cross-modal data evolution loop and a data-model evolution loop. The former loop expands the base dataset by generating complex multimodal problems that combine structured textual sub-problems with iteratively specified geometric diagrams, while the latter loop adaptively selects the generated problems based on the performance of the base model, to conduct supervised fine-tuning and reinforcement learning alternately. Consequently, our method continuously refines its model and training data, and consistently obtains considerable performance gains across multiple mathematical reasoning benchmarks. Our code, models, and datasets will be released.

cs.CV↗

arXiv2025

MoLink: Distributed and Efficient Serving Framework for Large Models

Large language models represent a groundbreaking shift in generative AI. Yet, these advances come with a significant challenge: the high cost of model serving. To mitigate these costs, consumer-grade GPUs emerge as a more affordable alternative. This presents an opportunity for more cost-efficient LLM serving by leveraging these GPUs. However, it is non-trivial to achieve high-efficiency LLM serving on consumer-grade GPUs, mainly due to two challenges: 1) these GPUs are often deployed in limited network conditions; 2) these GPUs often exhibit heterogeneity in host systems. To address these challenges, we present MoLink, a distributed LLM serving system for large models. It incorporates several key techniques, enabling efficient LLM serving on heterogeneous and weakly connected consumer-grade GPUs. Our experiments demonstrate that it achieves throughput improvements of up to 458\% and cost-profit margin improvements of up to 151\%, compared to state-of-the-art systems. MoLink allows users on Windows, Linux, and containerized VMs to seamlessly integrate GPUs with just a few lines of code over Ethernet or public networks. Currently, it supports 18 mainstream architectures of open-source large language models. The source code is publicly available https://github.com/oldcpple/MoLink.

cs.DC↗

arXiv2025

PaMi-VDPO: Mitigating Video Hallucinations by Prompt-Aware Multi-Instance Video Preference Learning

Direct Preference Optimization (DPO) helps reduce hallucinations in Video Multimodal Large Language Models (VLLMs), but its reliance on offline preference data limits adaptability and fails to capture true video-response misalignment. We propose Video Direct Preference Optimization (VDPO), an online preference learning framework that eliminates the need for preference annotation by leveraging video augmentations to generate rejected samples while keeping responses fixed. However, selecting effective augmentations is non-trivial, as some clips may be semantically identical to the original under specific prompts, leading to false rejections and disrupting alignment. To address this, we introduce Prompt-aware Multi-instance Learning VDPO (PaMi-VDPO), which selects augmentations based on prompt context. Instead of a single rejection, we construct a candidate set of augmented clips and apply a close-to-far selection strategy, initially ensuring all clips are semantically relevant while then prioritizing the most prompt-aware distinct clip. This allows the model to better capture meaningful visual differences, mitigating hallucinations, while avoiding false rejections, and improving alignment. PaMi-VDPOseamlessly integrates into existing VLLMs without additional parameters, GPT-4/human supervision. With only 10k SFT data, it improves the base model by 5.3% on VideoHallucer, surpassing GPT-4o, while maintaining stable performance on general video benchmarks.

cs.CV↗

arXiv2025

Exploiting Vulnerabilities in Speech Translation Systems through Targeted Adversarial Attacks

As speech translation (ST) systems become increasingly prevalent, understanding their vulnerabilities is crucial for ensuring robust and reliable communication. However, limited work has explored this issue in depth. This paper explores methods of compromising these systems through imperceptible audio manipulations. Specifically, we present two innovative approaches: (1) the injection of perturbation into source audio, and (2) the generation of adversarial music designed to guide targeted translation, while also conducting more practical over-the-air attacks in the physical world. Our experiments reveal that carefully crafted audio perturbations can mislead translation models to produce targeted, harmful outputs, while adversarial music achieve this goal more covertly, exploiting the natural imperceptibility of music. These attacks prove effective across multiple languages and translation models, highlighting a systemic vulnerability in current ST architectures. The implications of this research extend beyond immediate security concerns, shedding light on the interpretability and robustness of neural speech processing systems. Our findings underscore the need for advanced defense mechanisms and more resilient architectures in the realm of audio systems. More details and samples can be found at https://adv-st.github.io.

cs.SD↗

arXiv2025

Physics-informed deep learning for infectious disease forecasting

Accurate forecasting of contagious diseases is critical for public health policymaking and pandemic preparedness. We propose a new infectious disease forecasting model based on physics-informed neural networks (PINNs), an emerging scientific machine learning approach. By embedding a compartmental model into the loss function, our method integrates epidemiological theory with data, helping to prevent model overfitting. We further enhance the model with a sub-network that accounts for covariates such as mobility and cumulative vaccine doses, which influence the transmission rate. Using state-level COVID-19 data from California, we demonstrate that the PINN model accurately predicts cases, deaths, and hospitalizations, aligning well with existing benchmarks. Notably, the PINN model outperforms naive baseline forecasts and several sequence deep learning models, including Recurrent Neural Networks (RNNs), Long Short-Term Memory (LSTM) networks, Gated Recurrent Units (GRUs), and Transformers. It also achieves performance comparable to a sophisticated Gaussian infection state forecasting model that combines compartmental dynamics, a data observation model, and parameter regression. However, the PINN model features a simpler structure and is easier to implement. In summary, we systematically evaluate the PINN model's ability to forecast infectious disease dynamics, demonstrating its potential as an efficient computational tool to strengthen forecasting capabilities.

cs.LG↗

arXiv2024

Hierarchical Consensus-Based Multi-Agent Reinforcement Learning for Multi-Robot Cooperation Tasks

In multi-agent reinforcement learning (MARL), the Centralized Training with Decentralized Execution (CTDE) framework is pivotal but struggles due to a gap: global state guidance in training versus reliance on local observations in execution, lacking global signals. Inspired by human societal consensus mechanisms, we introduce the Hierarchical Consensus-based Multi-Agent Reinforcement Learning (HC-MARL) framework to address this limitation. HC-MARL employs contrastive learning to foster a global consensus among agents, enabling cooperative behavior without direct communication. This approach enables agents to form a global consensus from local observations, using it as an additional piece of information to guide collaborative actions during execution. To cater to the dynamic requirements of various tasks, consensus is divided into multiple layers, encompassing both short-term and long-term considerations. Short-term observations prompt the creation of an immediate, low-layer consensus, while long-term observations contribute to the formation of a strategic, high-layer consensus. This process is further refined through an adaptive attention mechanism that dynamically adjusts the influence of each consensus layer. This mechanism optimizes the balance between immediate reactions and strategic planning, tailoring it to the specific demands of the task at hand. Extensive experiments and real-world applications in multi-robot systems showcase our framework's superior performance, marking significant advancements over baselines.

cs.AI↗

arXiv2023

Segue: Side-information Guided Generative Unlearnable Examples for Facial Privacy Protection in Real World

The widespread use of face recognition technology has given rise to privacy concerns, as many individuals are worried about the collection and utilization of their facial data. To address these concerns, researchers are actively exploring the concept of ``unlearnable examples", by adding imperceptible perturbation to data in the model training stage, which aims to prevent the model from learning discriminate features of the target face. However, current methods are inefficient and cannot guarantee transferability and robustness at the same time, causing impracticality in the real world. To remedy it, we propose a novel method called Segue: Side-information guided generative unlearnable examples. Specifically, we leverage a once-trained multiple-used model to generate the desired perturbation rather than the time-consuming gradient-based method. To improve transferability, we introduce side information such as true labels and pseudo labels, which are inherently consistent across different scenarios. For robustness enhancement, a distortion layer is integrated into the training pipeline. Extensive experiments demonstrate that the proposed Segue is much faster than previous methods (1000$\times$) and achieves transferable effectiveness across different datasets and model architectures. Furthermore, it can resist JPEG compression, adversarial training, and some standard data augmentations.

cs.CR↗