arXiv subjects

Yu Tian

Publications and source records attributed to Yu Tian.

At least 145 records · Page 8Linked to original sources

arXiv2025

MultiMedEdit: A Scenario-Aware Benchmark for Evaluating Knowledge Editing in Medical VQA

Knowledge editing (KE) provides a scalable approach for updating factual knowledge in large language models without full retraining. While previous studies have demonstrated effectiveness in general domains and medical QA tasks, little attention has been paid to KE in multimodal medical scenarios. Unlike text-only settings, medical KE demands integrating updated knowledge with visual reasoning to support safe and interpretable clinical decisions. To address this gap, we propose MultiMedEdit, the first benchmark tailored to evaluating KE in clinical multimodal tasks. Our framework spans both understanding and reasoning task types, defines a three-dimensional metric suite (reliability, generality, and locality), and supports cross-paradigm comparisons across general and domain-specific models. We conduct extensive experiments under single-editing and lifelong-editing settings. Results suggest that current methods struggle with generalization and long-tail reasoning, particularly in complex clinical workflows. We further present an efficiency analysis (e.g., edit latency, memory footprint), revealing practical trade-offs in real-world deployment across KE paradigms. Overall, MultiMedEdit not only reveals the limitations of current approaches but also provides a solid foundation for developing clinically robust knowledge editing techniques in the future.

cs.AI↗

arXiv2025

Learning to Perform Low-Contact Autonomous Nasotracheal Intubation by Recurrent Action-Confidence Chunking with Transformer

Nasotracheal intubation (NTI) is critical for establishing artificial airways in clinical anesthesia and critical care. Current manual methods face significant challenges, including cross-infection, especially during respiratory infection care, and insufficient control of endoluminal contact forces, increasing the risk of mucosal injuries. While existing studies have focused on automated endoscopic insertion, the automation of NTI remains unexplored despite its unique challenges: Nasotracheal tubes exhibit greater diameter and rigidity than standard endoscopes, substantially increasing insertion complexity and patient risks. We propose a novel autonomous NTI system with two key components to address these challenges. First, an autonomous NTI system is developed, incorporating a prosthesis embedded with force sensors, allowing for safety assessment and data filtering. Then, the Recurrent Action-Confidence Chunking with Transformer (RACCT) model is developed to handle complex tube-tissue interactions and partial visual observations. Experimental results demonstrate that the RACCT model outperforms the ACT model in all aspects and achieves a 66% reduction in average peak insertion force compared to manual operations while maintaining equivalent success rates. This validates the system's potential for reducing infection risks and improving procedural safety.

cs.RO↗

arXiv2025

FairFedMed: Benchmarking Group Fairness in Federated Medical Imaging with FairLoRA

Fairness remains a critical concern in healthcare, where unequal access to services and treatment outcomes can adversely affect patient health. While Federated Learning (FL) presents a collaborative and privacy-preserving approach to model training, ensuring fairness is challenging due to heterogeneous data across institutions, and current research primarily addresses non-medical applications. To fill this gap, we establish the first experimental benchmark for fairness in medical FL, evaluating six representative FL methods across diverse demographic attributes and imaging modalities. We introduce FairFedMed, the first medical FL dataset specifically designed to study group fairness (i.e., demographics). It comprises two parts: FairFedMed-Oph, featuring 2D fundus and 3D OCT ophthalmology samples with six demographic attributes; and FairFedMed-Chest, which simulates real cross-institutional FL using subsets of CheXpert and MIMIC-CXR. Together, they support both simulated and real-world FL across diverse medical modalities and demographic groups. Existing FL models often underperform on medical images and overlook fairness across demographic groups. To address this, we propose FairLoRA, a fairness-aware FL framework based on SVD-based low-rank approximation. It customizes singular value matrices per demographic group while sharing singular vectors, ensuring both fairness and efficiency. Experimental results on the FairFedMed dataset demonstrate that FairLoRA not only achieves state-of-the-art performance in medical image classification but also significantly improves fairness across diverse populations. Our code and dataset can be accessible via link: https://wang.hms.harvard.edu/fairfedmed/.

cs.CY↗

arXiv2025

Seedream 3.0 Technical Report

We present Seedream 3.0, a high-performance Chinese-English bilingual image generation foundation model. We develop several technical improvements to address existing challenges in Seedream 2.0, including alignment with complicated prompts, fine-grained typography generation, suboptimal visual aesthetics and fidelity, and limited image resolutions. Specifically, the advancements of Seedream 3.0 stem from improvements across the entire pipeline, from data construction to model deployment. At the data stratum, we double the dataset using a defect-aware training paradigm and a dual-axis collaborative data-sampling framework. Furthermore, we adopt several effective techniques such as mixed-resolution training, cross-modality RoPE, representation alignment loss, and resolution-aware timestep sampling in the pre-training phase. During the post-training stage, we utilize diversified aesthetic captions in SFT, and a VLM-based reward model with scaling, thereby achieving outputs that well align with human preferences. Furthermore, Seedream 3.0 pioneers a novel acceleration paradigm. By employing consistent noise expectation and importance-aware timestep sampling, we achieve a 4 to 8 times speedup while maintaining image quality. Seedream 3.0 demonstrates significant improvements over Seedream 2.0: it enhances overall capabilities, in particular for text-rendering in complicated Chinese characters which is important to professional typography generation. In addition, it provides native high-resolution output (up to 2K), allowing it to generate images with high visual quality.

cs.CV↗

arXiv2025

Research on Model Parallelism and Data Parallelism Optimization Methods in Large Language Model-Based Recommendation Systems

With the rapid adoption of large language models (LLMs) in recommendation systems, the computational and communication bottlenecks caused by their massive parameter sizes and large data volumes have become increasingly prominent. This paper systematically investigates two classes of optimization methods-model parallelism and data parallelism-for distributed training of LLMs in recommendation scenarios. For model parallelism, we implement both tensor parallelism and pipeline parallelism, and introduce an adaptive load-balancing mechanism to reduce cross-device communication overhead. For data parallelism, we compare synchronous and asynchronous modes, combining gradient compression and sparsification techniques with an efficient aggregation communication framework to significantly improve bandwidth utilization. Experiments conducted on a real-world recommendation dataset in a simulated service environment demonstrate that our proposed hybrid parallelism scheme increases training throughput by over 30% and improves resource utilization by approximately 20% compared to traditional single-mode parallelism, while maintaining strong scalability and robustness. Finally, we discuss trade-offs among different parallel strategies in online deployment and outline future directions involving heterogeneous hardware integration and automated scheduling technologies.

cs.DC↗

arXiv2025

Root Defence Strategies: Ensuring Safety of LLM at the Decoding Level

Large language models (LLMs) have demonstrated immense utility across various industries. However, as LLMs advance, the risk of harmful outputs increases due to incorrect or malicious instruction prompts. While current methods effectively address jailbreak risks, they share common limitations: 1) Judging harmful responses from the prefill-level lacks utilization of the model's decoding outputs, leading to relatively lower effectiveness and robustness. 2) Rejecting potentially harmful responses based on a single evaluation can significantly impair the model's helpfulness.This paper examines the LLMs' capability to recognize harmful outputs, revealing and quantifying their proficiency in assessing the danger of previous tokens. Motivated by pilot experiment results, we design a robust defense mechanism at the decoding level. Our novel decoder-oriented, step-by-step defense architecture corrects harmful queries directly rather than rejecting them outright. We introduce speculative decoding to enhance usability and facilitate deployment to boost secure decoding speed. Extensive experiments demonstrate that our approach improves model security without compromising reasoning speed. Notably, our method leverages the model's ability to discern hazardous information, maintaining its helpfulness compared to existing methods.

cs.CL↗

arXiv2025

Translation Consistent Semi-supervised Segmentation for 3D Medical Images

3D medical image segmentation methods have been successful, but their dependence on large amounts of voxel-level annotated data is a disadvantage that needs to be addressed given the high cost to obtain such annotation. Semi-supervised learning (SSL) solve this issue by training models with a large unlabelled and a small labelled dataset. The most successful SSL approaches are based on consistency learning that minimises the distance between model responses obtained from perturbed views of the unlabelled data. These perturbations usually keep the spatial input context between views fairly consistent, which may cause the model to learn segmentation patterns from the spatial input contexts instead of the segmented objects. In this paper, we introduce the Translation Consistent Co-training (TraCoCo) which is a consistency learning SSL method that perturbs the input data views by varying their spatial input context, allowing the model to learn segmentation patterns from visual objects. Furthermore, we propose the replacement of the commonly used mean squared error (MSE) semi-supervised loss by a new Cross-model confident Binary Cross entropy (CBC) loss, which improves training convergence and keeps the robustness to co-training pseudo-labelling mistakes. We also extend CutMix augmentation to 3D SSL to further improve generalisation. Our TraCoCo shows state-of-the-art results for the Left Atrium (LA) and Brain Tumor Segmentation (BRaTS19) datasets with different backbones. Our code is available at https://github.com/yyliu01/TraCoCo.

cs.CV↗

arXiv2025

Modality Curation: Building Universal Embeddings for Advanced Multimodal Information Retrieval

Multimodal information retrieval (MIR) faces inherent challenges due to the heterogeneity of data sources and the complexity of cross-modal alignment. While previous studies have identified modal gaps in feature spaces, a systematic approach to address these challenges remains unexplored. In this work, we introduce UNITE, a universal framework that tackles these challenges through two critical yet underexplored aspects: data curation and modality-aware training configurations. Our work provides the first comprehensive analysis of how modality-specific data properties influence downstream task performance across diverse scenarios. Moreover, we propose Modal-Aware Masked Contrastive Learning (MAMCL) to mitigate the competitive relationships among the instances of different modalities. Our framework achieves state-of-the-art results on multiple multimodal retrieval benchmarks, outperforming existing methods by notable margins. Through extensive experiments, we demonstrate that strategic modality curation and tailored training protocols are pivotal for robust cross-modal representation learning. This work not only advances MIR performance but also provides a foundational blueprint for future research in multimodal systems. Our project is available at https://friedrichor.github.io/projects/UNITE.

cs.CV↗

arXiv2025

TUNA: Comprehensive Fine-grained Temporal Understanding Evaluation on Dense Dynamic Videos

Videos are unique in their integration of temporal elements, including camera, scene, action, and attribute, along with their dynamic relationships over time. However, existing benchmarks for video understanding often treat these properties separately or narrowly focus on specific aspects, overlooking the holistic nature of video content. To address this, we introduce TUNA, a temporal-oriented benchmark for fine-grained understanding on dense dynamic videos, with two complementary tasks: captioning and QA. Our TUNA features diverse video scenarios and dynamics, assisted by interpretable and robust evaluation criteria. We evaluate several leading models on our benchmark, providing fine-grained performance assessments across various dimensions. This evaluation reveals key challenges in video temporal understanding, such as limited action description, inadequate multi-subject understanding, and insensitivity to camera motion, offering valuable insights for improving video understanding models. The data and code are available at https://friedrichor.github.io/projects/TUNA.

cs.CV↗

arXiv2025

Bayesian inferences on covariant density functionals from multimessenger astrophysical data: Nucleonic models

[Background] Bayesian inference frameworks incorporating multi-messenger astrophysical constraints have recently been applied to covariant density functional (CDF) models to constrain their parameters. Among these, frameworks utilizing CDFs with density-dependent meson-nucleon couplings furnishing the equation of state (EoS) of compact star (CS) matter have been explored. [Purpose] The aforementioned inference framework has not yet incorporated astrophysical objects with potentially extreme high masses or ultra-small radii among its constraints, leaving its flexibility and predictive power under such extreme parameters still unknown. [Method] We apply the Bayesian inference framework based on CDFs with density dependent couplings. The astrophysical data is expanded to include not only the latest multi-messenger constraints from NICER and gravitational wave events but also the highest measured mass to date for the ``black widow" pulsar PSR J0952-0607 and the mass-radius estimates for the ultra-compact, low-mass object HESS J1731-347. [Results] Our systematic Bayesian analysis indicates that our CDF models can support higher maximum masses for CSs, reaching up to $2.4$-$2.5\,M_{\odot}$. However, achieving sufficient softening of the EoS in the low-density regime to accommodate the HESS J1731-347 data remains challenging. Nonetheless, we are able to impose tighter constraints on the parameter space of CDF models, ensuring consistency with current nuclear experimental and astrophysical data. [Conclusions] CDF models with density-dependent meson-nucleon couplings encompass a wide range of nuclear and astrophysical phenomena, providing a robust theoretical framework for interpreting compact objects. However, the predicted lower limit for the radii of low-mass stars is approximately 12 km, which stems from the restricted degrees of freedom in the isovector sector.

nucl-th↗

arXiv2025

Landau Instability and soliton formation

Consider at a finite temperature $T$ a superfluid moving with a velocity $v$ relative to the thermal bath or its normal component. From Landau's argument there exists a critical $v_c (T)$ beyond which excitations can be spontaneously generated and the system becomes unstable. Identifying the final state induced by such an instability has been an outstanding open question. Using holographic duality we perform dynamical simulations of evolutions from initial unstable states, and find that the system settles to a homogenous superfluid state with a final velocity below the critical velocity. The dynamical evolution process appears to be highly chaotic, exhibiting transient turbulence. Nevertheless we are able to identify from the simulations a universal physical mechanism for the reduction of superfluid velocity, in terms of spontaneous nucleation of solitons. We also derive a simple analytic formula which relates the final velocity to the number of solitons nucleated during the evolution.

hep-th↗

arXiv2025

Broadband acousto-optic modulators on Silicon Nitride

Stress-optic modulators are emerging as a necessary building block of photonic integrated circuits tasked with controlling and manipulating classical and quantum optical systems. While photonic platforms such as lithium niobate and silicon on insulator have well developed modulator ecosystems, silicon nitride so far does not. As silicon nitride has favorable optical properties, such as ultra-low-loss and a large optical transparency window, a rich ecosystem of potential photonic integrated circuits are therefore inhibited. Here we demonstrate a traveling wave optically broadband acousto-optic spiral modulator architecture at a wavelength of 1550 nm using 90 nm thick silicon nitride waveguides and demonstrate their use in an optomechanical sensing system. The spiral weaves the light repeatedly through the acoustic field up to 38 times, factoring in the time evolution of the acoustic field during the light's transit through spirals up to 26 cm in length. These modulators avoid heterogeneous integration, release processes, complicated fabrication procedures, and modifications of the commercial foundry fabricated photonic layer stack by exploiting ultra-low-loss waveguides to enable long phonon-photon interaction lengths required for efficient modulation. The design allows for thick top oxide cladding of 4 $μ$m such that the low loss optical properties of thin silicon nitride can be preserved, ultimately achieving a $V_π$ of 8.98 V at 704 MHz with 1.13 dB of insertion loss. Our modulators are the first optically broadband high frequency acousto-optic modulators on thin silicon nitride, and the novel architecture is accessible to any low loss photonic platform. We demonstrate an immediate use case for these devices in a high-Q optomechanical sensing system.

physics.optics↗

arXiv2025

Bayesian constraints on covariant density functional equations of state of compact stars with new NICER mass-radius measurements

Recent advancements in astrophysical observations of compact stars, particularly the new and updated NICER constraints, have provided mass-radius ($M$-$R$) data for pulsars spanning masses from 1 to $2\,M_{\odot }$. These data offer a unique opportunity to test modern theories of dense matter using multi-messenger constraints. Covariant density functional (CDF) models of nuclear matter, which capture a broad range of nuclear and astrophysical phenomena, provide a robust theoretical framework to interpret these observations. This study applies the Bayesian framework to a class of CDF models with density-dependent meson-nucleon couplings, specifically those based on nucleonic degrees of freedom. By incorporating the latest multi-messenger constraints, we impose tighter limits on the parameter space of these models and assess their consistency with observational data. Our analysis advances previous efforts by refining the density-dependence parameterization and integrating recent $M$-$R$ ellipses. This enables more stringent evaluations of dense matter models in light of new astrophysical observations.

hep-ph↗

arXiv2025

Discharge structure theory of highly electronegative plasma and its hierarchy and interdisciplinary meanings

In this work the systematic theory of discharge structure is built for highly electronegative plasma, by means of self-consistent fluid model simulation that is based on the finite element method. The highly electronegative plasma is selected to be the inductively coupled Ar/SF6 plasma source with 10% the reactive SF6 concentration and in a pressure range of 10~90mTorr. The discharge structure is classified the transport dominated regime, transport and chemistry self-balanced regime and chemistry dominated regime. At low pressure of 10mTorr, the parabola feature of core plasma, stratification of whole discharge area into electronegative core and electropositive halo, anion potential barrel, and the dipole and capacitor models of double layer characterize the discharge structure of transport dominated regime. At increasing the pressure, the recombination loss of ions becomes significant and the discharge structure is characterized by ellipse profile. Meanwhile, the regions of double layer and electropositive halo are strikingly shrunk, which means that the plasma of transport and chemistry self-balanced regime is a close system and probably do not need the shield of chamber anymore. The dimensional analysis shows the recombination can be transformed into drift flux, which balances the ambi-polar diffusion of plasma species. In the range of pressure considered, simulation shows astro-structures are inlayed in the parabolic and elliptic profiles. At observing the characteristics of the astro-structures, the self-coagulation theory and quasi-Helmholtz equation are built based on the free diffusion and negative chemical source. This is the chemistry dominated regime and defined as a tight type of self-balance since the inertia is lost automatically in the unsteady state continuity equations of anions after counteracting the diffusion and recombination.

physics.plasm-ph↗

arXiv2025

Feature-Aware Malicious Output Detection and Mitigation

The rapid advancement of large language models (LLMs) has brought significant benefits to various domains while introducing substantial risks. Despite being fine-tuned through reinforcement learning, LLMs lack the capability to discern malicious content, limiting their defense against jailbreak. To address these safety concerns, we propose a feature-aware method for harmful response rejection (FMM), which detects the presence of malicious features within the model's feature space and adaptively adjusts the model's rejection mechanism. By employing a simple discriminator, we detect potential malicious traits during the decoding phase. Upon detecting features indicative of toxic tokens, FMM regenerates the current token. By employing activation patching, an additional rejection vector is incorporated during the subsequent token generation, steering the model towards a refusal response. Experimental results demonstrate the effectiveness of our approach across multiple language models and diverse attack techniques, while crucially maintaining the models' standard generation capabilities.

cs.CL↗

arXiv2025

FairDiffusion: Enhancing Equity in Latent Diffusion Models via Fair Bayesian Perturbation

Recent progress in generative AI, especially diffusion models, has demonstrated significant utility in text-to-image synthesis. Particularly in healthcare, these models offer immense potential in generating synthetic datasets and training medical students. However, despite these strong performances, it remains uncertain if the image generation quality is consistent across different demographic subgroups. To address this critical concern, we present the first comprehensive study on the fairness of medical text-to-image diffusion models. Our extensive evaluations of the popular Stable Diffusion model reveal significant disparities across gender, race, and ethnicity. To mitigate these biases, we introduce FairDiffusion, an equity-aware latent diffusion model that enhances fairness in both image generation quality as well as the semantic correlation of clinical features. In addition, we also design and curate FairGenMed, the first dataset for studying the fairness of medical generative models. Complementing this effort, we further evaluate FairDiffusion on two widely-used external medical datasets: HAM10000 (dermatoscopic images) and CheXpert (chest X-rays) to demonstrate FairDiffusion's effectiveness in addressing fairness concerns across diverse medical imaging modalities. Together, FairDiffusion and FairGenMed significantly advance research in fair generative learning, promoting equitable benefits of generative AI in healthcare.

cs.CV↗

arXiv2025

Open-Qwen2VL: Compute-Efficient Pre-Training of Fully-Open Multimodal LLMs on Academic Resources

The reproduction of state-of-the-art multimodal LLM pre-training faces barriers at every stage of the pipeline, including high-quality data filtering, multimodal data mixture strategies, sequence packing techniques, and training frameworks. We introduce Open-Qwen2VL, a fully open-source 2B-parameter Multimodal Large Language Model pre-trained efficiently on 29M image-text pairs using only 220 A100-40G GPU hours. Our approach employs low-to-high dynamic image resolution and multimodal sequence packing to significantly enhance pre-training efficiency. The training dataset was carefully curated using both MLLM-based filtering techniques (e.g., MLM-Filter) and conventional CLIP-based filtering methods, substantially improving data quality and training efficiency. The Open-Qwen2VL pre-training is conducted on academic level 8xA100-40G GPUs at UCSB on 5B packed multimodal tokens, which is 0.36% of 1.4T multimodal pre-training tokens of Qwen2-VL. The final instruction-tuned Open-Qwen2VL outperforms partially-open state-of-the-art MLLM Qwen2-VL-2B on various multimodal benchmarks of MMBench, SEEDBench, MMstar, and MathVista, indicating the remarkable training efficiency of Open-Qwen2VL. We open-source all aspects of our work, including compute-efficient and data-efficient training details, data filtering methods, sequence packing scripts, pre-training data in WebDataset format, FSDP-based training codebase, and both base and instruction-tuned model checkpoints. We redefine "fully open" for multimodal LLMs as the complete release of: 1) the training codebase, 2) detailed data filtering techniques, and 3) all pre-training and supervised fine-tuning data used to develop the model.

cs.CL↗

arXiv2025

Time and Memory Trade-off of KV-Cache Compression in Tensor Transformer Decoding

The key-value (KV) cache in the tensor version of transformers presents a significant bottleneck during inference. While previous work analyzes the fundamental space complexity barriers in standard attention mechanisms [Haris and Onak, 2025], our work generalizes the space complexity barriers result to tensor attention version. Our theoretical contributions rely on a reduction from communication complexity and deduce the memory lower bound for tensor-structured attention mechanisms when $d = Ω(\log n)$. Furthermore, we introduce two types of tensor attention cache and present a trade-off between time and memory for two scenarios. Overall, our work provides a theoretical foundation for us to understand the time-memory tradeoff of KV-Cache compression in tensor attention decoding and offers more perspectives in developing more memory-efficient tensor attention Transformer architectures.

cs.LG↗