Xuanzhao Dong

Xiwen Chen, Wenhui Zhu, Jingjing Wang et al.

Aligning Large Language Models (LLMs) with human preferences is often formulated via Direct Preference Optimization (DPO). However, the standard Bradley-Terry instantiation of DPO is limited in modeling common departures from transitivity in human preferences. To address this, recent work has introduced Self-Play Preference Optimization (SPPO), which iteratively refines the policy by training on self-generated win-lose pairs. Our investigation, however, reveals a critical instability in SPPO: the optimization is prone to policy degeneration when the preference oracle assigns overly confident wins to semantically indistinguishable responses. To mitigate this, we propose S-SPPO, a dual-space semantic calibration framework comprising: i) Supervision Calibration via semantic gating, which anneals win rate targets toward the maximum-entropy baseline as semantic overlap increases; and ii) Representation Calibration via latent repulsion to enforce geometric diversity to prevent manifold collapse and maintain latent diversity between chosen and rejected samples. Theoretically, we show that the calibration preserves the constant-sum game structure, facilitating convergence to a Nash Equilibrium. Empirically, S-SPPO avoids the performance degradation seen in prior methods, achieving 52.19% win rate and 47.46% length-controlled win rate on AlpacaEval 2.0 with Llama-3-8B, without using additional human-annotated preferences during training. The code will be available at https://github.com/xiwenc1/s-sppo.

Xuanzhao Dong, Wenhui Zhu, Peijie Qiu et al.

Despite their popularity and success, Multimodal Large Language Models (MLLMs) often struggle to interpret images accurately, which limits their reasoning capability in complex scenarios (e.g., high object density and complex background clutter). Prior work mainly addresses this limitation by incorporating explicit visual cues like bounding boxes that require extra annotations. In addition, the resulting low-resolution crops often miss fine-grained details that MLLMs require for accurate reasoning. Therefore, we propose Mags-RL, an Agentic Reinforcement Learning (RL) framework that equips MLLMs with an external super-resolution "magnifying glass" agent for high-resolution fine-grained inspection. Specifically, the model performs two-round reasoning: in the first round, it generates an initial rationale and autonomously identifies regions of interest without relying on additional annotations; in the second round, it invokes a super-resolution agent to crop and upscale those regions, then revisits and verifies its earlier reasoning to produce the final answer. We also introduce a novel curriculum learning strategy that enables data-efficient RL training, needing as few as only 40 training samples to achieve reasonable performance. Experiments on VSR, TallyQA, and GQA subsets show its superior performance against recent strong competing methods, demonstrating high-quality reasoning with precise visual grounding. Code and weights will be released soon.

Xuanzhao Dong, Wenhui Zhu, Xiwen Chen et al.

The advancement of general medical Multimodal Large Language Models (MLLMs) has shown great potential for building conversational assistants to support clinical diagnosis. However, their adaptation to highly specialized domains such as ophthalmology remains underexplored, primarily due to the scarcity of large-scale, domain-specific instruction-tuning data. Existing ophthalmic datasets for conversational agents are often limited in scale and largely rely on images from established public benchmarks, limiting the scalability of ophthalmic MLLMs and their ability to capture real-world clinical complexity. To address this gap, we propose $\textbf{OphIn-Engine}$, an ophthalmology-specific instruction data curation pipeline that constructs high-quality instruction data from open-access ophthalmology web-scale videos. The pipeline integrates multimodal transcription for extracting image-transcript pairs, visual cue separation and scoring for identifying clinically relevant visual descriptions, and instruction synthesis with quality control for generating accurate and diverse clinical dialogues. Using this engine, we introduce $\textbf{OphIn-500K}$, a large-scale multimodal ophthalmology instruction-tuning dataset containing over 500,000 instruction instances and more than 151,000 unique images from over 29,000 video clips, formatted as visual question answering (VQA), multi-turn conversational interactions, and chain-of-thought (CoT) reasoning. Built upon this dataset, we further develop $\textbf{OphIn-VL}$, an ophthalmology-specific MLLM with advanced visual understanding and conversational capabilities. Comprehensive experiments and case studies demonstrate that OphIn-VL achieves superior performance compared with state-of-the-art general medical and domain-specific MLLMs.

Xuanzhao Dong, Vamsi Krishna Vasa, Wenhui Zhu et al.

Retinal fundus photography is significant in diagnosing and monitoring retinal diseases. However, systemic imperfections and operator/patient-related factors can hinder the acquisition of high-quality retinal images. Previous efforts in retinal image enhancement primarily relied on GANs, which are limited by the trade-off between training stability and output diversity. In contrast, the Schrödinger Bridge (SB), offers a more stable solution by utilizing Optimal Transport (OT) theory to model a stochastic differential equation (SDE) between two arbitrary distributions. This allows SB to effectively transform low-quality retinal images into their high-quality counterparts. In this work, we leverage the SB framework to propose an image-to-image translation pipeline for retinal image enhancement. Additionally, previous methods often fail to capture fine structural details, such as blood vessels. To address this, we enhance our pipeline by introducing Dynamic Snake Convolution, whose tortuous receptive field can better preserve tubular structures. We name the resulting retinal fundus image enhancement framework the Context-aware Unpaired Neural Schrödinger Bridge (CUNSB-RFIE). To the best of our knowledge, this is the first endeavor to use the SB approach for retinal image enhancement. Experimental results on a large-scale dataset demonstrate the advantage of the proposed method compared to several state-of-the-art supervised and unsupervised methods in terms of image quality and performance on downstream tasks.The code is available at https://github.com/Retinal-Research/CUNSB-RFIE .

Wenhui Zhu, Xuanzhao Dong, Xiwen Chen et al.

The rapid deployment of open-source frameworks has significantly advanced the development of modern multi-agent systems. However, expanded action spaces, including uncontrolled privilege exposure and hidden inter-system interactions, pose severe security challenges. Specifically, Indirect Prompt Injections (IPI), which conceal malicious instructions within third-party content, can trigger unauthorized actions such as data exfiltration during normal operations. While current security evaluations predominantly rely on isolated single-turn benchmarks, the systemic vulnerabilities of these agents within complex dynamic environments remain critically underexplored. To bridge this gap, we systematically evaluate six defense strategies against four sophisticated IPI attack vectors across nine LLM backbones. Crucially, we conduct our evaluation entirely within dynamic multi-step tool-calling environments to capture the true attack surface of modern autonomous agents. Moving beyond binary success rates, our multidimensional analysis reveals a pronounced fragility. Advanced injections successfully bypass nearly all baseline defenses, and some surface-level mitigations even produce counterproductive side effects. Furthermore, while agents execute malicious instructions almost instantaneously, their internal states exhibit abnormally high decision entropy. Motivated by this latent hesitation, we investigate Representation Engineering (RepE) as a robust detection strategy. By extracting hidden states at the tool-input position, we revealed that the RepE-based circuit breaker successfully identifies and intercepts unauthorized actions before the agent commits to them, achieving high detection accuracy across diverse LLM backbones. This study exposes the limitations of current IPI defenses and provides a highly practical paradigm for building resilient multi-agent architectures.

Wenhui Zhu, Xiwen Chen, Zhipeng Wang et al.

Long-horizon LLM agents require memory systems that remain accurate under fixed context budgets. However, existing systems struggle with two persistent challenges in long-term dialogue: (i) \textbf{disconnected evidence}, where multi-hop answers require linking facts distributed across time, and (ii) \textbf{state updates}, where evolving information (e.g., schedule changes) creates conflicts with older static logs. We propose AriadneMem, a structured memory system that addresses these failure modes via a decoupled two-phase pipeline. In the \textbf{offline construction phase}, AriadneMem employs \emph{entropy-aware gating} to filter noise and low-information message before LLM extraction and applies \emph{conflict-aware coarsening} to merge static duplicates while preserving state transitions as temporal edges. In the \textbf{online reasoning phase}, rather than relying on expensive iterative planning, AriadneMem executes \emph{algorithmic bridge discovery} to reconstruct missing logical paths between retrieved facts, followed by \emph{single-call topology-aware synthesis}. On LoCoMo experiments with GPT-4o, AriadneMem improves \textbf{Multi-Hop F1 by 15.2\%} and \textbf{Average F1 by 9.0\%} over strong baselines. Crucially, by offloading reasoning to the graph layer, AriadneMem reduces \textbf{total runtime by 77.8\%} using only \textbf{497} context tokens. The code is available at https://github.com/LLM-VLM-GSL/AriadneMem.

Yanxi Chen, Wenhui Zhu, Xiwen Chen et al.

Although Large Audio-Language Models (LALMs) deliver state-of-the-art (SOTA) performance, they frequently suffer from hallucinations, e.g. generating text not grounded in the audio input. We analyze these grounding failures and identify a distinct taxonomy: Event Omission, False Event Identity, Temporal Relation Error, and Quantitative Temporal Error. To address this, we introduce the AHA (Audio Hallucination Alignment) framework. By leveraging counterfactual hard negative mining, our pipeline constructs a high-quality preference dataset that forces models to distinguish strict acoustic evidence from linguistically plausible fabrications. Additionally, we establish AHA-Eval, a diagnostic benchmark designed to rigorously test these fine-grained temporal reasoning capabilities. We apply this data to align Qwen2.5-Omni. The resulting model, Qwen-Audio-AHA, achieves a 13.7% improvement on AHA-Eval. Crucially, this benefit generalizes beyond our diagnostic set. Our model shows substantial gains on public benchmarks, including 1.3% on MMAU-Test and 1.6% on MMAR, outperforming latest SOTA methods. The model and dataset are open-sourced at https://github.com/LLM-VLM-GSL/AHA.

Xiwen Chen, Jingjing Wang, Wenhui Zhu et al.

Black-box knowledge distillation for large language models presents a strict trade-off. Simple off-policy methods (e.g., sequence-level knowledge distillation) struggle to correct the student's inherent errors. Fully on-policy methods (e.g., Generative Adversarial Distillation) solve this via adversarial training but introduce well-known training instability and crippling computational overhead. To address this dilemma, we propose SODA (Semi On-policy Distillation with Alignment), a highly efficient alternative motivated by the inherent capability gap between frontier teachers and much smaller base models. Because a compact student model's natural, zero-shot responses are almost strictly inferior to the powerful teacher's targets, we can construct a highly effective contrastive signal simply by pairing the teacher's optimal response with a one-time static snapshot of the student's outputs. This demonstrates that exposing the small student to its own static inferior behaviors is sufficient for high-quality distribution alignment, eliminating the need for costly dynamic rollouts and fragile adversarial balancing. Extensive evaluations across four compact Qwen2.5 and Llama-3 models validate this semi on-policy paradigm. SODA matches or outperforms the state-of-the-art methods on 15 out of 16 benchmark results. More importantly, it achieves this superior distillation quality while training 10 times faster, consuming 27% less peak GPU memory, and completely eliminating adversarial instability.

Xin Li, Wenhui Zhu, Xuanzhao Dong et al.

Retinal imaging is a critical, non-invasive modality for the early detection and monitoring of ocular and systemic diseases. Deep learning, particularly convolutional neural networks (CNNs), has significant progress in automated retinal analysis, supporting tasks such as fundus image classification, lesion detection, and vessel segmentation. As a representative lightweight network, nnMobileNet has demonstrated strong performance across multiple retinal benchmarks while remaining computationally efficient. However, purely convolutional architectures inherently struggle to capture long-range dependencies and model the irregular lesions and elongated vascular patterns that characterize on retinal images, despite the critical importance of vascular features for reliable clinical diagnosis. To further advance this line of work and extend the original vision of nnMobileNet, we propose nnMobileNet++, a hybrid architecture that progressively bridges convolutional and transformer representations. The framework integrates three key components: (i) dynamic snake convolution for boundary-aware feature extraction, (ii) stage-specific transformer blocks introduced after the second down-sampling stage for global context modeling, and (iii) retinal image pretraining to improve generalization. Experiments on multiple public retinal datasets for classification, together with ablation studies, demonstrate that nnMobileNet++ achieves state-of-the-art or highly competitive accuracy while maintaining low computational cost, underscoring its potential as a lightweight yet effective framework for retinal image analysis.

Xiwen Chen, Wenhui Zhu, Peijie Qiu et al.

Recent advances in reinforcement learning for language model post-training, such as Group Relative Policy Optimization (GRPO), have shown promise in low-resource settings. However, GRPO typically relies on solution-level and scalar reward signals that fail to capture the semantic diversity among sampled completions. This leads to what we identify as a diversity-quality inconsistency, where distinct reasoning paths may receive indistinguishable rewards. To address this limitation, we propose $\textit{Diversity-aware Reward Adjustment}$ (DRA), a method that explicitly incorporates semantic diversity into the reward computation. DRA uses Submodular Mutual Information (SMI) to downweight redundant completions and amplify rewards for diverse ones. This encourages better exploration during learning, while maintaining stable exploitation of high-quality samples. Our method integrates seamlessly with both GRPO and its variant DR.~GRPO, resulting in $\textit{DRA-GRPO}$ and $\textit{DGA-DR.~GRPO}$. We evaluate our method on five mathematical reasoning benchmarks and find that it outperforms recent strong baselines. It achieves state-of-the-art performance with an average accuracy of 58.2%, using only 7,000 fine-tuning samples and a total training cost of approximately $55. The code is available at https://github.com/xiwenc1/DRA-GRPO.

Wenhui Zhu, Xin Li, Xiwen Chen et al.

Recently, Multimodal Large Language Models (MLLMs) have gained significant attention for their remarkable ability to process and analyze non-textual data, such as images, videos, and audio. Notably, several adaptations of general-domain MLLMs to the medical field have been explored, including LLaVA-Med. However, these medical adaptations remain insufficiently advanced in understanding and interpreting retinal images. In contrast, medical experts emphasize the importance of quantitative analyses for disease detection and interpretation. This underscores a gap between general-domain and medical-domain MLLMs: while general-domain MLLMs excel in broad applications, they lack the specialized knowledge necessary for precise diagnostic and interpretative tasks in the medical field. To address these challenges, we introduce \textit{RetinalGPT}, a multimodal conversational assistant for clinically preferred quantitative analysis of retinal images. Specifically, we achieve this by compiling a large retinal image dataset, developing a novel data pipeline, and employing customized visual instruction tuning to enhance both retinal analysis and enrich medical knowledge. In particular, RetinalGPT outperforms MLLM in the generic domain by a large margin in the diagnosis of retinal diseases in 8 benchmark retinal datasets. Beyond disease diagnosis, RetinalGPT features quantitative analyses and lesion localization, representing a pioneering step in leveraging LLMs for an interpretable and end-to-end clinical research framework. The code is available at https://github.com/Retinal-Research/RetinalGPT

Zhangsihao Yang, Mengyi Shan, Mohammad Farazi et al.

Human video generation task has gained significant attention with the advancement of deep generative models. Generating realistic videos with human movements is challenging in nature, due to the intricacies of human body topology and sensitivity to visual artifacts. The extensively studied 2D media generation methods take advantage of massive human media datasets, but struggle with 3D-aware control; whereas 3D avatar-based approaches, while offering more freedom in control, lack photorealism and cannot be harmonized seamlessly with background scene. We propose AMG, a method that combines the 2D photorealism and 3D controllability by conditioning video diffusion models on controlled rendering of 3D avatars. We additionally introduce a novel data processing pipeline that reconstructs and renders human avatar movements from dynamic camera videos. AMG is the first method that enables multi-person diffusion video generation with precise control over camera positions, human motions, and background style. We also demonstrate through extensive evaluation that it outperforms existing human video generation methods conditioned on pose sequences or driving videos in terms of realism and adaptability.

Xuanzhao Dong, Wenhui Zhu, Hao Wang et al.

Medical question answering (QA) is a reasoning-intensive task that remains challenging for large language models (LLMs) due to hallucinations and outdated domain knowledge. Retrieval-Augmented Generation (RAG) provides a promising post-training solution by leveraging external knowledge. However, existing medical RAG systems suffer from two key limitations: (1) a lack of modeling for human-like reasoning behaviors during information retrieval, and (2) reliance on suboptimal medical corpora, which often results in the retrieval of irrelevant or noisy snippets. To overcome these challenges, we propose Discuss-RAG, a plug-and-play module designed to enhance the medical QA RAG system through collaborative agent-based reasoning. Our method introduces a summarizer agent that orchestrates a team of medical experts to emulate multi-turn brainstorming, thereby improving the relevance of retrieved content. Additionally, a decision-making agent evaluates the retrieved snippets before their final integration. Experimental results on four benchmark medical QA datasets show that Discuss-RAG consistently outperforms MedRAG, especially significantly improving answer accuracy by up to 16.67% on BioASQ and 12.20% on PubMedQA. The code is available at: https://github.com/LLM-VLM-GSL/Discuss-RAG.

Xuanzhao Dong, Wenhui Zhu, Xiwen Chen et al.

Autoregressive models (ARMs) have long dominated the landscape of biomedical vision-language models (VLMs). Recently, masked diffusion models such as LLaDA have emerged as promising alternatives, yet their application in the biomedical domain remains largely underexplored. To bridge this gap, we introduce \textbf{LLaDA-MedV}, the first large language diffusion model tailored for biomedical image understanding through vision instruction tuning. LLaDA-MedV achieves relative performance gains of 7.855\% over LLaVA-Med and 1.867\% over LLaDA-V in the open-ended biomedical visual conversation task, and sets new state-of-the-art accuracy on the closed-form subset of three VQA benchmarks: 84.93\% on VQA-RAD, 92.31\% on SLAKE, and 95.15\% on PathVQA. Furthermore, a detailed comparison with LLaVA-Med suggests that LLaDA-MedV is capable of generating reasonably longer responses by explicitly controlling response length, which can lead to more informative outputs. We also conduct an in-depth analysis of both the training and inference stages, highlighting the critical roles of initialization weight selection, fine-tuning strategies, and the interplay between sampling steps and response repetition. The code and model weight is released at https://github.com/LLM-VLM-GSL/LLaDA-MedV.

Yujian Xiong, Mohammad Farazi, Yanxi Chen et al.

Representation learning on large-scale unstructured volumetric and surface meshes poses significant challenges in neuroimaging, especially when models must incorporate diverse vertex-level morphometric descriptors, such as cortical thickness, curvature, sulcal depth, and myelin content, which carry subtle disease-related signals. Current approaches either ignore these clinically informative features or support only a single mesh topology, restricting their use across imaging pipelines. We introduce a hierarchical transformer framework designed for heterogeneous mesh analysis that operates on spatially adaptive tree partitions constructed from simplicial complexes of arbitrary order. This design accommodates both volumetric and surface discretizations within a single architecture, enabling efficient multi-scale attention without topology-specific modifications. A feature projection module maps variable-length per-vertex clinical descriptors into the spatial hierarchy, separating geometric structure from feature dimensionality and allowing seamless integration of different neuroimaging feature sets. Self-supervised pretraining via masked reconstruction of both coordinates and morphometric channels on large unlabeled cohorts yields a transferable encoder backbone applicable to diverse downstream tasks and mesh modalities. We validate our approach on Alzheimer's disease classification and amyloid burden prediction using volumetric brain meshes from ADNI, as well as focal cortical dysplasia detection on cortical surface meshes from the MELD dataset, achieving state-of-the-art results across all benchmarks.

Xuanzhao Dong, Wenhui Zhu, Xiwen Chen et al.

Over the past decade, generative models have demonstrated success in enhancing fundus images. However, the evaluation of these models remains a challenge. A benchmark for fundus image enhancement is needed for three main reasons:(1) Conventional denoising metrics such as PSNR and SSIM fail to capture clinically relevant features, such as lesion preservation and vessel morphology consistency, limiting their applicability in real-world settings; (2) There is a lack of unified evaluation protocols that address both paired and unpaired enhancement methods, particularly those guided by clinical expertise; and (3) An evaluation framework should provide actionable insights to guide future advancements in clinically aligned enhancement models. To address these gaps, we introduce EyeBench-V2, a benchmark designed to bridge the gap between enhancement model performance and clinical utility. Our work offers three key contributions:(1) Multi-dimensional clinical-alignment through downstream evaluations: Beyond standard enhancement metrics, we assess performance across clinically meaningful tasks including vessel segmentation, diabetic retinopathy (DR) grading, generalization to unseen noise patterns, and lesion segmentation. (2) Expert-guided evaluation design: We curate a novel dataset enabling fair comparisons between paired and unpaired enhancement methods, accompanied by a structured manual assessment protocol by medical experts, which evaluates clinically critical aspects such as lesion structure alterations, background color shifts, and the introduction of artificial structures. (3) Actionable insights: Our benchmark provides a rigorous, task-oriented analysis of existing generative models, equipping clinical researchers with the evidence needed to make informed decisions, while also identifying limitations in current methods to inform the design of next-generation enhancement models.

Jingjing Wang, Xiwen Chen, Wenhui Zhu et al.

LLM-powered coding agents spend the majority of their token budget reading repository files, yet much of the retrieved code is irrelevant to the task at hand. Existing learned pruners compress this context with a single-objective sequence labeler, collapsing all facets of code relevance into one score and one transition matrix. We show that this formulation creates a modeling bottleneck: a single CRF transition prior must serve heterogeneous retention patterns, including contiguous semantic spans and sparse structural support lines. We propose LaMR (Latent Multi-Rubric), a structured pruning framework that decomposes code relevance into two interpretable quality dimensions, semantic evidence and dependency support, each modeled by a dedicated CRF with dimension-specific transition dynamics. A mixture-of-experts gating network dynamically weights the per-rubric emissions conditioned on the query, and a final CRF layer on the fused emissions produces the aggregate keep-or-prune decision. To supervise each dimension without additional annotation cost, we derive multi-rubric labels from the existing training corpus via AST-based program analysis, simultaneously denoising the teacher's binary labels. By effectively filtering distracting noise, LaMR frequently matches or even outperforms unpruned full-context baselines. Experiments on four benchmarks (SWE-Bench Verified, SWE-QA, LCC, LongCodeQA) show that LaMR wins 12 of 16 head-to-head multi-turn comparisons. It saves up to 31% more tokens on multi-turn agent tasks and improves Exact Match by up to +3.5 on single-turn tasks, while performance is frequently enhanced by denoising the context, and any remaining drops are marginal.

Xuanzhao Dong, Wenhui Zhu, Xin Li et al.

Retinal fundus photography enhancement is important for diagnosing and monitoring retinal diseases. However, early approaches to retinal image enhancement, such as those based on Generative Adversarial Networks (GANs), often struggle to preserve the complex topological information of blood vessels, resulting in spurious or missing vessel structures. The persistence diagram, which captures topological features based on the persistence of topological structures under different filtrations, provides a promising way to represent the structure information. In this work, we propose a topology-preserving training paradigm that regularizes blood vessel structures by minimizing the differences of persistence diagrams. We call the resulting framework Topology Preserving Optimal Transport (TPOT). Experimental results on a large-scale dataset demonstrate the superiority of the proposed method compared to several state-of-the-art supervised and unsupervised techniques, both in terms of image quality and performance in the downstream blood vessel segmentation task. The code is available at https://github.com/Retinal-Research/TPOT.

Wenhui Zhu, Xuanzhao Dong, Xin Li et al.

Over the past decade, generative models have achieved significant success in enhancement fundus images.However, the evaluation of these models still presents a considerable challenge. A comprehensive evaluation benchmark for fundus image enhancement is indispensable for three main reasons: 1) The existing denoising metrics (e.g., PSNR, SSIM) are hardly to extend to downstream real-world clinical research (e.g., Vessel morphology consistency). 2) There is a lack of comprehensive evaluation for both paired and unpaired enhancement methods, along with the need for expert protocols to accurately assess clinical value. 3) An ideal evaluation system should provide insights to inform future developments of fundus image enhancement. To this end, we propose a novel comprehensive benchmark, EyeBench, to provide insights that align enhancement models with clinical needs, offering a foundation for future work to improve the clinical relevance and applicability of generative models for fundus image enhancement. EyeBench has three appealing properties: 1) multi-dimensional clinical alignment downstream evaluation: In addition to evaluating the enhancement task, we provide several clinically significant downstream tasks for fundus images, including vessel segmentation, DR grading, denoising generalization, and lesion segmentation. 2) Medical expert-guided evaluation design: We introduce a novel dataset that promote comprehensive and fair comparisons between paired and unpaired methods and includes a manual evaluation protocol by medical experts. 3) Valuable insights: Our benchmark study provides a comprehensive and rigorous evaluation of existing methods across different downstream tasks, assisting medical experts in making informed choices. Additionally, we offer further analysis of the challenges faced by existing methods. The code is available at \url{https://github.com/Retinal-Research/EyeBench}

Vamsi Krishna Vasa, Wenhui Zhu, Xiwen Chen et al.

In recent years, significant progress has been made in the medical image analysis domain using convolutional neural networks (CNNs). In particular, deep neural networks based on a U-shaped architecture (UNet) with skip connections have been adopted for several medical imaging tasks, including organ segmentation. Despite their great success, CNNs are not good at learning global or semantic features. Especially ones that require human-like reasoning to understand the context. Many UNet architectures attempted to adjust with the introduction of Transformer-based self-attention mechanisms, and notable gains in performance have been noted. However, the transformers are inherently flawed with redundancy to learn at shallow layers, which often leads to an increase in the computation of attention from the nearby pixels offering limited information. The recently introduced Super Token Attention (STA) mechanism adapts the concept of superpixels from pixel space to token space, using super tokens as compact visual representations. This approach tackles the redundancy by learning efficient global representations in vision transformers, especially for the shallow layers. In this work, we introduce the STA module in the UNet architecture (STA-UNet), to limit redundancy without losing rich information. Experimental results on four publicly available datasets demonstrate the superiority of STA-UNet over existing state-of-the-art architectures in terms of Dice score and IOU for organ segmentation tasks. The code is available at \url{https://github.com/Retinal-Research/STA-UNet}.

Xiwen Chen, Wenhui Zhu, Gen Li et al.

Multi-modal large language models (MLLMs) achieve strong visual-language reasoning but suffer from high inference cost due to redundant visual tokens. Recent work explores visual token pruning to accelerate inference, while existing pruning methods overlook the underlying distributional structure of visual representations. We propose OTPrune, a training-free framework that formulates pruning as distribution alignment via optimal transport (OT). By minimizing the 2-Wasserstein distance between the full and pruned token distributions, OTPrune preserves both local diversity and global representativeness while reducing inference cost. Moreover, we derive a tractable submodular objective that enables efficient optimization, and theoretically prove its monotonicity and submodularity, providing a principled foundation for stable and efficient pruning. We further provide a comprehensive analysis that explains how distributional alignment contributes to stable and semantically faithful pruning. Comprehensive experiments on wider benchmarks demonstrate that OTPrune achieves superior performance-efficiency tradeoffs compared to state-of-the-art methods. The code is available at https://github.com/xiwenc1/OTPrune.

Xuanzhao Dong, Wenhui Zhu, Yujian Xiong et al.

Color fundus photography (CFP) is central to diagnosing and monitoring retinal disease, yet its acquisition variability (e.g., illumination changes) often degrades image quality, which motivates robust enhancement methods. Unpaired enhancement pipelines are typically GAN-based, however, they can distort clinically critical vasculature, altering vessel topology and endpoint integrity. Motivated by these structural alterations, we propose Vessel-Aware Optimal Transport (\textbf{VAOT}), a framework that combines an optimal-transport objective with two structure-preserving regularizers: (i) a skeleton-based loss to maintain global vascular connectivity and (ii) an endpoint-aware loss to stabilize local termini. These constraints guide learning in the unpaired setting, reducing noise while preserving vessel structure. Experimental results on synthetic degradation benchmark and downstream evaluations in vessel and lesion segmentation demonstrate the superiority of the proposed methods against several state-of-the art baselines. The code is available at https://github.com/Retinal-Research/VAOT

Xiwen Chen, Peijie Qiu, Wenhui Zhu et al.

While multiple instance learning (MIL) has shown to be a promising approach for histopathological whole slide image (WSI) analysis, its reliance on permutation invariance significantly limits its capacity to effectively uncover semantic correlations between instances within WSIs. Based on our empirical and theoretical investigations, we argue that approaches that are not permutation-invariant but better capture spatial correlations between instances can offer more effective solutions. In light of these findings, we propose a novel alternative to existing MIL for WSI analysis by learning to restore the order of instances from their randomly shuffled arrangement. We term this task as cracking an instance jigsaw puzzle problem, where semantic correlations between instances are uncovered. To tackle the instance jigsaw puzzles, we propose a novel Siamese network solution, which is theoretically justified by optimal transport theory. We validate the proposed method on WSI classification and survival prediction tasks, where the proposed method outperforms the recent state-of-the-art MIL competitors. The code is available at https://github.com/xiwenc1/MIL-JigsawPuzzles.

Wenhui Zhu, Xuanzhao Dong, Xin Li et al.

Recently, reinforcement learning (RL)-based tuning has shifted the trajectory of Multimodal Large Language Models (MLLMs), particularly following the introduction of Group Relative Policy Optimization (GRPO). However, directly applying it to medical tasks remains challenging for achieving clinically grounded model behavior. Motivated by the need to align model response with clinical expectations, we investigate four critical dimensions that affect the effectiveness of RL-based tuning in medical visual question answering (VQA): base model initialization strategy, the role of medical semantic alignment, the impact of length-based rewards on long-chain reasoning, and the influence of bias. We conduct extensive experiments to analyze these factors for medical MLLMs, providing new insights into how models are domain-specifically fine-tuned. Additionally, our results also demonstrate that GRPO-based RL tuning consistently outperforms standard supervised fine-tuning (SFT) in both accuracy and reasoning quality.

Xin Li, Wenhui Zhu, Xuanzhao Dong et al.

With the rapid development of deep learning, CNN-based U-shaped networks have succeeded in medical image segmentation and are widely applied for various tasks. However, their limitations in capturing global features hinder their performance in complex segmentation tasks. The rise of Vision Transformer (ViT) has effectively compensated for this deficiency of CNNs and promoted the application of ViT-based U-networks in medical image segmentation. However, the high computational demands of ViT make it unsuitable for many medical devices and mobile platforms with limited resources, restricting its deployment on resource-constrained and edge devices. To address this, we propose EViT-UNet, an efficient ViT-based segmentation network that reduces computational complexity while maintaining accuracy, making it ideal for resource-constrained medical devices. EViT-UNet is built on a U-shaped architecture, comprising an encoder, decoder, bottleneck layer, and skip connections, combining convolutional operations with self-attention mechanisms to optimize efficiency. Experimental results demonstrate that EViT-UNet achieves high accuracy in medical image segmentation while significantly reducing computational complexity.

Hao Wang, Wenhui Zhu, Xuanzhao Dong et al.

In this work, we propose Many-MobileNet, an efficient model fusion strategy for retinal disease classification using lightweight CNN architecture. Our method addresses key challenges such as overfitting and limited dataset variability by training multiple models with distinct data augmentation strategies and different model complexities. Through this fusion technique, we achieved robust generalization in data-scarce domains while balancing computational efficiency with feature extraction capabilities.

Yujian Xiong, Xuanzhao Dong, Wenhui Zhu et al.

Retinal fundus photography is indispensable for ophthalmic screening and diagnosis, yet image quality is often degraded by noise, artifacts, and uneven illumination. Recent GAN- and diffusion-based enhancement methods improve perceptual quality by aligning degraded images with high-quality distributions, but our analysis shows that this focus can distort intra-class geometry: clinically related samples become dispersed, disease-class boundaries blur, and downstream tasks such as grading or lesion detection are harmed. The Gromov Wasserstein (GW) discrepancy offers a principled solution by aligning distributions through internal pairwise distances, naturally preserving intra-class structure, but its high computational cost restricts practical use. To overcome this, we propose SGW-GAN, the first framework to incorporate Sliced GW (SGW) into retinal image enhancement. SGW approximates GW via random projections, retaining relational fidelity while greatly reducing cost. Experiments on public datasets show that SGW-GAN produces visually compelling enhancements, achieves superior diabetic retinopathy grading, and reports the lowest GW discrepancy across disease labels, demonstrating both efficiency and clinical fidelity for unpaired medical image enhancement.

Hao Wang, Ashish Bastola, Chaoyi Zhou et al.

As generative models become increasingly capable of producing high-fidelity visual content, the demand for efficient, interpretable, and editable image representations has grown substantially. Recent advances in 2D Gaussian Splatting (2DGS) have emerged as a promising solution, offering explicit control, high interpretability, and real-time rendering capabilities (>1000 FPS). However, high-quality 2DGS typically requires post-optimization. Existing methods adopt random or heuristics (e.g., gradient maps), which are often insensitive to image complexity and lead to slow convergence (>10s). More recent approaches introduce learnable networks to predict initial Gaussian configurations, but at the cost of increased computational and architectural complexity. To bridge this gap, we present Fast-2DGS, a lightweight framework for efficient Gaussian image representation. Specifically, we introduce Deep Gaussian Prior, implemented as a conditional network to capture the spatial distribution of Gaussian primitives under different complexities. In addition, we propose an attribute regression network to predict dense Gaussian properties. Experiments demonstrate that this disentangled architecture achieves high-quality reconstruction in a single forward pass, followed by minimal fine-tuning. More importantly, our approach significantly reduces computational cost without compromising visual quality, bringing 2DGS closer to industry-ready deployment.

Wenhui Zhu, Xiwen Chen, Zhipeng Wang et al.

Instructed Visual Segmentation (IVS) tasks require segmenting objects in images or videos based on natural language instructions. While recent multimodal large language models (MLLMs) have achieved strong performance on IVS, their inference cost remains a major bottleneck, particularly in video. We empirically analyze visual token sampling in MLLMs and observe a strong correlation between subset token coverage and segmentation performance. This motivates our design of a simple and effective token pruning method that selects a compact yet spatially representative subset of tokens to accelerate inference. In this paper, we introduce a novel visual token pruning method for IVS, called EVTP-IV, which builds upon the k-center by integrating spatial information to ensure better coverage. We further provide an information-theoretic analysis to support our design. Experiments on standard IVS benchmarks show that our method achieves up to 5X speed-up on video tasks and 3.5X on image tasks, while maintaining comparable accuracy using only 20% of the tokens. Our method also consistently outperforms state-of-the-art pruning baselines under varying pruning ratios.

Yujian Xiong, Xuanzhao Dong, Sebastian Waz et al.

Ultra-high-field (7 Tesla) BOLD fMRI offers exceptional detail in both spatial and temporal domains, along with robust signal-to-noise characteristics, making it a powerful modality for studying visual information processing in the brain. However, due to the limited accessibility of 7T scanners, the majority of neuroimaging studies are still conducted using 3T systems, which inherently suffer from reduced fidelity in both resolution and SNR. To mitigate this limitation, we introduce a new computational approach designed to enhance the quality of 3T BOLD fMRI acquisitions. Specifically, we project both 3T and 7T datasets, sourced from different individuals and experimental setups, into a shared low-dimensional representation space. Within this space, we employ a lightweight, unsupervised Schrödinger Bridge framework to infer a high-SNR, high-resolution counterpart of the 3T data, without relying on paired supervision. This methodology is evaluated across multiple fMRI retinotopy datasets, including synthetically generated samples, and demonstrates a marked improvement in the reliability and fit of population receptive field (pRF) models applied to the enhanced 3T outputs. Our findings suggest that it is feasible to computationally approximate 7T-level quality from standard 3T acquisitions.