Yiqing Shen

Haoyu Wang, Sizheng Guo, Jin Ye et al.

Existing volumetric medical image segmentation models are typically task-specific, excelling at specific target but struggling to generalize across anatomical structures or modalities. This limitation restricts their broader clinical use. In this paper, we introduce SAM-Med3D for general-purpose segmentation on volumetric medical images. Given only a few 3D prompt points, SAM-Med3D can accurately segment diverse anatomical structures and lesions across various modalities. To achieve this, we gather and process a large-scale 3D medical image dataset, SA-Med3D-140K, from a blend of public sources and licensed private datasets. This dataset includes 22K 3D images and 143K corresponding 3D masks. Then SAM-Med3D, a promptable segmentation model characterized by the fully learnable 3D structure, is trained on this dataset using a two-stage procedure and exhibits impressive performance on both seen and unseen segmentation targets. We comprehensively evaluate SAM-Med3D on 16 datasets covering diverse medical scenarios, including different anatomical structures, modalities, targets, and zero-shot transferability to new/unseen tasks. The evaluation shows the efficiency and efficacy of SAM-Med3D, as well as its promising application to diverse downstream tasks as a pre-trained model. Our approach demonstrates that substantial medical resources can be utilized to develop a general-purpose medical AI for various potential applications. Our dataset, code, and models are available at https://github.com/uni-medical/SAM-Med3D.

Yiqing Shen, Yulin Luo, Dinggang Shen et al.

Stain variations often decrease the generalization ability of deep learning based approaches in digital histopathology analysis. Two separate proposals, namely stain normalization (SN) and stain augmentation (SA), have been spotlighted to reduce the generalization error, where the former alleviates the stain shift across different medical centers using template image and the latter enriches the accessible stain styles by the simulation of more stain variations. However, their applications are bounded by the selection of template images and the construction of unrealistic styles. To address the problems, we unify SN and SA with a novel RandStainNA scheme, which constrains variable stain styles in a practicable range to train a stain agnostic deep learning model. The RandStainNA is applicable to stain normalization in a collection of color spaces i.e. HED, HSV, LAB. Additionally, we propose a random color space selection scheme to gain extra performance improvement. We evaluate our method by two diagnostic tasks i.e. tissue subtype classification and nuclei segmentation, with various network backbones. The performance superiority over both SA and SN yields that the proposed RandStainNA can consistently improve the generalization ability, that our models can cope with more incoming clinical datasets with unpredicted stain styles. The codes is available at https://github.com/yiqings/RandStainNA.

Zhenqi He, Mathias Unberath, Jing Ke et al.

Nuclei appear small in size, yet, in real clinical practice, the global spatial information and correlation of the color or brightness contrast between nuclei and background, have been considered a crucial component for accurate nuclei segmentation. However, the field of automatic nuclei segmentation is dominated by Convolutional Neural Networks (CNNs), meanwhile, the potential of the recently prevalent Transformers has not been fully explored, which is powerful in capturing local-global correlations. To this end, we make the first attempt at a pure Transformer framework for nuclei segmentation, called TransNuSeg. Different from prior work, we decouple the challenging nuclei segmentation task into an intrinsic multi-task learning task, where a tri-decoder structure is employed for nuclei instance, nuclei edge, and clustered edge segmentation respectively. To eliminate the divergent predictions from different branches in previous work, a novel self distillation loss is introduced to explicitly impose consistency regulation between branches. Moreover, to formulate the high correlation between branches and also reduce the number of parameters, an efficient attention sharing scheme is proposed by partially sharing the self-attention heads amongst the tri-decoders. Finally, a token MLP bottleneck replaces the over-parameterized Transformer bottleneck for a further reduction in model complexity. Experiments on two datasets of different modalities, including MoNuSeg have shown that our methods can outperform state-of-the-art counterparts such as CA2.5-Net by 2-3% Dice with 30% fewer parameters. In conclusion, TransNuSeg confirms the strength of Transformer in the context of nuclei segmentation, which thus can serve as an efficient solution for real clinical practice. Code is available at https://github.com/zhenqi-he/transnuseg.

Zhenqi He, Junjun He, Jin Ye et al.

Histological whole slide images (WSIs) can be usually compromised by artifacts, such as tissue folding and bubbles, which will increase the examination difficulty for both pathologists and Computer-Aided Diagnosis (CAD) systems. Existing approaches to restoring artifact images are confined to Generative Adversarial Networks (GANs), where the restoration process is formulated as an image-to-image transfer. Those methods are prone to suffer from mode collapse and unexpected mistransfer in the stain style, leading to unsatisfied and unrealistic restored images. Innovatively, we make the first attempt at a denoising diffusion probabilistic model for histological artifact restoration, namely ArtiFusion.Specifically, ArtiFusion formulates the artifact region restoration as a gradual denoising process, and its training relies solely on artifact-free images to simplify the training complexity.Furthermore, to capture local-global correlations in the regional artifact restoration, a novel Swin-Transformer denoising architecture is designed, along with a time token scheme. Our extensive evaluations demonstrate the effectiveness of ArtiFusion as a pre-processing method for histology analysis, which can successfully preserve the tissue structures and stain style in artifact-free regions during the restoration. Code is available at https://github.com/zhenqi-he/ArtiFusion.

Yiqing Shen, Bingxin Zhou, Xinye Xiong et al.

Gigapixel medical images provide massive data, both morphological textures and spatial information, to be mined. Due to the large data scale in histology, deep learning methods play an increasingly significant role as feature extractors. Existing solutions heavily rely on convolutional neural networks (CNNs) for global pixel-level analysis, leaving the underlying local geometric structure such as the interaction between cells in the tumor microenvironment unexplored. The topological structure in medical images, as proven to be closely related to tumor evolution, can be well characterized by graphs. To obtain a more comprehensive representation for downstream oncology tasks, we propose a fusion framework for enhancing the global image-level representation captured by CNNs with the geometry of cell-level spatial information learned by graph neural networks (GNN). The fusion layer optimizes an integration between collaborative features of global images and cell graphs. Two fusion strategies have been developed: one with MLP which is simple but turns out efficient through fine-tuning, and the other with Transformer gains a champion in fusing multiple networks. We evaluate our fusion strategies on histology datasets curated from large patient cohorts of colorectal and gastric cancers for three biomarker prediction tasks. Both two models outperform plain CNNs or GNNs, reaching a consistent AUC improvement of more than 5% on various network backbones. The experimental results yield the necessity for combining image-level morphological features with cell spatial relations in medical image analysis. Codes are available at https://github.com/yiqings/HEGnnEnhanceCnn.

Zhongying Deng, Cheng Tang, Ziyan Huang et al. · pku

Foundation models have demonstrated remarkable success across diverse domains and tasks, primarily due to the thrive of large-scale, diverse, and high-quality datasets. However, in the field of medical imaging, the curation and assembling of such medical datasets are highly challenging due to the reliance on clinical expertise and strict ethical and privacy constraints, resulting in a scarcity of large-scale unified medical datasets and hindering the development of powerful medical foundation models. In this work, we present the largest survey to date of medical image datasets, covering over 1,000 open-access datasets with a systematic catalog of their modalities, tasks, anatomies, annotations, limitations, and potential for integration. Our analysis exposes a landscape that is modest in scale, fragmented across narrowly scoped tasks, and unevenly distributed across organs and modalities, which in turn limits the utility of existing medical image datasets for developing versatile and robust medical foundation models. To turn fragmentation into scale, we propose a metadata-driven fusion paradigm (MDFP) that integrates public datasets with shared modalities or tasks, thereby transforming multiple small data silos into larger, more coherent resources. Building on MDFP, we release an interactive discovery portal that enables end-to-end, automated medical image dataset integration, and compile all surveyed datasets into a unified, structured table that clearly summarizes their key characteristics and provides reference links, offering the community an accessible and comprehensive repository. By charting the current terrain and offering a principled path to dataset consolidation, our survey provides a practical roadmap for scaling medical imaging corpora, supporting faster data discovery, more principled dataset creation, and more capable medical foundation models.

Yiqing Shen, Pengfei Guo, Jingpu Wu et al.

The synergy of long-range dependencies from transformers and local representations of image content from convolutional neural networks (CNNs) has led to advanced architectures and increased performance for various medical image analysis tasks due to their complementary benefits. However, compared with CNNs, transformers require considerably more training data, due to a larger number of parameters and an absence of inductive bias. The need for increasingly large datasets continues to be problematic, particularly in the context of medical imaging, where both annotation efforts and data protection result in limited data availability. In this work, inspired by the human decision-making process of correlating new evidence with previously memorized experience, we propose a Memorizing Vision Transformer (MoViT) to alleviate the need for large-scale datasets to successfully train and deploy transformer-based architectures. MoViT leverages an external memory structure to cache history attention snapshots during the training stage. To prevent overfitting, we incorporate an innovative memory update scheme, attention temporal moving average, to update the stored external memories with the historical moving average. For inference speedup, we design a prototypical attention learning method to distill the external memory into smaller representative subsets. We evaluate our method on a public histology image dataset and an in-house MRI dataset, demonstrating that MoViT applied to varied medical image analysis tasks, can outperform vanilla transformer models across varied data regimes, especially in cases where only a small amount of annotated data is available. More importantly, MoViT can reach a competitive performance of ViT with only 3.0% of the training data.

Yiqing Shen, Yuyin Zhou, Lequan Yu

Federated learning (FL) is a distributed learning paradigm that enables multiple clients to collaboratively learn a shared global model. Despite the recent progress, it remains challenging to deal with heterogeneous data clients, as the discrepant data distributions usually prevent the global model from delivering good generalization ability on each participating client. In this paper, we propose CD^2-pFed, a novel Cyclic Distillation-guided Channel Decoupling framework, to personalize the global model in FL, under various settings of data heterogeneity. Different from previous works which establish layer-wise personalization to overcome the non-IID data across different clients, we make the first attempt at channel-wise assignment for model personalization, referred to as channel decoupling. To further facilitate the collaboration between private and shared weights, we propose a novel cyclic distillation scheme to impose a consistent regularization between the local and global model representations during the federation. Guided by the cyclical distillation, our channel decoupling framework can deliver more accurate and generalized results for different kinds of heterogeneity, such as feature skew, label distribution skew, and concept shift. Comprehensive experiments on four benchmarks, including natural image and medical image analysis tasks, demonstrate the consistent effectiveness of our method on both local and external validations.

Jun Peng, Chuanlei Zhai, Guoxi Ni et al.

Due to its excellent shock-capturing capability and high resolution, the WENO scheme family has been widely used in varieties of compressive flow simulation. However, for problems containing strong shocks and contact discontinuities, such as the Lax shock tube problem, the WENO scheme still produces numerical oscillations. To avoid such numerical oscillations, the characteristic-wise construction method should be applied. Compared to component-wise reconstruction, characteristic-wise reconstruction leads to much more computational cost and thus is not suite for large scale simulation such as direct numeric simulation of turbulence. In this paper, an adaptive characteristic-wise reconstruction WENO scheme, i.e. the AdaWENO scheme, is proposed to improve the computational efficiency of the characteristic-wise reconstruction method. The new scheme performs characteristic-wise reconstruction near discontinuities while switching to component-wise reconstruction for smooth regions. Meanwhile, a new calculation strategy for the WENO smoothness indicators is implemented to reduce over-all computational cost. Several one dimensional and two dimensional numerical tests are performed to validate and evaluate the AdaWENO scheme. Numerical results show that AdaWENO maintains essentially non-oscillatory flow field near discontinuities as the characteristic-wise reconstruction method. Besieds, compared to the component-wise reconstruction, AdaWENO is about 40\% faster which indicates its excellent efficiency.

Kai Yi, Bingxin Zhou, Yiqing Shen et al.

Inverse protein folding is challenging due to its inherent one-to-many mapping characteristic, where numerous possible amino acid sequences can fold into a single, identical protein backbone. This task involves not only identifying viable sequences but also representing the sheer diversity of potential solutions. However, existing discriminative models, such as transformer-based auto-regressive models, struggle to encapsulate the diverse range of plausible solutions. In contrast, diffusion probabilistic models, as an emerging genre of generative approaches, offer the potential to generate a diverse set of sequence candidates for determined protein backbones. We propose a novel graph denoising diffusion model for inverse protein folding, where a given protein backbone guides the diffusion process on the corresponding amino acid residue types. The model infers the joint distribution of amino acids conditioned on the nodes' physiochemical properties and local environment. Moreover, we utilize amino acid replacement matrices for the diffusion forward process, encoding the biologically-meaningful prior knowledge of amino acids from their spatial and sequential neighbors as well as themselves, which reduces the sampling space of the generative process. Our model achieves state-of-the-art performance over a set of popular baseline methods in sequence recovery and exhibits great potential in generating diverse protein sequences for a determined protein backbone structure.

Open-H-Embodiment Consortium, Nigel Nelson, Juo-Tung Chen et al.

Autonomous medical robots hold promise to improve patient outcomes, reduce provider workload, democratize access to care, and enable superhuman precision. However, autonomous medical robotics has been limited by a fundamental data problem: existing medical robotic datasets are small, single-embodiment, and rarely shared openly, restricting the development of foundation models that the field needs to advance. We introduce Open-H-Embodiment, the largest open dataset of medical robotic video with synchronized kinematics to date, spanning more than 49 institutions and multiple robotic platforms including the CMR Versius, Intuitive Surgical's da Vinci, da Vinci Research Kit (dVRK), Rob Surgical BiTrack, Virtual Incision's MIRA, Moon Surgical Maestro, and a variety of custom systems, spanning surgical manipulation, robotic ultrasound, and endoscopy procedures. We demonstrate the research enabled by this dataset through two foundation models. GR00T-H is the first open foundation vision-language-action model for medical robotics, which is the only evaluated model to achieve full end-to-end task completion on a structured suturing benchmark (25% of trials vs. 0% for all others) and achieves 64% average success across a 29-step ex vivo suturing sequence. We also train Cosmos-H-Surgical-Simulator, the first action-conditioned world model to enable multi-embodiment surgical simulation from a single checkpoint, spanning nine robotic platforms and supporting in silico policy evaluation and synthetic data generation for the medical domain. These results suggest that open, large-scale medical robot data collection can serve as critical infrastructure for the research community, enabling advances in robot learning, world modeling, and beyond.

Yiqing Shen, Zan Chen, Michail Mamalakis et al.

The structural similarities between protein sequences and natural languages have led to parallel advancements in deep learning across both domains. While large language models (LLMs) have achieved much progress in the domain of natural language processing, their potential in protein engineering remains largely unexplored. Previous approaches have equipped LLMs with protein understanding capabilities by incorporating external protein encoders, but this fails to fully leverage the inherent similarities between protein sequences and natural languages, resulting in sub-optimal performance and increased model complexity. To address this gap, we present TourSynbio-7B, the first multi-modal large model specifically designed for protein engineering tasks without external protein encoders. TourSynbio-7B demonstrates that LLMs can inherently learn to understand proteins as language. The model is post-trained and instruction fine-tuned on InternLM2-7B using ProteinLMDataset, a dataset comprising 17.46 billion tokens of text and protein sequence for self-supervised pretraining and 893K instructions for supervised fine-tuning. TourSynbio-7B outperforms GPT-4 on the ProteinLMBench, a benchmark of 944 manually verified multiple-choice questions, with 62.18% accuracy. Leveraging TourSynbio-7B's enhanced protein sequence understanding capability, we introduce TourSynbio-Agent, an innovative framework capable of performing various protein engineering tasks, including mutation analysis, inverse folding, protein folding, and visualization. TourSynbio-Agent integrates previously disconnected deep learning models in the protein engineering domain, offering a unified conversational user interface for improved usability. Finally, we demonstrate the efficacy of TourSynbio-7B and TourSynbio-Agent through two wet lab case studies on vanilla key enzyme modification and steroid compound catalysis.

Han Zhang, Yiqing Shen, Roger D. Soberanis-Mukul et al.

Developing embodied AI for intelligent surgical systems requires safe, controllable environments for continual learning and evaluation. However, safety regulations and operational constraints in operating rooms (ORs) limit embodied agents from freely perceiving and interacting in realistic settings. Digital twins provide high-fidelity, risk-free environments for exploration and training. How we may create photorealistic and dynamic digital representations of ORs that capture relevant spatial, visual, and behavioral complexity remains unclear. We introduce TwinOR, a framework for constructing photorealistic, dynamic digital twins of ORs for embodied AI research. The system reconstructs static geometry from pre-scan videos and continuously models human and equipment motion through multi-view perception of OR activities. The static and dynamic components are fused into an immersive 3D environment that supports controllable simulation and embodied exploration. The proposed framework reconstructs complete OR geometry with centimeter level accuracy while preserving dynamic interaction across surgical workflows, enabling realistic renderings and a virtual playground for embodied AI systems. In our experiments, TwinOR simulates stereo and monocular sensor streams for geometry understanding and visual localization tasks. Models such as FoundationStereo and ORB-SLAM3 on TwinOR-synthesized data achieve performance within their reported accuracy on real indoor datasets, demonstrating that TwinOR provides sensor-level realism sufficient for perception and localization challenges. By establishing a real-to-sim pipeline for constructing dynamic, photorealistic digital twins of OR environments, TwinOR enables the safe, scalable, and data-efficient development and benchmarking of embodied AI, ultimately accelerating the deployment of embodied AI from sim-to-real.

Yiqing Shen, Liwu Xu, Yuzhe Yang et al.

Mixed Sample Regularization (MSR), such as MixUp or CutMix, is a powerful data augmentation strategy to generalize convolutional neural networks. Previous empirical analysis has illustrated an orthogonal performance gain between MSR and conventional offline Knowledge Distillation (KD). To be more specific, student networks can be enhanced with the involvement of MSR in the training stage of sequential distillation. Yet, the interplay between MSR and online knowledge distillation, where an ensemble of peer students learn mutually from each other, remains unexplored. To bridge the gap, we make the first attempt at incorporating CutMix into online distillation, where we empirically observe a significant improvement. Encouraged by this fact, we propose an even stronger MSR specifically for online distillation, named as Cut\textsuperscript{n}Mix. Furthermore, a novel online distillation framework is designed upon Cut\textsuperscript{n}Mix, to enhance the distillation with feature level mutual learning and a self-ensemble teacher. Comprehensive evaluations on CIFAR10 and CIFAR100 with six network architectures show that our approach can consistently outperform state-of-the-art distillation methods.

Yiqing Shen, Baiyun Liu, Ruize Yu et al.

Chronic Obstructive Pulmonary Disease (COPD) is the fourth leading cause of death worldwide. Yet, COPD diagnosis heavily relies on spirometric examination as well as functional airway limitation, which may cause a considerable portion of COPD patients underdiagnosed especially at the early stage. Recent advance in deep learning (DL) has shown their promising potential in COPD identification from CT images. However, with heterogeneous syndromes and distinct phenotypes, DL models trained with CTs from one data center fail to generalize on images from another center. Due to privacy regularizations, a collaboration of distributed CT images into one centralized center is not feasible. Federated learning (FL) approaches enable us to train with distributed private data. Yet, routine FL solutions suffer from performance degradation in the case where COPD CTs are not independent and identically distributed (Non-IID). To address this issue, we propose a novel personalized federated learning (PFL) method based on vision transformer (ViT) for distributed and heterogeneous COPD CTs. To be more specific, we partially personalize some heads in multiheaded self-attention layers to learn the personalized attention for local data and retain the other heads shared to extract the common attention. To the best of our knowledge, this is the first proposal of a PFL framework specifically for ViT to identify COPD. Our evaluation of a dataset set curated from six medical centers shows our method outperforms the PFL approaches for convolutional neural networks.

Erfan Darzi, Yiqing Shen, Yangming Ou et al.

Optimization-based regularization methods have been effective in addressing the challenges posed by data heterogeneity in medical federated learning, particularly in improving the performance of underrepresented clients. However, these methods often lead to lower overall model accuracy and slower convergence rates. In this paper, we demonstrate that using Vision Transformers can substantially improve the performance of underrepresented clients without a significant trade-off in overall accuracy. This improvement is attributed to the Vision transformer's ability to capture long-range dependencies within the input data.

Daqin Luo, Chengjian Feng, Yuxuan Nong et al.

Automated Machine Learning (AutoML) offers a promising approach to streamline the training of machine learning models. However, existing AutoML frameworks are often limited to unimodal scenarios and require extensive manual configuration. Recent advancements in Large Language Models (LLMs) have showcased their exceptional abilities in reasoning, interaction, and code generation, presenting an opportunity to develop a more automated and user-friendly framework. To this end, we introduce AutoM3L, an innovative Automated Multimodal Machine Learning framework that leverages LLMs as controllers to automatically construct multimodal training pipelines. AutoM3L comprehends data modalities and selects appropriate models based on user requirements, providing automation and interactivity. By eliminating the need for manual feature engineering and hyperparameter optimization, our framework simplifies user engagement and enables customization through directives, addressing the limitations of previous rule-based AutoML approaches. We evaluate the performance of AutoM3L on six diverse multimodal datasets spanning classification, regression, and retrieval tasks, as well as a comprehensive set of unimodal datasets. The results demonstrate that AutoM3L achieves competitive or superior performance compared to traditional rule-based AutoML methods. Furthermore, a user study highlights the user-friendliness and usability of our framework, compared to the rule-based AutoML methods.

Chong Wang, Mengyao Li, Junjun He et al.

Recent breakthroughs in large language models (LLMs) offer unprecedented natural language understanding and generation capabilities. However, existing surveys on LLMs in biomedicine often focus on specific applications or model architectures, lacking a comprehensive analysis that integrates the latest advancements across various biomedical domains. This review, based on an analysis of 484 publications sourced from databases including PubMed, Web of Science, and arXiv, provides an in-depth examination of the current landscape, applications, challenges, and prospects of LLMs in biomedicine, distinguishing itself by focusing on the practical implications of these models in real-world biomedical contexts. Firstly, we explore the capabilities of LLMs in zero-shot learning across a broad spectrum of biomedical tasks, including diagnostic assistance, drug discovery, and personalized medicine, among others, with insights drawn from 137 key studies. Then, we discuss adaptation strategies of LLMs, including fine-tuning methods for both uni-modal and multi-modal LLMs to enhance their performance in specialized biomedical contexts where zero-shot fails to achieve, such as medical question answering and efficient processing of biomedical literature. Finally, we discuss the challenges that LLMs face in the biomedicine domain including data privacy concerns, limited model interpretability, issues with dataset quality, and ethics due to the sensitive nature of biomedical data, the need for highly reliable model outputs, and the ethical implications of deploying AI in healthcare. To address these challenges, we also identify future research directions of LLM in biomedicine including federated learning methods to preserve data privacy and integrating explainable AI methodologies to enhance the transparency of LLMs.

Yizhu Wen, Kai Yi, Jing Ke et al.

Tabular data plays a crucial role in various domains but often suffers from missing values, thereby curtailing its potential utility. Traditional imputation techniques frequently yield suboptimal results and impose substantial computational burdens, leading to inaccuracies in subsequent modeling tasks. To address these challenges, we propose DiffImpute, a novel Denoising Diffusion Probabilistic Model (DDPM). Specifically, DiffImpute is trained on complete tabular datasets, ensuring that it can produce credible imputations for missing entries without undermining the authenticity of the existing data. Innovatively, it can be applied to various settings of Missing Completely At Random (MCAR) and Missing At Random (MAR). To effectively handle the tabular features in DDPM, we tailor four tabular denoising networks, spanning MLP, ResNet, Transformer, and U-Net. We also propose Harmonization to enhance coherence between observed and imputed data by infusing the data back and denoising them multiple times during the sampling stage. To enable efficient inference while maintaining imputation performance, we propose a refined non-Markovian sampling process that works along with Harmonization. Empirical evaluations on seven diverse datasets underscore the prowess of DiffImpute. Specifically, when paired with the Transformer as the denoising network, it consistently outperforms its competitors, boasting an average ranking of 1.7 and the most minimal standard deviation. In contrast, the next best method lags with a ranking of 2.8 and a standard deviation of 0.9. The code is available at https://github.com/Dendiiiii/DiffImpute.

Aiza Maksutova, Lalithkumar Seenivasan, Hao Ding et al.

Surgical action automation has progressed rapidly toward achieving surgeon-like dexterous control, driven primarily by advances in learning from demonstration and vision-language-action models. While these have demonstrated success in table-top experiments, translating them to clinical deployment remains challenging: current methods offer limited predictability on where instruments will interact on tissue surfaces and lack explicit conditioning inputs to enforce tool-action-specific safe interaction regions. Addressing this gap, we introduce AffordTissue, a multimodal framework for predicting tool-action specific tissue affordance regions as dense heatmaps during cholecystectomy. Our approach combines a temporal vision encoder capturing tool motion and tissue dynamics across multiple viewpoints, language conditioning enabling generalization across diverse instrument-action pairs, and a DiT-style decoder for dense affordance prediction. We establish the first tissue affordance benchmark by curating and annotating 15,638 video clips across 103 cholecystectomy procedures, covering six unique tool-action pairs involving four instruments (hook, grasper, scissors, clipper) and their associated tasks: dissection, grasping, clipping, and cutting. Experiments demonstrate substantial improvement over vision-language model baselines (20.6 px ASSD vs. 60.2 px for Molmo-VLM), showing that our task-specific architecture outperforms large-scale foundation models for dense surgical affordance prediction. By predicting tool-action specific tissue affordance regions, AffordTissue provides explicit spatial reasoning for safe surgical automation, potentially unlocking explicit policy guidance toward appropriate tissue regions and early safe stop when instruments deviate outside predicted safe zones.

Zijie Lin, Yiqing Shen, Qilin Cai et al.

Machine Learning (ML) research is spread through academic papers featuring rich multimodal content, including text, diagrams, and tabular results. However, translating these multimodal elements into executable code remains a challenging and time-consuming process that requires substantial ML expertise. We introduce ``Paper-to-Code'' (P2C), a novel task that transforms the multimodal content of scientific publications into fully executable code repositories, which extends beyond the existing formulation of code generation that merely converts textual descriptions into isolated code snippets. To automate the P2C process, we propose AutoP2C, a multi-agent framework based on large language models that processes both textual and visual content from research papers to generate complete code repositories. Specifically, AutoP2C contains four stages: (1) repository blueprint extraction from established codebases, (2) multimodal content parsing that integrates information from text, equations, and figures, (3) hierarchical task decomposition for structured code generation, and (4) iterative feedback-driven debugging to ensure functionality and performance. Evaluation on a benchmark of eight research papers demonstrates the effectiveness of AutoP2C, which can successfully generate executable code repositories for all eight papers, while OpenAI-o1 or DeepSeek-R1 can only produce runnable code for one paper. The code is available at https://github.com/shoushouyu/Automated-Paper-to-Code.

Dominik Schneider, Lalithkumar Seenivasan, Sampath Rapuri et al.

Purpose: Curating large-scale datasets of operating room (OR) workflow, encompassing rare, safety-critical, or atypical events, remains operationally and ethically challenging. This data bottleneck complicates the development of ambient intelligence for detecting, understanding, and mitigating rare or safety-critical events in the OR. Methods: This work presents an OR video diffusion framework that enables controlled synthesis of rare and safety-critical events. The framework integrates a geometric abstraction module, a conditioning module, and a fine-tuned diffusion model to first transform OR scenes into abstract geometric representations, then condition the synthesis process, and finally generate realistic OR event videos. Using this framework, we also curate a synthetic dataset to train and validate AI models for detecting near-misses of sterile-field violations. Results: In synthesizing routine OR events, our method outperforms off-the-shelf video diffusion baselines, achieving lower FVD/LPIPS and higher SSIM/PSNR in both in- and out-of-domain datasets. Through qualitative results, we illustrate its ability for controlled video synthesis of counterfactual events. An AI model trained and validated on the generated synthetic data achieved a RECALL of 70.13% in detecting near safety-critical events. Finally, we conduct an ablation study to quantify performance gains from key design choices. Conclusion: Our solution enables controlled synthesis of routine and rare OR events from abstract geometric representations. Beyond demonstrating its capability to generate rare and safety-critical scenarios, we show its potential to support the development of ambient intelligence models.

Yuangang Li, Yiqing Shen, Yi Nian et al.

Large language models (LLMs) exhibit logically inconsistent hallucinations that appear coherent yet violate reasoning principles, with recent research suggesting an inverse relationship between causal reasoning capabilities and such hallucinations. However, existing reasoning approaches in LLMs, such as Chain-of-Thought (CoT) and its graph-based variants, operate at the linguistic token level rather than modeling the underlying causal relationships between variables, lacking the ability to represent conditional independencies or satisfy causal identification assumptions. To bridge this gap, we introduce causal-DAG construction and reasoning (CDCR-SFT), a supervised fine-tuning framework that trains LLMs to explicitly construct variable-level directed acyclic graph (DAG) and then perform reasoning over it. Moreover, we present a dataset comprising 25,368 samples (CausalDR), where each sample includes an input question, explicit causal DAG, graph-based reasoning trace, and validated answer. Experiments on four LLMs across eight tasks show that CDCR-SFT improves the causal reasoning capability with the state-of-the-art 95.33% accuracy on CLADDER (surpassing human performance of 94.8% for the first time) and reduces the hallucination on HaluEval with 10% improvements. It demonstrates that explicit causal structure modeling in LLMs can effectively mitigate logical inconsistencies in LLM outputs. Code is available at https://github.com/MrLYG/CDCR-SFT.

Yiqing Shen, Chenxiao Fan, Chenjia Li et al.

The goal of text-to-video retrieval is to search large databases for relevant videos based on text queries. Existing methods have progressed to handling explicit queries where the visual content of interest is described explicitly; however, they fail with implicit queries where identifying videos relevant to the query requires reasoning. We introduce reasoning text-to-video retrieval, a paradigm that extends traditional retrieval to process implicit queries through reasoning while providing object-level grounding masks that identify which entities satisfy the query conditions. Instead of relying on vision-language models directly, we propose representing video content as digital twins, i.e., structured scene representations that decompose salient objects through specialist vision models. This approach is beneficial because it enables large language models to reason directly over long-horizon video content without visual token compression. Specifically, our two-stage framework first performs compositional alignment between decomposed sub-queries and digital twin representations for candidate identification, then applies large language model-based reasoning with just-in-time refinement that invokes additional specialist models to address information gaps. We construct a benchmark of 447 manually created implicit queries with 135 videos (ReasonT2VBench-135) and another more challenging version of 1000 videos (ReasonT2VBench-1000). Our method achieves 81.2% R@1 on ReasonT2VBench-135, outperforming the strongest baseline by greater than 50 percentage points, and maintains 81.7% R@1 on the extended configuration while establishing state-of-the-art results in three conventional benchmarks (MSR-VTT, MSVD, and VATEX).

Yiqing Shen, Mathias Unberath

Reasoning segmentation enables open-set object segmentation via implicit text queries, therefore serving as a foundation for embodied agents that should operate autonomously in real-world environments. However, existing methods for reasoning segmentation require multimodal large language models with billions of parameters that exceed the computational capabilities of edge devices that typically deploy the embodied AI systems. Distillation offers a pathway to compress these models while preserving their capabilities. Yet, existing distillation approaches fail to transfer the multi-step reasoning capabilities that reasoning segmentation demands, as they focus on matching output predictions and intermediate features rather than preserving reasoning chains. The emerging paradigm of reasoning over digital twin representations presents an opportunity for more effective distillation by re-framing the problem. Consequently, we propose FastReasonSeg, which employs digital twin representations that decouple perception from reasoning to enable more effective distillation. Our distillation scheme first relies on supervised fine-tuning on teacher-generated reasoning chains. Then it is followed by reinforcement fine-tuning with joint rewards evaluating both segmentation accuracy and reasoning quality alignment. Experiments on two video (JiTBench, RVTBench) and two image benchmarks (ReasonSeg, LLM-Seg40K) demonstrate that our FastReasonSeg achieves state-of-the-art reasoning segmentation performance. Moreover, the distilled 0.6B variant outperforms models with 20 times more parameters while achieving 7.79 FPS throughput with only 2.1GB memory consumption. This efficiency enables deployment in resource-constrained environments to enable real-time reasoning segmentation.

Yiqing Shen, Mathias Unberath

Visual reasoning may require models to interpret images and videos and respond to implicit text queries across diverse output formats, from pixel-level segmentation masks to natural language descriptions. Existing approaches rely on supervised fine-tuning with task-specific architectures. For example, reasoning segmentation, grounding, summarization, and visual question answering each demand distinct model designs and training, preventing unified solutions and limiting cross-task and cross-modality generalization. Hence, we propose DT-R1, a reinforcement learning framework that trains large language models to construct digital twin representations of complex multi-modal visual inputs and then reason over these high-level representations as a unified approach to visual reasoning. Specifically, we train DT-R1 using GRPO with a novel reward that validates both structural integrity and output accuracy. Evaluations in six visual reasoning benchmarks, covering two modalities and four task types, demonstrate that DT-R1 consistently achieves improvements over state-of-the-art task-specific models. DT-R1 opens a new direction where visual reasoning emerges from reinforcement learning with digital twin representations.

Xinyuan Shao, Yiqing Shen, Mathias Unberath

Segment Anything Models (SAMs) have gained increasing attention in medical image analysis due to their zero-shot generalization capability in segmenting objects of unseen classes and domains when provided with appropriate user prompts. Addressing this performance gap is important to fully leverage the pre-trained weights of SAMs, particularly in the domain of volumetric medical image segmentation, where accuracy is important but well-annotated 3D medical data for fine-tuning is limited. In this work, we investigate whether introducing the memory mechanism as a plug-in, specifically the ability to memorize and recall internal representations of past inputs, can improve the performance of SAM with limited computation cost. To this end, we propose Memorizing SAM, a novel 3D SAM architecture incorporating a memory Transformer as a plug-in. Unlike conventional memorizing Transformers that save the internal representation during training or inference, our Memorizing SAM utilizes existing highly accurate internal representation as the memory source to ensure the quality of memory. We evaluate the performance of Memorizing SAM in 33 categories from the TotalSegmentator dataset, which indicates that Memorizing SAM can outperform state-of-the-art 3D SAM variant i.e., FastSAM3D with an average Dice increase of 11.36% at the cost of only 4.38 millisecond increase in inference time. The source code is publicly available at https://github.com/swedfr/memorizingSAM

Yijia Wang, Yiqing Shen, Weiming Chen et al.

Existing image editing methods can handle simple editing instructions very well. To deal with complex editing instructions, they often need to jointly fine-tune the large language models (LLMs) and diffusion models (DMs), which involves very high computational complexity and training cost. To address this issue, we propose a new method, called \textbf{C}omplex \textbf{I}mage \textbf{E}diting via \textbf{L}LM \textbf{R}easoning (CIELR), which converts a complex user instruction into a set of simple and explicit editing actions, eliminating the need for jointly fine-tuning the large language models and diffusion models. Specifically, we first construct a structured semantic representation of the input image using foundation models. Then, we introduce an iterative update mechanism that can progressively refine this representation, obtaining a fine-grained visual representation of the image scene. This allows us to perform complex and flexible image editing tasks. Extensive experiments on the SmartEdit Reasoning Scenario Set show that our method surpasses the previous state-of-the-art by 9.955 dB in PSNR, indicating its superior preservation of regions that should remain consistent. Due to the limited number of samples of public datasets of complex image editing with reasoning, we construct a benchmark named CIEBench, containing 86 image samples, together with a metric specifically for reasoning-based image editing. CIELR also outperforms previous methods on this benchmark. The code and dataset are available at \href{https://github.com/Jia-shao/Reasoning-Editing}{https://github.com/Jia-shao/Reasoning-Editing}.

Congying Liu, Xingyuan Wei, Peipei Liu et al.

Biomedical queries often rely on a deep understanding of specialized knowledge such as gene regulatory mechanisms and pathological processes of diseases. They require detailed analysis of complex physiological processes and effective integration of information from multiple data sources to support accurate retrieval and reasoning. Although large language models (LLMs) perform well in general reasoning tasks, their generated biomedical content often lacks scientific rigor due to the inability to access authoritative biomedical databases and frequently fabricates protein functions, interactions, and structural details that deviate from authentic information. Therefore, we present BioMedSearch, a multi-source biomedical information retrieval framework based on LLMs. The method integrates literature retrieval, protein database and web search access to support accurate and efficient handling of complex biomedical queries. Through sub-queries decomposition, keywords extraction, task graph construction, and multi-source information filtering, BioMedSearch generates high-quality question-answering results. To evaluate the accuracy of question answering, we constructed a multi-level dataset, BioMedMCQs, consisting of 3,000 questions. The dataset covers three levels of reasoning: mechanistic identification, non-adjacent semantic integration, and temporal causal reasoning, and is used to assess the performance of BioMedSearch and other methods on complex QA tasks. Experimental results demonstrate that BioMedSearch consistently improves accuracy over all baseline models across all levels. Specifically, at Level 1, the average accuracy increases from 59.1% to 91.9%; at Level 2, it rises from 47.0% to 81.0%; and at the most challenging Level 3, the average accuracy improves from 36.3% to 73.4%. The code and BioMedMCQs are available at: https://github.com/CyL-ucas/BioMed_Search

Yiqing Shen, Jingxing Li, Xinyuan Shao et al.

Segment anything models (SAMs) are gaining attention for their zero-shot generalization capability in segmenting objects of unseen classes and in unseen domains when properly prompted. Interactivity is a key strength of SAMs, allowing users to iteratively provide prompts that specify objects of interest to refine outputs. However, to realize the interactive use of SAMs for 3D medical imaging tasks, rapid inference times are necessary. High memory requirements and long processing delays remain constraints that hinder the adoption of SAMs for this purpose. Specifically, while 2D SAMs applied to 3D volumes contend with repetitive computation to process all slices independently, 3D SAMs suffer from an exponential increase in model parameters and FLOPS. To address these challenges, we present FastSAM3D which accelerates SAM inference to 8 milliseconds per 128*128*128 3D volumetric image on an NVIDIA A100 GPU. This speedup is accomplished through 1) a novel layer-wise progressive distillation scheme that enables knowledge transfer from a complex 12-layer ViT-B to a lightweight 6-layer ViT-Tiny variant encoder without training from scratch; and 2) a novel 3D sparse flash attention to replace vanilla attention operators, substantially reducing memory needs and improving parallelization. Experiments on three diverse datasets reveal that FastSAM3D achieves a remarkable speedup of 527.38x compared to 2D SAMs and 8.75x compared to 3D SAMs on the same volumes without significant performance decline. Thus, FastSAM3D opens the door for low-cost truly interactive SAM-based 3D medical imaging segmentation with commonly used GPU hardware. Code is available at https://github.com/arcadelab/FastSAM3D.

Yiqing Shen, Liwu Xu, Yuzhe Yang et al.

Knowledge distillation (KD) shows a bright promise as a powerful regularization strategy to boost generalization ability by leveraging learned sample-level soft targets. Yet, employing a complex pre-trained teacher network or an ensemble of peer students in existing KD is both time-consuming and computationally costly. Various self KD methods have been proposed to achieve higher distillation efficiency. However, they either require extra network architecture modification or are difficult to parallelize. To cope with these challenges, we propose an efficient and reliable self-distillation framework, named Self-Distillation from Last Mini-Batch (DLB). Specifically, we rearrange the sequential sampling by constraining half of each mini-batch coinciding with the previous iteration. Meanwhile, the rest half will coincide with the upcoming iteration. Afterwards, the former half mini-batch distills on-the-fly soft targets generated in the previous iteration. Our proposed mechanism guides the training stability and consistency, resulting in robustness to label noise. Moreover, our method is easy to implement, without taking up extra run-time memory or requiring model structure modification. Experimental results on three classification benchmarks illustrate that our approach can consistently outperform state-of-the-art self-distillation approaches with different network architectures. Additionally, our method shows strong compatibility with augmentation strategies by gaining additional performance improvement. The code is available at https://github.com/Meta-knowledge-Lab/DLB.

Yufei Wang, Yiqing Shen, Meng Yuan et al.

Optical Coherence Tomography Angiography (OCTA) is a non-invasive and non-contacting imaging technique providing visualization of microvasculature of retina and optic nerve head in human eyes in vivo. The adequate image quality of OCTA is the prerequisite for the subsequent quantification of retinal microvasculature. Traditionally, the image quality score based on signal strength is used for discriminating low quality. However, it is insufficient for identifying artefacts such as motion and off-centration, which rely specialized knowledge and need tedious and time-consuming manual identification. One of the most primary issues in OCTA analysis is to sort out the foveal avascular zone (FAZ) region in the retina, which highly correlates with any visual acuity disease. However, the variations in OCTA visual quality affect the performance of deep learning in any downstream marginally. Moreover, filtering the low-quality OCTA images out is both labor-intensive and time-consuming. To address these issues, we develop an automated computer-aided OCTA image processing system using deep neural networks as the classifier and segmentor to help ophthalmologists in clinical diagnosis and research. This system can be an assistive tool as it can process OCTA images of different formats to assess the quality and segment the FAZ area. The source code is freely available at https://github.com/shanzha09/COIPS.git. Another major contribution is the large-scale OCTA dataset, namely OCTA-25K-IQA-SEG we publicize for performance evaluation. It is comprised of four subsets, namely sOCTA-3$\times$3-10k, sOCTA-6$\times$6-14k, sOCTA-3$\times$3-1.1k-seg, and dOCTA-6$\times$6-1.1k-seg, which contains a total number of 25,665 images. The large-scale OCTA dataset is available at https://doi.org/10.5281/zenodo.5111975, https://doi.org/10.5281/zenodo.5111972.

Congying Liu, Taihao Li, Ming Huang et al.

Protein analysis tasks arising in healthcare settings often require accurate reasoning under protein sequence constraints, involving tasks such as functional interpretation of disease-related variants, protein-level analysis for clinical research, and similar scenarios. To address such tasks, search agents are introduced to search protein-related information, providing support for disease-related variant analysis and protein function reasoning in protein-centric inference. However, such search agents are mostly limited to single-round, text-only modality search, which prevents the protein sequence modality from being incorporated as a multimodal input into the search decision-making process. Meanwhile, their reliance on reinforcement learning (RL) supervision that focuses solely on the final answer results in a lack of search process constraints, making deviations in keyword selection and reasoning directions difficult to identify and correct in a timely manner. To address these limitations, we propose ProtRLSearch, a multi-round protein search agent trained with multi-dimensional reward based RL, which jointly leverages protein sequence and text as multimodal inputs during real-time search to produce high quality reports. To evaluate the ability of models to integrate protein sequence information and text-based multimodal inputs in realistic protein query settings, we construct ProtMCQs, a benchmark of 3,000 multiple choice questions (MCQs) organized into three difficulty levels. The benchmark evaluates protein query tasks that range from sequence constrained reasoning about protein function and phenotype changes to comprehensive protein reasoning that integrates multi-dimensional sequence features with signal pathways and regulatory networks.

Haoxiang Gao, Zhongruo Wang, Yaqian Li et al.

The advent of foundation models has revolutionized the fields of natural language processing and computer vision, paving the way for their application in autonomous driving (AD). This survey presents a comprehensive review of more than 40 research papers, demonstrating the role of foundation models in enhancing AD. Large language models contribute to planning and simulation in AD, particularly through their proficiency in reasoning, code generation and translation. In parallel, vision foundation models are increasingly adapted for critical tasks such as 3D object detection and tracking, as well as creating realistic driving scenarios for simulation and testing. Multi-modal foundation models, integrating diverse inputs, exhibit exceptional visual understanding and spatial reasoning, crucial for end-to-end AD. This survey not only provides a structured taxonomy, categorizing foundation models based on their modalities and functionalities within the AD domain but also delves into the methods employed in current research. It identifies the gaps between existing foundation models and cutting-edge AD approaches, thereby charting future research directions and proposing a roadmap for bridging these gaps.

Miao Zhang, Jun Yin, Pengyu Zeng et al.

Deep learning-based image enhancement methods show significant advantages in reducing noise and improving visibility in low-light conditions. These methods are typically based on one-to-one mapping, where the model learns a direct transformation from low light to specific enhanced images. Therefore, these methods are inflexible as they do not allow highly personalized mapping, even though an individual's lighting preferences are inherently personalized. To overcome these limitations, we propose a new light enhancement task and a new framework that provides customized lighting control through prompt-driven, semantic-level, and quantitative brightness adjustments. The framework begins by leveraging a Large Language Model (LLM) to understand natural language prompts, enabling it to identify target objects for brightness adjustments. To localize these target objects, the Retinex-based Reasoning Segment (RRS) module generates precise target localization masks using reflection images. Subsequently, the Text-based Brightness Controllable (TBC) module adjusts brightness levels based on the generated illumination map. Finally, an Adaptive Contextual Compensation (ACC) module integrates multi-modal inputs and controls a conditional diffusion model to adjust the lighting, ensuring seamless and precise enhancements accurately. Experimental results on benchmark datasets demonstrate our framework's superior performance at increasing visibility, maintaining natural color balance, and amplifying fine details without creating artifacts. Furthermore, its robust generalization capabilities enable complex semantic-level lighting adjustments in diverse open-world environments through natural language interactions.

Yiqing Shen, Bohan Liu, Chenjia Li et al.

Reasoning segmentation (RS) aims to identify and segment objects of interest based on implicit text queries. As such, RS is a catalyst for embodied AI agents, enabling them to interpret high-level commands without requiring explicit step-by-step guidance. However, current RS approaches rely heavily on the visual perception capabilities of multimodal large language models (LLMs), leading to several major limitations. First, they struggle with queries that require multiple steps of reasoning or those that involve complex spatial/temporal relationships. Second, they necessitate LLM fine-tuning, which may require frequent updates to maintain compatibility with contemporary LLMs and may increase risks of catastrophic forgetting during fine-tuning. Finally, being primarily designed for static images or offline video processing, they scale poorly to online video data. To address these limitations, we propose an agent framework that disentangles perception and reasoning for online video RS without LLM fine-tuning. Our innovation is the introduction of a just-in-time digital twin concept, where -- given an implicit query -- a LLM plans the construction of a low-level scene representation from high-level video using specialist vision models. We refer to this approach to creating a digital twin as "just-in-time" because the LLM planner will anticipate the need for specific information and only request this limited subset instead of always evaluating every specialist model. The LLM then performs reasoning on this digital twin representation to identify target objects. To evaluate our approach, we introduce a new comprehensive video reasoning segmentation benchmark comprising 200 videos with 895 implicit text queries. The benchmark spans three reasoning categories (semantic, spatial, and temporal) with three different reasoning chain complexity.

Yiqing Shen, Chenjia Li, Fei Xiong et al.

Reasoning Segmentation (RS) aims to delineate objects based on implicit text queries, the interpretation of which requires reasoning and knowledge integration. Unlike the traditional formulation of segmentation problems that relies on fixed semantic categories or explicit prompting, RS bridges the gap between visual perception and human-like reasoning capabilities, facilitating more intuitive human-AI interaction through natural language. Our work presents the first comprehensive survey of RS for image and video processing, examining 26 state-of-the-art methods together with a review of the corresponding evaluation metrics, as well as 29 datasets and benchmarks. We also explore existing applications of RS across diverse domains and identify their potential extensions. Finally, we identify current research gaps and highlight promising future directions.

Yiqing Shen, Chenjia Li, Chenxiao Fan et al.

Visual reasoning, the capability to interpret visual input in response to implicit text query through multi-step reasoning, remains a challenge for deep learning models due to the lack of relevant benchmarks. Previous work in visual reasoning has primarily focused on reasoning segmentation, where models aim to segment objects based on implicit text queries. This paper introduces reasoning visual tasks (RVTs), a unified formulation that extends beyond traditional video reasoning segmentation to a diverse family of visual language reasoning problems, which can therefore accommodate multiple output formats including bounding boxes, natural language descriptions, and question-answer pairs. Correspondingly, we identify the limitations in current benchmark construction methods that rely solely on large language models (LLMs), which inadequately capture complex spatial-temporal relationships and multi-step reasoning chains in video due to their reliance on token representation, resulting in benchmarks with artificially limited reasoning complexity. To address this limitation, we propose a novel automated RVT benchmark construction pipeline that leverages digital twin (DT) representations as structured intermediaries between perception and the generation of implicit text queries. Based on this method, we construct RVTBench, a RVT benchmark containing 3,896 queries of over 1.2 million tokens across four types of RVT (segmentation, grounding, VQA and summary), three reasoning categories (semantic, spatial, and temporal), and four increasing difficulty levels, derived from 200 video sequences. Finally, we propose RVTagent, an agent framework for RVT that allows for zero-shot generalization across various types of RVT without task-specific fine-tuning.

Yiqing Shen, Hao Ding, Lalithkumar Seenivasan et al.

Current foundation models (FMs) rely on token representations that directly fragment continuous real-world multimodal data into discrete tokens. They limit FMs to learning real-world knowledge and relationships purely through statistical correlation rather than leveraging explicit domain knowledge. Consequently, current FMs struggle with maintaining semantic coherence across modalities, capturing fine-grained spatial-temporal dynamics, and performing causal reasoning. These limitations cannot be overcome by simply scaling up model size or expanding datasets. This position paper argues that the machine learning community should consider digital twin (DT) representations, which are outcome-driven digital representations that serve as building blocks for creating virtual replicas of physical processes, as an alternative to the token representation for building FMs. Finally, we discuss how DT representations can address these challenges by providing physically grounded representations that explicitly encode domain knowledge and preserve the continuous nature of real-world processes.

Xu Liu, Tong Zhou, Yuanxin Wang et al.

The advent of foundation models, which are pre-trained on vast datasets, has ushered in a new era of computer vision, characterized by their robustness and remarkable zero-shot generalization capabilities. Mirroring the transformative impact of foundation models like large language models (LLMs) in natural language processing, visual foundation models (VFMs) have become a catalyst for groundbreaking developments in computer vision. This review paper delineates the pivotal trajectories of VFMs, emphasizing their scalability and proficiency in generative tasks such as text-to-image synthesis, as well as their adeptness in discriminative tasks including image segmentation. While generative and discriminative models have historically charted distinct paths, we undertake a comprehensive examination of the recent strides made by VFMs in both domains, elucidating their origins, seminal breakthroughs, and pivotal methodologies. Additionally, we collate and discuss the extensive resources that facilitate the development of VFMs and address the challenges that pave the way for future research endeavors. A crucial direction for forthcoming innovation is the amalgamation of generative and discriminative paradigms. The nascent application of generative models within discriminative contexts signifies the early stages of this confluence. This survey aspires to be a contemporary compendium for scholars and practitioners alike, charting the course of VFMs and illuminating their multifaceted landscape.

Ming Hu, Chenglong Ma, Wei Li et al. · pku

Scientific Large Language Models (Sci-LLMs) are transforming how knowledge is represented, integrated, and applied in scientific research, yet their progress is shaped by the complex nature of scientific data. This survey presents a comprehensive, data-centric synthesis that reframes the development of Sci-LLMs as a co-evolution between models and their underlying data substrate. We formulate a unified taxonomy of scientific data and a hierarchical model of scientific knowledge, emphasizing the multimodal, cross-scale, and domain-specific challenges that differentiate scientific corpora from general natural language processing datasets. We systematically review recent Sci-LLMs, from general-purpose foundations to specialized models across diverse scientific disciplines, alongside an extensive analysis of over 270 pre-/post-training datasets, showing why Sci-LLMs pose distinct demands -- heterogeneous, multi-scale, uncertainty-laden corpora that require representations preserving domain invariance and enabling cross-modal reasoning. On evaluation, we examine over 190 benchmark datasets and trace a shift from static exams toward process- and discovery-oriented assessments with advanced evaluation protocols. These data-centric analyses highlight persistent issues in scientific data development and discuss emerging solutions involving semi-automated annotation pipelines and expert validation. Finally, we outline a paradigm shift toward closed-loop systems where autonomous agents based on Sci-LLMs actively experiment, validate, and contribute to a living, evolving knowledge base. Collectively, this work provides a roadmap for building trustworthy, continually evolving artificial intelligence (AI) systems that function as a true partner in accelerating scientific discovery.

Yiqing Shen, Outongyi Lv, Houying Zhu et al.

Large language models (LLMs) have garnered considerable attention for their proficiency in tackling intricate tasks, particularly leveraging their capacities for zero-shot and in-context learning. However, their utility has been predominantly restricted to general tasks due to an absence of domain-specific knowledge. This constraint becomes particularly pertinent in the realm of protein engineering, where specialized expertise is required for tasks such as protein function prediction, protein evolution analysis, and protein design, with a level of specialization that existing LLMs cannot furnish. In response to this challenge, we introduce \textsc{ProteinEngine}, a human-centered platform aimed at amplifying the capabilities of LLMs in protein engineering by seamlessly integrating a comprehensive range of relevant tools, packages, and software via API calls. Uniquely, \textsc{ProteinEngine} assigns three distinct roles to LLMs, facilitating efficient task delegation, specialized task resolution, and effective communication of results. This design fosters high extensibility and promotes the smooth incorporation of new algorithms, models, and features for future development. Extensive user studies, involving participants from both the AI and protein engineering communities across academia and industry, consistently validate the superiority of \textsc{ProteinEngine} in augmenting the reliability and precision of deep learning in protein engineering tasks. Consequently, our findings highlight the potential of \textsc{ProteinEngine} to bride the disconnected tools for future research in the protein engineering domain.

Meng-Hao Guo, Xuanyu Chu, Qianrui Yang et al.

The rapid advancement of native multi-modal models and omni-models, exemplified by GPT-4o, Gemini, and o3, with their capability to process and generate content across modalities such as text and images, marks a significant milestone in the evolution of intelligence. Systematic evaluation of their multi-modal output capabilities in visual thinking processes (also known as multi-modal chain of thought, M-CoT) becomes critically important. However, existing benchmarks for evaluating multi-modal models primarily focus on assessing multi-modal inputs and text-only reasoning while neglecting the importance of reasoning through multi-modal outputs. In this paper, we present a benchmark, dubbed RBench-V, designed to assess models' vision-indispensable reasoning abilities. To construct RBench-V, we carefully hand-pick 803 questions covering math, physics, counting, and games. Unlike previous benchmarks that typically specify certain input modalities, RBench-V presents problems centered on multi-modal outputs, which require image manipulation such as generating novel images and constructing auxiliary lines to support the reasoning process. We evaluate numerous open- and closed-source models on RBench-V, including o3, Gemini 2.5 Pro, Qwen2.5-VL, etc. Even the best-performing model, o3, achieves only 25.8% accuracy on RBench-V, far below the human score of 82.3%, highlighting that current models struggle to leverage multi-modal reasoning. Data and code are available at https://evalmodels.github.io/rbenchv

Jinzheng Li, Jingshu Zhang, Hongguang Li et al.

Financial decision-making requires processing vast amounts of real-time information while understanding their complex temporal relationships. While traditional search engines excel at providing real-time information access, they often struggle to comprehend sophisticated user intentions and contextual nuances. Conversely, Large Language Models (LLMs) demonstrate reasoning and interaction capabilities but may generate unreliable outputs without access to current data. While recent attempts have been made to combine LLMs with search capabilities, they suffer from (1) restricted access to specialized financial data, (2) static query structures that cannot adapt to dynamic market conditions, and (3) insufficient temporal awareness in result generation. To address these challenges, we present FinSearch, a novel agent-based search framework specifically designed for financial applications that interface with diverse financial data sources including market, stock, and news data. Innovatively, FinSearch comprises four components: (1) an LLM-based multi-step search pre-planner that decomposes user queries into structured sub-queries mapped to specific data sources through a graph representation; (2) a search executor with an LLM-based adaptive query rewriter that executes the searching of each sub-query while dynamically refining the sub-queries in its subsequent node based on intermediate search results; (3) a temporal weighting mechanism that prioritizes information relevance based on the deduced time context from the user's query; (4) an LLM-based response generator that synthesizes results into coherent, contextually appropriate outputs. To evaluate FinSearch, we construct FinSearchBench-24, a benchmark of 1,500 four-choice questions across the stock market, rate changes, monetary policy, and industry developments spanning from June to October 2024.

Yiqing Shen, Chenjia Li, Chenxiao Fan et al.

Conventional approaches to video segmentation are confined to predefined object categories and cannot identify out-of-vocabulary objects, let alone objects that are not identified explicitly but only referred to implicitly in complex text queries. This shortcoming limits the utility for video segmentation in complex and variable scenarios, where a closed set of object categories is difficult to define and where users may not know the exact object category that will appear in the video. Such scenarios can arise in operating room video analysis, where different health systems may use different workflows and instrumentation, requiring flexible solutions for video analysis. Reasoning segmentation (RS) now offers promise towards such a solution, enabling natural language text queries as interaction for identifying object to segment. However, existing video RS formulation assume that target objects remain contextually relevant throughout entire video sequences. This assumption is inadequate for real-world scenarios in which objects of interest appear, disappear or change relevance dynamically based on temporal context, such as surgical instruments that become relevant only during specific procedural phases or anatomical structures that gain importance at particular moments during surgery. Our first contribution is the introduction of temporally-constrained video reasoning segmentation, a novel task formulation that requires models to implicitly infer when target objects become contextually relevant based on text queries that incorporate temporal reasoning. Since manual annotation of temporally-constrained video RS datasets would be expensive and limit scalability, our second contribution is an innovative automated benchmark construction method. Finally, we present TCVideoRSBenchmark, a temporally-constrained video RS dataset containing 52 samples using the videos from the MVOR dataset.

Yunhao Liu, Suyang Xi, Shiqi Liu et al.

Medical image analysis increasingly relies on the integration of multiple imaging modalities to capture complementary anatomical and functional information, enabling more accurate diagnosis and treatment planning. Achieving aligned feature representations across these diverse modalities is therefore important for effective multimodal analysis. While contrastive language-image pre-training (CLIP) and its variant have enabled image-text alignments, they require explicitly paired data between arbitrary two modalities, which is difficult to acquire in medical contexts. To address the gap, we present Multimodal Medical Image Binding with Text (M\textsuperscript{3}Bind), a novel pre-training framework that enables seamless alignment of multiple medical imaging modalities through a shared text representation space without requiring explicit paired data between any two medical image modalities. Specifically, based on the insight that different images can naturally bind with text, M\textsuperscript{3}Bind first fine-tunes pre-trained CLIP-like image-text models to align their modality-specific text embedding space while preserving their original image-text alignments. Subsequently, we distill these modality-specific text encoders into a unified model, creating a shared text embedding space. Experiments on X-ray, CT, retina, ECG, and pathological images on multiple downstream tasks demonstrate that M\textsuperscript{3}Bind achieves state-of-the-art performance in zero-shot, few-shot classification and cross-modal retrieval tasks compared to its CLIP-like counterparts. These results validate M\textsuperscript{3}Bind's effectiveness in achieving cross-image-modal alignment for medical analysis.

Shijie Han, Jingshu Zhang, Yiqing Shen et al.

Current financial large language models (FinLLMs) struggle with two critical limitations: the absence of objective evaluation metrics to assess the quality of stock analysis reports and a lack of depth in stock analysis, which impedes their ability to generate professional-grade insights. To address these challenges, this paper introduces FinSphere, a stock analysis agent, along with three major contributions: (1) AnalyScore, a systematic evaluation framework for assessing stock analysis quality, (2) Stocksis, a dataset curated by industry experts to enhance LLMs' stock analysis capabilities, and (3) FinSphere, an AI agent that can generate high-quality stock analysis reports in response to user queries. Experiments demonstrate that FinSphere achieves superior performance compared to both general and domain-specific LLMs, as well as existing agent-based systems, even when they are enhanced with real-time data access and few-shot guidance. The integrated framework, which combines real-time data feeds, quantitative tools, and an instruction-tuned LLM, yields substantial improvements in both analytical quality and practical applicability for real-world stock analysis.

Blanca Inigo, Yiqing Shen, Benjamin D. Killeen et al.

Vertebral compression fractures (VCFs) are a common and potentially serious consequence of osteoporosis. Yet, they often remain undiagnosed. Opportunistic screening, which involves automated analysis of medical imaging data acquired primarily for other purposes, is a cost-effective method to identify undiagnosed VCFs. In high-stakes scenarios like opportunistic medical diagnosis, model interpretability is a key factor for the adoption of AI recommendations. Rule-based methods are inherently explainable and closely align with clinical guidelines, but they are not immediately applicable to high-dimensional data such as CT scans. To address this gap, we introduce a neurosymbolic approach for VCF detection in CT volumes. The proposed model combines deep learning (DL) for vertebral segmentation with a shape-based algorithm (SBA) that analyzes vertebral height distributions in salient anatomical regions. This allows for the definition of a rule set over the height distributions to detect VCFs. Evaluation of VerSe19 dataset shows that our method achieves an accuracy of 96% and a sensitivity of 91% in VCF detection. In comparison, a black box model, DenseNet, achieved an accuracy of 95% and sensitivity of 91% in the same dataset. Our results demonstrate that our intrinsically explainable approach can match or surpass the performance of black box deep neural networks while providing additional insights into why a prediction was made. This transparency can enhance clinician's trust thus, supporting more informed decision-making in VCF diagnosis and treatment planning.

Yiqing Shen, Chenjia Li, Mathias Unberath

Text-driven video editing enables users to modify video content only using text queries. While existing methods can modify video content if explicit descriptions of editing targets with precise spatial locations and temporal boundaries are provided, these requirements become impractical when users attempt to conceptualize edits through implicit queries referencing semantic properties or object relationships. We introduce reasoning video editing, a task where video editing models must interpret implicit queries through multi-hop reasoning to infer editing targets before executing modifications, and a first model attempting to solve this complex task, RIVER (Reasoning-based Implicit Video Editor). RIVER decouples reasoning from generation through digital twin representations of video content that preserve spatial relationships, temporal trajectories, and semantic attributes. A large language model then processes this representation jointly with the implicit query, performing multi-hop reasoning to determine modifications, then outputs structured instructions that guide a diffusion-based editor to execute pixel-level changes. RIVER training uses reinforcement learning with rewards that evaluate reasoning accuracy and generation quality. Finally, we introduce RVEBenchmark, a benchmark of 100 videos with 519 implicit queries spanning three levels and categories of reasoning complexity specifically for reasoning video editing. RIVER demonstrates best performance on the proposed RVEBenchmark and also achieves state-of-the-art performance on two additional video editing benchmarks (VegGIE and FiVE), where it surpasses six baseline methods.

Yiqing Shen, Aiza Maksutova, Chenjia Li et al.

World models learn to predict the temporal evolution of visual observations given a control signal, potentially enabling agents to reason about environments through forward simulation. Because of the focus on forward simulation, current world models generate predictions based on factual observations. For many emerging applications, such as comprehensive evaluations of physical AI behavior under varying conditions, the ability of world models to answer counterfactual queries, such as "what would happen if this object was removed?", is of increasing importance. We formalize counterfactual world models that additionally take interventions as explicit inputs, predicting temporal sequences under hypothetical modifications to observed scene properties. Traditional world models operate directly on entangled pixel-space representations where object properties and relationships cannot be selectively modified. This modeling choice prevents targeted interventions on specific scene properties. We introduce CWMDT, a framework to overcome those limitations, turning standard video diffusion models into effective counterfactual world models. First, CWMDT constructs digital twins of observed scenes to explicitly encode objects and their relationships, represented as structured text. Second, CWMDT applies large language models to reason over these representations and predict how a counterfactual intervention propagates through time to alter the observed scene. Third, CWMDT conditions a video diffusion model with the modified representation to generate counterfactual visual sequences. Evaluations on two benchmarks show that the CWMDT approach achieves state-of-the-art performance, suggesting that alternative representations of videos, such as the digital twins considered here, offer powerful control signals for video forward simulation-based world models.