Hongbin Liu

Aaron Blakeman, Aarti Basant, Abhinav Khattar et al. · nvidia

As inference-time scaling becomes critical for enhanced reasoning capabilities, it is increasingly becoming important to build models that are efficient to infer. We introduce Nemotron-H, a family of 8B and 56B/47B hybrid Mamba-Transformer models designed to reduce inference cost for a given accuracy level. To achieve this goal, we replace the majority of self-attention layers in the common Transformer model architecture with Mamba layers that perform constant computation and require constant memory per generated token. We show that Nemotron-H models offer either better or on-par accuracy compared to other similarly-sized state-of-the-art open-sourced Transformer models (e.g., Qwen-2.5-7B/72B and Llama-3.1-8B/70B), while being up to 3$\times$ faster at inference. To further increase inference speed and reduce the memory required at inference time, we created Nemotron-H-47B-Base from the 56B model using a new compression via pruning and distillation technique called MiniPuzzle. Nemotron-H-47B-Base achieves similar accuracy to the 56B model, but is 20% faster to infer. In addition, we introduce an FP8-based training recipe and show that it can achieve on par results with BF16-based training. This recipe is used to train the 56B model. We are releasing Nemotron-H base model checkpoints with support in Hugging Face and NeMo.

Zhen Chen, Zongming Zhang, Wenwu Guo et al.

Surgical instrument segmentation is crucial in surgical scene understanding, thereby facilitating surgical safety. Existing algorithms directly detected all instruments of pre-defined categories in the input image, lacking the capability to segment specific instruments according to the surgeon's intention. During different stages of surgery, surgeons exhibit varying preferences and focus toward different surgical instruments. Therefore, an instrument segmentation algorithm that adheres to the surgeon's intention can minimize distractions from irrelevant instruments and assist surgeons to a great extent. The recent Segment Anything Model (SAM) reveals the capability to segment objects following prompts, but the manual annotations for prompts are impractical during the surgery. To address these limitations in operating rooms, we propose an audio-driven surgical instrument segmentation framework, named ASI-Seg, to accurately segment the required surgical instruments by parsing the audio commands of surgeons. Specifically, we propose an intention-oriented multimodal fusion to interpret the segmentation intention from audio commands and retrieve relevant instrument details to facilitate segmentation. Moreover, to guide our ASI-Seg segment of the required surgical instruments, we devise a contrastive learning prompt encoder to effectively distinguish the required instruments from the irrelevant ones. Therefore, our ASI-Seg promotes the workflow in the operating rooms, thereby providing targeted support and reducing the cognitive load on surgeons. Extensive experiments are performed to validate the ASI-Seg framework, which reveals remarkable advantages over classical state-of-the-art and medical SAMs in both semantic segmentation and intention-oriented segmentation. The source code is available at https://github.com/Zonmgin-Zhang/ASI-Seg.

Lihua Zhou, Mao Ye, Xiatian Zhu et al.

Open-vocabulary object detection with vision-language models (VLMs) such as Grounding DINO suffers from performance degradation under test-time distribution shifts, primarily due to semantic misalignment between text embeddings and shifted visual embeddings of region proposals. While recent test-time adaptive object detection methods for VLM-based either rely on costly backpropagation or bypass semantic misalignment via external memory, none directly and efficiently align text and vision in a training-free manner. To address this, we propose Reward-Guided Semantic Evolution (RGSE), a training-free framework that directly refines the text embeddings at test time. Inspired by evolutionary search, RGSE treats text embedding adaptation as a semantic search process: it perturbs text embeddings as candidate variants, evaluates them via cosine similarity with current and historical high-confidence visual proposals as a reward signal, and fuses them into a refined embedding through reward-weighted averaging. Without any backpropagation, RGSE achieves state-of-the-art performance across multiple detection benchmarks while adding minimal computational overhead. Our code will be open source upon publication.

Hongbin Liu, Jinyuan Jia, Neil Zhenqiang Gong

Contrastive learning pre-trains an image encoder using a large amount of unlabeled data such that the image encoder can be used as a general-purpose feature extractor for various downstream tasks. In this work, we propose PoisonedEncoder, a data poisoning attack to contrastive learning. In particular, an attacker injects carefully crafted poisoning inputs into the unlabeled pre-training data, such that the downstream classifiers built based on the poisoned encoder for multiple target downstream tasks simultaneously classify attacker-chosen, arbitrary clean inputs as attacker-chosen, arbitrary classes. We formulate our data poisoning attack as a bilevel optimization problem, whose solution is the set of poisoning inputs; and we propose a contrastive-learning-tailored method to approximately solve it. Our evaluation on multiple datasets shows that PoisonedEncoder achieves high attack success rates while maintaining the testing accuracy of the downstream classifiers built upon the poisoned encoder for non-attacker-chosen inputs. We also evaluate five defenses against PoisonedEncoder, including one pre-processing, three in-processing, and one post-processing defenses. Our results show that these defenses can decrease the attack success rate of PoisonedEncoder, but they also sacrifice the utility of the encoder or require a large clean pre-training dataset.

Zhen Chen, Xingjian Luo, Jinlin Wu et al.

Surgical phase recognition is critical for assisting surgeons in understanding surgical videos. Existing studies focused more on online surgical phase recognition, by leveraging preceding frames to predict the current frame. Despite great progress, they formulated the task as a series of frame-wise classification, which resulted in a lack of global context of the entire procedure and incoherent predictions. Moreover, besides online analysis, accurate offline surgical phase recognition is also in significant clinical need for retrospective analysis, and existing online algorithms do not fully analyze the entire video, thereby limiting accuracy in offline analysis. To overcome these challenges and enhance both online and offline inference capabilities, we propose a universal Surgical Phase Localization Network, named SurgPLAN++, with the principle of temporal detection. To ensure a global understanding of the surgical procedure, we devise a phase localization strategy for SurgPLAN++ to predict phase segments across the entire video through phase proposals. For online analysis, to generate high-quality phase proposals, SurgPLAN++ incorporates a data augmentation strategy to extend the streaming video into a pseudo-complete video through mirroring, center-duplication, and down-sampling. For offline analysis, SurgPLAN++ capitalizes on its global phase prediction framework to continuously refine preceding predictions during each online inference step, thereby significantly improving the accuracy of phase recognition. We perform extensive experiments to validate the effectiveness, and our SurgPLAN++ achieves remarkable performance in both online and offline modes, which outperforms state-of-the-art methods. The source code is available at https://github.com/franciszchen/SurgPLAN-Plus.

Jinghuai Zhang, Hongbin Liu, Jinyuan Jia et al.

Contrastive learning (CL) pre-trains general-purpose encoders using an unlabeled pre-training dataset, which consists of images or image-text pairs. CL is vulnerable to data poisoning based backdoor attacks (DPBAs), in which an attacker injects poisoned inputs into the pre-training dataset so the encoder is backdoored. However, existing DPBAs achieve limited effectiveness. In this work, we take the first step to analyze the limitations of existing backdoor attacks and propose new DPBAs called CorruptEncoder to CL. CorruptEncoder introduces a new attack strategy to create poisoned inputs and uses a theory-guided method to maximize attack effectiveness. Our experiments show that CorruptEncoder substantially outperforms existing DPBAs. In particular, CorruptEncoder is the first DPBA that achieves more than 90% attack success rates with only a few (3) reference images and a small poisoning ratio 0.5%. Moreover, we also propose a defense, called localized cropping, to defend against DPBAs. Our results show that our defense can reduce the effectiveness of DPBAs, but it sacrifices the utility of the encoder, highlighting the need for new defenses.

Qingyao Tian, Zhen Chen, Huai Liao et al.

Single-image depth estimation is essential for endoscopy tasks such as localization, reconstruction, and augmented reality. Most existing methods in surgical scenes focus on in-domain depth estimation, limiting their real-world applicability. This constraint stems from the scarcity and inferior labeling quality of medical data for training. In this work, we present EndoOmni, the first foundation model for zero-shot cross-domain depth estimation for endoscopy. To harness the potential of diverse training data, we refine the advanced self-learning paradigm that employs a teacher model to generate pseudo-labels, guiding a student model trained on large-scale labeled and unlabeled data. To address training disturbance caused by inherent noise in depth labels, we propose a robust training framework that leverages both depth labels and estimated confidence from the teacher model to jointly guide the student model training. Moreover, we propose a weighted scale-and-shift invariant loss to adaptively adjust learning weights based on label confidence, thus imposing learning bias towards cleaner label pixels while reducing the influence of highly noisy pixels. Experiments on zero-shot relative depth estimation show that our EndoOmni improves state-of-the-art methods in medical imaging for 33\% and existing foundation models for 34\% in terms of absolute relative error on specific datasets. Furthermore, our model provides strong initialization for fine-tuning metric depth estimation, maintaining superior performance in both in-domain and out-of-domain scenarios. The source code is publicly available at https://github.com/TianCuteQY/EndoOmni.

Xinlei He, Hongbin Liu, Neil Zhenqiang Gong et al.

Semi-supervised learning (SSL) leverages both labeled and unlabeled data to train machine learning (ML) models. State-of-the-art SSL methods can achieve comparable performance to supervised learning by leveraging much fewer labeled data. However, most existing works focus on improving the performance of SSL. In this work, we take a different angle by studying the training data privacy of SSL. Specifically, we propose the first data augmentation-based membership inference attacks against ML models trained by SSL. Given a data sample and the black-box access to a model, the goal of membership inference attack is to determine whether the data sample belongs to the training dataset of the model. Our evaluation shows that the proposed attack can consistently outperform existing membership inference attacks and achieves the best performance against the model trained by SSL. Moreover, we uncover that the reason for membership leakage in SSL is different from the commonly believed one in supervised learning, i.e., overfitting (the gap between training and testing accuracy). We observe that the SSL model is well generalized to the testing data (with almost 0 overfitting) but ''memorizes'' the training data by giving a more confident prediction regardless of its correctness. We also explore early stopping as a countermeasure to prevent membership inference attacks against SSL. The results show that early stopping can mitigate the membership inference attack, but with the cost of model's utility degradation.

Zedian Shao, Hongbin Liu, Yuepeng Hu et al.

Multimodal large language models (MLLMs) have become the cornerstone of today's generative AI ecosystem, sparking intense competition among tech giants and startups. In particular, an MLLM generates a text response given a prompt consisting of an image and a question. While state-of-the-art MLLMs use safety filters and alignment techniques to refuse unsafe prompts, in this work, we introduce MLLM-Refusal, the first method that induces refusals for safe prompts. In particular, our MLLM-Refusal optimizes a nearly-imperceptible refusal perturbation and adds it to an image, causing target MLLMs to likely refuse a safe prompt containing the perturbed image and a safe question. Specifically, we formulate MLLM-Refusal as a constrained optimization problem and propose an algorithm to solve it. Our method offers competitive advantages for MLLM model providers by potentially disrupting user experiences of competing MLLMs, since competing MLLM's users will receive unexpected refusals when they unwittingly use these perturbed images in their prompts. We evaluate MLLM-Refusal on four MLLMs across four datasets, demonstrating its effectiveness in causing competing MLLMs to refuse safe prompts while not affecting non-competing MLLMs. Furthermore, we explore three potential countermeasures-adding Gaussian noise, DiffPure, and adversarial training. Our results show that though they can mitigate MLLM-Refusal's effectiveness, they also sacrifice the accuracy and/or efficiency of the competing MLLM. The code is available at https://github.com/Sadcardation/MLLM-Refusal.

Jinghuai Zhang, Jinyuan Jia, Hongbin Liu et al.

Point cloud classification is an essential component in many security-critical applications such as autonomous driving and augmented reality. However, point cloud classifiers are vulnerable to adversarially perturbed point clouds. Existing certified defenses against adversarial point clouds suffer from a key limitation: their certified robustness guarantees are probabilistic, i.e., they produce an incorrect certified robustness guarantee with some probability. In this work, we propose a general framework, namely PointCert, that can transform an arbitrary point cloud classifier to be certifiably robust against adversarial point clouds with deterministic guarantees. PointCert certifiably predicts the same label for a point cloud when the number of arbitrarily added, deleted, and/or modified points is less than a threshold. Moreover, we propose multiple methods to optimize the certified robustness guarantees of PointCert in three application scenarios. We systematically evaluate PointCert on ModelNet and ScanObjectNN benchmark datasets. Our results show that PointCert substantially outperforms state-of-the-art certified defenses even though their robustness guarantees are probabilistic.

Hongbin Liu, Wenjie Qu, Jinyuan Jia et al.

Classifiers in supervised learning have various security and privacy issues, e.g., 1) data poisoning attacks, backdoor attacks, and adversarial examples on the security side as well as 2) inference attacks and the right to be forgotten for the training data on the privacy side. Various secure and privacy-preserving supervised learning algorithms with formal guarantees have been proposed to address these issues. However, they suffer from various limitations such as accuracy loss, small certified security guarantees, and/or inefficiency. Self-supervised learning is an emerging technique to pre-train encoders using unlabeled data. Given a pre-trained encoder as a feature extractor, supervised learning can train a simple yet accurate classifier using a small amount of labeled training data. In this work, we perform the first systematic, principled measurement study to understand whether and when a pre-trained encoder can address the limitations of secure or privacy-preserving supervised learning algorithms. Our key findings are that a pre-trained encoder substantially improves 1) both accuracy under no attacks and certified security guarantees against data poisoning and backdoor attacks of state-of-the-art secure learning algorithms (i.e., bagging and KNN), 2) certified security guarantees of randomized smoothing against adversarial examples without sacrificing its accuracy under no attacks, 3) accuracy of differentially private classifiers, and 4) accuracy and/or efficiency of exact machine unlearning.

Baoru Huang, Tuan Vo, Chayun Kongtongvattana et al.

Real-time visual feedback from catheterization analysis is crucial for enhancing surgical safety and efficiency during endovascular interventions. However, existing datasets are often limited to specific tasks, small scale, and lack the comprehensive annotations necessary for broader endovascular intervention understanding. To tackle these limitations, we introduce CathAction, a large-scale dataset for catheterization understanding. Our CathAction dataset encompasses approximately 500,000 annotated frames for catheterization action understanding and collision detection, and 25,000 ground truth masks for catheter and guidewire segmentation. For each task, we benchmark recent related works in the field. We further discuss the challenges of endovascular intentions compared to traditional computer vision tasks and point out open research questions. We hope that CathAction will facilitate the development of endovascular intervention understanding methods that can be applied to real-world applications. The dataset is available at https://airvlab.github.io/cathaction/.

Wenwu Guo, Jinlin Wu, Zhen Chen et al.

Vision-based surgical navigation has received increasing attention due to its non-invasive, cost-effective, and flexible advantages. In particular, a critical element of the vision-based navigation system is tracking surgical instruments. Compared with 2D instrument tracking methods, 3D instrument tracking has broader value in clinical practice, but is also more challenging due to weak texture, occlusion, and lack of Computer-Aided Design (CAD) models for 3D registration. To solve these challenges, we propose the SurgTrack, a two-stage 3D instrument tracking method for CAD-free and robust real-world applications. In the first registration stage, we incorporate an Instrument Signed Distance Field (SDF) modeling the 3D representation of instruments, achieving CAD-freed 3D registration. Due to this, we can obtain the location and orientation of instruments in the 3D space by matching the video stream with the registered SDF model. In the second tracking stage, we devise a posture graph optimization module, leveraging the historical tracking results of the posture memory pool to optimize the tracking results and improve the occlusion robustness. Furthermore, we collect the Instrument3D dataset to comprehensively evaluate the 3D tracking of surgical instruments. The extensive experiments validate the superiority and scalability of our SurgTrack, by outperforming the state-of-the-arts with a remarkable improvement. The code and dataset are available at https://github.com/wenwucode/SurgTrack.

Rongfeng Wei, Jinlin Wu, Xuexue Bai et al.

In minimally invasive surgery, surgical instrument localization is a crucial task for endoscopic videos, which enables various applications for improving surgical outcomes. However, annotating the instrument localization in endoscopic videos is tedious and labor-intensive. In contrast, obtaining the category information is easy and efficient in real-world applications. To fully utilize the category information and address the localization problem, we propose a weakly supervised localization framework named WS-YOLO for surgical instruments. By leveraging the instrument category information as the weak supervision, our WS-YOLO framework adopts an unsupervised multi-round training strategy for the localization capability training. We validate our WS-YOLO framework on the Endoscopic Vision Challenge 2023 dataset, which achieves remarkable performance in the weakly supervised surgical instrument localization. The source code is available at https://github.com/Breezewrf/WS-YOLO.

Ziyang Song, Zelin Zang, Zuyao Chen et al.

Multimodal Large Language Models (MLLMs) have achieved impressive progress in natural image reasoning, yet their potential in medical imaging remains underexplored, especially in clinical anatomical surgical images. Anatomy understanding tasks demand precise understanding and clinically coherent answers, which are difficult to achieve due to the complexity of medical data and the scarcity of high-quality expert annotations. These challenges limit the effectiveness of conventional Supervised Fine-Tuning (SFT) strategies. While recent work has demonstrated that Group Relative Policy Optimization (GRPO) can enhance reasoning in MLLMs without relying on large amounts of data, we find two weaknesses that hinder GRPO's reasoning performance in anatomy recognition: 1) knowledge cannot be effectively shared between different anatomical structures, resulting in uneven information gain and preventing the model from converging, and 2) the model quickly converges to a single reasoning path, suppressing the exploration of diverse strategies. To overcome these challenges, we propose two novel methods. First, we implement a progressive learning strategy called Anatomical Similarity Curriculum Learning by controlling question difficulty via the similarity of answer choices, enabling the model to master complex problems incrementally. Second, we utilize question augmentation referred to as Group Diversity Question Augmentation to expand the model's search space for difficult queries, mitigating the tendency to produce uniform responses. Comprehensive experiments on the SGG-VQA and OmniMedVQA benchmarks show our method achieves a significant improvement across the two benchmarks, demonstrating its effectiveness in enhancing the medical reasoning capabilities of MLLMs. The code can be found in https://github.com/tomato996/Anatomy-R1

Zhenrong Zhang, Jianan Liu, Yuxuan Xia et al.

Online multi-object tracking (MOT) plays a pivotal role in autonomous systems. The state-of-the-art approaches usually employ a tracking-by-detection method, and data association plays a critical role. This paper proposes a learning and graph-optimized (LEGO) modular tracker to improve data association performance in the existing literature. The proposed LEGO tracker integrates graph optimization and self-attention mechanisms, which efficiently formulate the association score map, facilitating the accurate and efficient matching of objects across time frames. To further enhance the state update process, the Kalman filter is added to ensure consistent tracking by incorporating temporal coherence in the object states. Our proposed method utilizing LiDAR alone has shown exceptional performance compared to other online tracking approaches, including LiDAR-based and LiDAR-camera fusion-based methods. LEGO ranked 1st at the time of submitting results to KITTI object tracking evaluation ranking board and remains 2nd at the time of submitting this paper, among all online trackers in the KITTI MOT benchmark for cars1

Xingjian Luo, You Pang, Zhen Chen et al.

Surgical phase recognition is crucial to providing surgery understanding in smart operating rooms. Despite great progress in automatic surgical phase recognition, most existing methods are still restricted by two problems. First, these methods cannot capture discriminative visual features for each frame and motion information with simple 2D networks. Second, the frame-by-frame recognition paradigm degrades the performance due to unstable predictions within each phase, termed as phase shaking. To address these two challenges, we propose a Surgical Phase LocAlization Network, named SurgPLAN, to facilitate a more accurate and stable surgical phase recognition with the principle of temporal detection. Specifically, we first devise a Pyramid SlowFast (PSF) architecture to serve as the visual backbone to capture multi-scale spatial and temporal features by two branches with different frame sampling rates. Moreover, we propose a Temporal Phase Localization (TPL) module to generate the phase prediction based on temporal region proposals, which ensures accurate and consistent predictions within each surgical phase. Extensive experiments confirm the significant advantages of our SurgPLAN over frame-by-frame approaches in terms of both accuracy and stability.

Qingyao Tian, Zhen Chen, Huai Liao et al.

Accurate bronchoscope localization is essential for pulmonary interventions, by providing six degrees of freedom (DOF) in airway navigation. However, the robustness of current vision-based methods is often compromised in clinical practice, and they struggle to perform in real-time and to generalize across cases unseen during training. To overcome these challenges, we propose a novel Probabilistic Airway Navigation System (PANS), leveraging Monte-Carlo method with pose hypotheses and likelihoods to achieve robust and real-time bronchoscope localization. Specifically, our PANS incorporates diverse visual representations (\textit{e.g.}, odometry and landmarks) by leveraging two key modules, including the Depth-based Motion Inference (DMI) and the Bronchial Semantic Analysis (BSA). To generate the pose hypotheses of bronchoscope for PANS, we devise the DMI to accurately propagate the estimation of pose hypotheses over time. Moreover, to estimate the accurate pose likelihood, we devise the BSA module by effectively distinguishing between similar bronchial regions in endoscopic images, along with a novel metric to assess the congruence between estimated depth maps and the segmented airway structure. Under this probabilistic formulation, our PANS is capable of achieving the 6-DOF bronchoscope localization with superior accuracy and robustness. Extensive experiments on the collected pulmonary intervention dataset comprising 10 clinical cases confirm the advantage of our PANS over state-of-the-arts, in terms of both robustness and generalization in localizing deeper airway branches and the efficiency of real-time inference. The proposed PANS reveals its potential to be a reliable tool in the operating room, promising to enhance the quality and safety of pulmonary interventions.

Wen Huang, Hongbin Liu, Minxin Guo et al.

Visual hallucination (VH) means that a multi-modal LLM (MLLM) imagines incorrect details about an image in visual question answering. Existing studies find VH instances only in existing image datasets, which results in biased understanding of MLLMs' performance under VH due to limited diversity of such VH instances. In this work, we propose a tool called VHTest to generate a diverse set of VH instances. Specifically, VHTest finds some initial VH instances in existing image datasets (e.g., COCO), generates a text description for each VH mode, and uses a text-to-image generative model (e.g., DALL-E-3) to generate VH images based on the text descriptions. We collect a benchmark dataset with 1,200 VH instances in 8 VH modes using VHTest. We find that existing MLLMs such as GPT-4V, LLaVA-1.5, and MiniGPT-v2 hallucinate for a large fraction of the instances in our benchmark. Moreover, we find that fine-tuning an MLLM using our benchmark dataset reduces its likelihood to hallucinate without sacrificing its performance on other benchmarks. Our benchmarks are publicly available: https://github.com/wenhuang2000/VHTest.

Simran Khanuja, Hongbin Liu, Shujian Zhang et al.

LLMs are deployed globally, yet produce responses biased towards cultures with abundant training data. Existing cultural localization approaches such as prompting or post-training alignment are black-box, hard to control, and do not reveal whether failures reflect missing knowledge or poor elicitation. In this paper, we address these gaps using mechanistic interpretability to uncover and manipulate cultural representations in LLMs. Leveraging sparse autoencoders, we identify interpretable features that encode culturally salient information and aggregate them into Cultural Embeddings (CuE). We use CuE both to analyze implicit cultural biases under underspecified prompts and to construct white-box steering interventions. Across multiple models, we show that CuE-based steering increases cultural faithfulness and elicits significantly rarer, long-tail cultural concepts than prompting alone. Notably, CuE-based steering is complementary to black-box localization methods, offering gains when applied on top of prompt-augmented inputs. This also suggests that models do benefit from better elicitation strategies, and don't necessarily lack long-tail knowledge representation, though this varies across cultures. Our results provide both diagnostic insight into cultural representations in LLMs and a controllable method to steer towards desired cultures.

Hongbin Liu, Zhengyuan Jiang, Cheng Hong et al.

A vision foundation model outputs an embedding vector for an image, which can be affected by common editing operations (e.g., JPEG compression, brightness, contrast adjustments). These common perturbations alter embedding vectors and may impact the performance of downstream tasks using these embeddings. In this work, we present the first systematic study on foundation models' robustness to such perturbations. We propose three robustness metrics and formulate five desired mathematical properties for these metrics, analyzing which properties they satisfy or violate. Using these metrics, we evaluate six industry-scale foundation models (OpenAI, Meta) across nine common perturbation categories, finding them generally non-robust. We also show that common perturbations degrade downstream application performance (e.g., classification accuracy) and that robustness values can predict performance impacts. Finally, we propose a fine-tuning approach to improve robustness without sacrificing utility.

Yuqi Jia, Minghong Fang, Hongbin Liu et al.

Poisoning attacks compromise the training phase of federated learning (FL) such that the learned global model misclassifies attacker-chosen inputs called target inputs. Existing defenses mainly focus on protecting the training phase of FL such that the learnt global model is poison free. However, these defenses often achieve limited effectiveness when the clients' local training data is highly non-iid or the number of malicious clients is large, as confirmed in our experiments. In this work, we propose FLForensics, the first poison-forensics method for FL. FLForensics complements existing training-phase defenses. In particular, when training-phase defenses fail and a poisoned global model is deployed, FLForensics aims to trace back the malicious clients that performed the poisoning attack after a misclassified target input is identified. We theoretically show that FLForensics can accurately distinguish between benign and malicious clients under a formal definition of poisoning attack. Moreover, we empirically show the effectiveness of FLForensics at tracing back both existing and adaptive poisoning attacks on five benchmark datasets.

Jinlin Wu, Felix Holm, Chuxi Chen et al.

While foundation models have advanced surgical video analysis, current approaches rely predominantly on pixel-level reconstruction objectives that waste model capacity on low-level visual details - such as smoke, specular reflections, and fluid motion - rather than semantic structures essential for surgical understanding. We present UniSurg, a video-native foundation model that shifts the learning paradigm from pixel-level reconstruction to latent motion prediction. Built on the Video Joint Embedding Predictive Architecture (V-JEPA), UniSurg introduces three key technical innovations tailored to surgical videos: 1) motion-guided latent prediction to prioritize semantically meaningful regions, 2) spatiotemporal affinity self-distillation to enforce relational consistency, and 3) feature diversity regularization to prevent representation collapse in texture-sparse surgical scenes. To enable large-scale pretraining, we curate UniSurg-15M, the largest surgical video dataset to date, comprising 3,658 hours of video from 50 sources across 13 anatomical regions. Extensive experiments across 17 benchmarks demonstrate that UniSurg significantly outperforms state-of-the-art methods on surgical workflow recognition (+14.6% F1 on EgoSurgery, +10.3% on PitVis), action triplet recognition (39.54% mAP-IVT on CholecT50), skill assessment, polyp segmentation, and depth estimation. These results establish UniSurg as a new standard for universal, motion-oriented surgical video understanding.

Zhen Sun, Huan Xu, Jinlin Wu et al.

In surgical procedures, correct instrument counting is essential. Instance segmentation is a location method that locates not only an object's bounding box but also each pixel's specific details. However, obtaining mask-level annotations is labor-intensive in instance segmentation. To address this issue, we propose a novel yet effective weakly-supervised surgical instrument instance segmentation approach, named Point-based Weakly-supervised Instance Segmentation (PWISeg). PWISeg adopts an FCN-based architecture with point-to-box and point-to-mask branches to model the relationships between feature points and bounding boxes, as well as feature points and segmentation masks on FPN, accomplishing instrument detection and segmentation jointly in a single model. Since mask level annotations are hard to available in the real world, for point-to-mask training, we introduce an unsupervised projection loss, utilizing the projected relation between predicted masks and bboxes as supervision signal. On the other hand, we annotate a few pixels as the key pixel for each instrument. Based on this, we further propose a key pixel association loss and a key pixel distribution loss, driving the point-to-mask branch to generate more accurate segmentation predictions. To comprehensively evaluate this task, we unveil a novel surgical instrument dataset with manual annotations, setting up a benchmark for further research. Our comprehensive research trial validated the superior performance of our PWISeg. The results show that the accuracy of surgical instrument segmentation is improved, surpassing most methods of instance segmentation via weakly supervised bounding boxes. This improvement is consistently observed in our proposed dataset and when applied to the public HOSPI-Tools dataset.

Ziyang Song, Zelin Zang, Xiaofan Ye et al.

Multimodal Large Language Models (MLLMs) have shown significant potential in surgical video understanding. With improved zero-shot performance and more effective human-machine interaction, they provide a strong foundation for advancing surgical education and assistance. However, existing research and datasets primarily focus on understanding surgical procedures and workflows, while paying limited attention to the critical role of anatomical comprehension. In clinical practice, surgeons rely heavily on precise anatomical understanding to interpret, review, and learn from surgical videos. To fill this gap, we introduce the Neurosurgical Anatomy Benchmark (NeuroABench), the first multimodal benchmark explicitly created to evaluate anatomical understanding in the neurosurgical domain. NeuroABench consists of 9 hours of annotated neurosurgical videos covering 89 distinct procedures and is developed using a novel multimodal annotation pipeline with multiple review cycles. The benchmark evaluates the identification of 68 clinical anatomical structures, providing a rigorous and standardized framework for assessing model performance. Experiments on over 10 state-of-the-art MLLMs reveal significant limitations, with the best-performing model achieving only 40.87% accuracy in anatomical identification tasks. To further evaluate the benchmark, we extract a subset of the dataset and conduct an informative test with four neurosurgical trainees. The results show that the best-performing student achieves 56% accuracy, with the lowest scores of 28% and an average score of 46.5%. While the best MLLM performs comparably to the lowest-scoring student, it still lags significantly behind the group's average performance. This comparison underscores both the progress of MLLMs in anatomical understanding and the substantial gap that remains in achieving human-level performance.

Zijie Yan, Hongxiao Bai, Xin Yao et al.

Scaling Mixture-of-Experts (MoE) training introduces systems challenges absent in dense models. Because each token activates only a subset of experts, this sparsity allows total parameters to grow much faster than per-token computation, creating coupled constraints across memory, communication, and computation. Optimizing one dimension often shifts pressure to another, demanding co-design across the full system stack. We address these challenges for MoE training through integrated optimizations spanning memory (fine-grained recomputation, offloading, etc.), communication (optimized dispatchers, overlapping, etc.), and computation (Grouped GEMM, fusions, CUDA Graphs, etc.). The framework also provides Parallel Folding for flexible multi-dimensional parallelism, low-precision training support for FP8 and NVFP4, and efficient long-context training. On NVIDIA GB300 and GB200, it achieves 1,233/1,048 TFLOPS/GPU for DeepSeek-V3-685B and 974/919 TFLOPS/GPU for Qwen3-235B. As a performant, scalable, and production-ready open-source solution, it has been used across academia and industry for training MoE models ranging from billions to trillions of parameters on clusters scaling up to thousands of GPUs. This report explains how these techniques work, their trade-offs, and their interactions at the systems level, providing practical guidance for scaling MoE models with Megatron Core.

Haoyan Gong, Yuzheng Feng, Zhenrong Zhang et al.

Vehicle license plate recognition is a crucial task in intelligent traffic management systems. However, the challenge of achieving accurate recognition persists due to motion blur from fast-moving vehicles. Despite the widespread use of image synthesis approaches in existing deblurring and recognition algorithms, their effectiveness in real-world scenarios remains unproven. To address this, we introduce the first large-scale license plate deblurring dataset named License Plate Blur (LPBlur), captured by a dual-camera system and processed through a post-processing pipeline to avoid misalignment issues. Then, we propose a License Plate Deblurring Generative Adversarial Network (LPDGAN) to tackle the license plate deblurring: 1) a Feature Fusion Module to integrate multi-scale latent codes; 2) a Text Reconstruction Module to restore structure through textual modality; 3) a Partition Discriminator Module to enhance the model's perception of details in each letter. Extensive experiments validate the reliability of the LPBlur dataset for both model training and testing, showcasing that our proposed model outperforms other state-of-the-art motion deblurring methods in realistic license plate deblurring scenarios. The dataset and code are available at https://github.com/haoyGONG/LPDGAN.

Yiming Huang, Beilei Cui, Long Bai et al.

Simultaneous Localization and Mapping (SLAM) is essential for precise surgical interventions and robotic tasks in minimally invasive procedures. While recent advancements in 3D Gaussian Splatting (3DGS) have improved SLAM with high-quality novel view synthesis and fast rendering, these systems struggle with accurate depth and surface reconstruction due to multi-view inconsistencies. Simply incorporating SLAM and 3DGS leads to mismatches between the reconstructed frames. In this work, we present Endo-2DTAM, a real-time endoscopic SLAM system with 2D Gaussian Splatting (2DGS) to address these challenges. Endo-2DTAM incorporates a surface normal-aware pipeline, which consists of tracking, mapping, and bundle adjustment modules for geometrically accurate reconstruction. Our robust tracking module combines point-to-point and point-to-plane distance metrics, while the mapping module utilizes normal consistency and depth distortion to enhance surface reconstruction quality. We also introduce a pose-consistent strategy for efficient and geometrically coherent keyframe sampling. Extensive experiments on public endoscopic datasets demonstrate that Endo-2DTAM achieves an RMSE of $1.87\pm 0.63$ mm for depth reconstruction of surgical scenes while maintaining computationally efficient tracking, high-quality visual appearance, and real-time rendering. Our code will be released at github.com/lastbasket/Endo-2DTAM.

Gheorghe Comanici, Eric Bieber, Mike Schaekermann et al. · amazon-science, baidu

In this report, we introduce the Gemini 2.X model family: Gemini 2.5 Pro and Gemini 2.5 Flash, as well as our earlier Gemini 2.0 Flash and Flash-Lite models. Gemini 2.5 Pro is our most capable model yet, achieving SoTA performance on frontier coding and reasoning benchmarks. In addition to its incredible coding and reasoning skills, Gemini 2.5 Pro is a thinking model that excels at multimodal understanding and it is now able to process up to 3 hours of video content. Its unique combination of long context, multimodal and reasoning capabilities can be combined to unlock new agentic workflows. Gemini 2.5 Flash provides excellent reasoning abilities at a fraction of the compute and latency requirements and Gemini 2.0 Flash and Flash-Lite provide high performance at low latency and cost. Taken together, the Gemini 2.X model generation spans the full Pareto frontier of model capability vs cost, allowing users to explore the boundaries of what is possible with complex agentic problem solving.

Zedian Shao, Hongbin Liu, Jaden Mu et al.

Prompt injection attack, where an attacker injects a prompt into the original one, aiming to make an Large Language Model (LLM) follow the injected prompt to perform an attacker-chosen task, represent a critical security threat. Existing attacks primarily focus on crafting these injections at inference time, treating the LLM itself as a static target. Our experiments show that these attacks achieve some success, but there is still significant room for improvement. In this work, we introduces a more foundational attack vector: poisoning the LLM's alignment process to amplify the success of future prompt injection attacks. Specifically, we propose PoisonedAlign, a method that strategically creates poisoned alignment samples to poison an LLM's alignment dataset. Our experiments across five LLMs and two alignment datasets show that when even a small fraction of the alignment data is poisoned, the resulting model becomes substantially more vulnerable to a wide range of prompt injection attacks. Crucially, this vulnerability is instilled while the LLM's performance on standard capability benchmarks remains largely unchanged, making the manipulation difficult to detect through automated, general-purpose performance evaluations. The code for implementing the attack is available at https://github.com/Sadcardation/PoisonedAlign.

Zedian Shao, Hongbin Liu, Yuepeng Hu et al.

Multi-modal large language models (MLLMs) have emerged as powerful tools for analyzing Internet-scale image data, offering significant benefits but also raising critical safety and societal concerns. In particular, open-weight MLLMs may be misused to extract sensitive information from personal images at scale, such as identities, locations, or other private details. In this work, we propose ImageProtector, a user-side method that proactively protects images before sharing by embedding a carefully crafted, nearly imperceptible perturbation that acts as a visual prompt injection attack on MLLMs. As a result, when an adversary analyzes a protected image with an MLLM, the MLLM is consistently induced to generate a refusal response such as "I'm sorry, I can't help with that request." We empirically demonstrate the effectiveness of ImageProtector across six MLLMs and four datasets. Additionally, we evaluate three potential countermeasures, Gaussian noise, DiffPure, and adversarial training, and show that while they partially mitigate the impact of ImageProtector, they simultaneously degrade model accuracy and/or efficiency. Our study focuses on the practically important setting of open-weight MLLMs and large-scale automated image analysis, and highlights both the promise and the limitations of perturbation-based privacy protection.

Yimeng Geng, Gaofeng Meng, Mingcong Chen et al.

Accurate identification of arteries and veins in ultrasound images is crucial for vascular examinations and interventions in robotics-assisted surgeries. However, current methods for ultrasound vessel segmentation face challenges in distinguishing between arteries and veins due to their morphological similarities. To address this challenge, this study introduces a novel force sensing guided segmentation approach to enhance artery-vein segmentation accuracy by leveraging their distinct deformability. Our proposed method utilizes force magnitude to identify key frames with the most significant vascular deformation in a sequence of ultrasound images. These key frames are then integrated with the current frame through attention mechanisms, with weights assigned in accordance with force magnitude. Our proposed force sensing guided framework can be seamlessly integrated into various segmentation networks and achieves significant performance improvements in multiple U-shaped networks such as U-Net, Swin-unet and Transunet. Furthermore, we contribute the first multimodal ultrasound artery-vein segmentation dataset, Mus-V, which encompasses both force and image data simultaneously. The dataset comprises 3114 ultrasound images of carotid and femoral vessels extracted from 105 videos, with corresponding force data recorded by the force sensor mounted on the US probe. Our code and dataset will be publicly available.

Xiaotian Ren, Nuerxiati Abudurexiti, Zhengyong Jiang et al.

Portfolio optimization in non-stationary markets is challenging due to regime shifts, dynamic correlations, and the limited interpretability of deep reinforcement learning (DRL) policies. We propose a Segmented Allocation with Momentum-Adjusted Utility for Multi-agent Portfolio Management via Hierarchical Deep Reinforcement Learning (SAMP-HDRL). The framework first applies dynamic asset grouping to partition the market into high-quality and ordinary subsets. An upper-level agent extracts global market signals, while lower-level agents perform intra-group allocation under mask constraints. A utility-based capital allocation mechanism integrates risky and risk-free assets, ensuring coherent coordination between global and local decisions. backtests across three market regimes (2019--2021) demonstrate that SAMP-HDRL consistently outperforms nine traditional baselines and nine DRL benchmarks under volatile and oscillating conditions. Compared with the strongest baseline, our method achieves at least 5\% higher Return, 5\% higher Sharpe ratio, 5\% higher Sortino ratio, and 2\% higher Omega ratio, with substantially larger gains observed in turbulent markets. Ablation studies confirm that upper--lower coordination, dynamic clustering, and capital allocation are indispensable to robustness. SHAP-based interpretability further reveals a complementary ``diversified + concentrated'' mechanism across agents, providing transparent insights into decision-making. Overall, SAMP-HDRL embeds structural market constraints directly into the DRL pipeline, offering improved adaptability, robustness, and interpretability in complex financial environments.

Qingyao Tian, Huai Liao, Xinyan Huang et al.

Endoscopic video-based tasks, such as visual navigation and surgical phase recognition, play a crucial role in minimally invasive surgeries by providing real-time assistance. While recent video foundation models have shown promise, their applications are hindered by (1) computational inefficiencies and (2) suboptimal performance caused by limited data for pre-training in endoscopy. To address these issues, we present EndoMamba, a foundation model designed for real-time inference while learning generalized spatiotemporal representations. First, to mitigate computational inefficiencies, we propose the EndoMamba backbone, optimized for real-time inference. Inspired by recent advancements in state space models, EndoMamba integrates Bidirectional Mamba blocks for spatial modeling within individual frames and vanilla Mamba blocks for past-to-present reasoning across the temporal domain. This design enables both strong spatiotemporal modeling and efficient inference in online video streams. Second, we propose a self-supervised hierarchical pre-training diagram to enhance EndoMamba's representation learning using endoscopic videos and incorporating general video domain knowledge. Specifically, our approach combines masked reconstruction with auxiliary supervision, leveraging low-level reconstruction to capture spatial-temporal structures and high-level alignment to transfer broader knowledge from a pretrained general-video domain foundation model. Extensive experiments on four downstream tasks--classification, segmentation, surgical phase recognition, and localization--demonstrate that EndoMamba outperforms existing foundation models and task-specific methods while maintaining real-time inference speed. The source code is available at https://github.com/TianCuteQY/EndoMamba.

Wenxi Wang, Hongbin Liu, Mingqian Li et al.

Text-to-image generation using diffusion models has achieved remarkable success. However, users often possess clear visual intents but struggle to express them precisely in language, resulting in ambiguous prompts and misaligned images. Existing methods struggle to bridge this gap, typically relying on high-load textual dialogues, opaque black-box inferences, or expensive fine-tuning. They fail to simultaneously achieve low cognitive load, interpretable preference inference, and remain training-free and model-agnostic. To address this, we propose RFD, an interactive framework that adapts the relevance feedback mechanism from information retrieval to diffusion models. In RFD, users replace explicit textual dialogue with implicit, multi-select visual feedback to minimize cognitive load, easily expressing complex, multi-dimensional preferences. To translate feedback into precise generative guidance, we construct an expert-curated feature repository and introduce an information-theoretic weighted cumulative preference analysis. This white-box method calculates preferences from current-round feedback and incrementally accumulates them, avoiding the concatenation of historical interactions and preventing inference degradation caused by lengthy contexts. Furthermore, RFD employs a probabilistic sampling mechanism for prompt reconstruction to balance exploitation and exploration, preventing output homogenization. Crucially, RFD operates entirely within the external text space, making it strictly training-free and model-agnostic as a universal plug-and-play solution. Extensive experiments demonstrate that RFD effectively captures the user's true visual intent, significantly outperforming baselines in preference alignment.

Zhen Chen, Xingjian Luo, Kun Yuan et al.

Surgical video understanding is crucial for facilitating Computer-Assisted Surgery (CAS) systems. Despite significant progress in existing studies, two major limitations persist, including inadequate visual content perception and insufficient temporal awareness in surgical videos, and hinder the development of versatile CAS solutions. In this work, we propose the SurgLLM framework, an effective large multimodal model tailored for versatile surgical video understanding tasks with enhanced spatial focus and temporal awareness. Specifically, to empower the spatial focus of surgical videos, we first devise Surgical Context-aware Multimodal Pretraining (Surg-Pretrain) for the video encoder of SurgLLM, by performing instrument-centric Masked Video Reconstruction (MV-Recon) and subsequent multimodal alignment. To incorporate surgical temporal knowledge into SurgLLM, we further propose Temporal-aware Multimodal Tuning (TM-Tuning) to enhance temporal reasoning with interleaved multimodal embeddings. Moreover, to accommodate various understanding tasks of surgical videos without conflicts, we devise a Surgical Task Dynamic Ensemble to efficiently triage a query with optimal learnable parameters in our SurgLLM. Extensive experiments performed on diverse surgical video understanding tasks, including captioning, general VQA, and temporal VQA, demonstrate significant improvements over the state-of-the-art approaches, validating the effectiveness of our SurgLLM in versatile surgical video understanding. The source code is available at https://github.com/franciszchen/SurgLLM.

Lumin Chen, Zhiying Wu, Tianye Lei et al.

Pituitary tumors often cause deformation or encapsulation of adjacent vital structures. Anatomical structure segmentation can provide surgeons with early warnings of regions that pose surgical risks, thereby enhancing the safety of pituitary surgery. However, pixel-level annotated video stream datasets for pituitary surgeries are extremely rare. To address this challenge, we introduce a new dataset for Pituitary Anatomy Segmentation (PAS). PAS comprises 7,845 time-coherent images extracted from 120 videos. To mitigate class imbalance, we apply data augmentation techniques that simulate the presence of surgical instruments in the training data. One major challenge in pituitary anatomy segmentation is the inconsistency in feature representation due to occlusions, camera motion, and surgical bleeding. By incorporating a Feature Fusion module, F2PASeg is proposed to refine anatomical structure segmentation by leveraging both high-resolution image features and deep semantic embeddings, enhancing robustness against intraoperative variations. Experimental results demonstrate that F2PASeg consistently segments critical anatomical structures in real time, providing a reliable solution for intraoperative pituitary surgery planning. Code: https://github.com/paulili08/F2PASeg.

Guankun Wang, Junyi Wang, Wenjin Mo et al.

Surgical scene understanding is critical for surgical training and robotic decision-making in robot-assisted surgery. Recent advances in Multimodal Large Language Models (MLLMs) have demonstrated great potential for advancing scene perception in the medical domain, facilitating surgeons to understand surgical scenes and procedures. However, these methods are primarily oriented towards image-based analysis or global video understanding, overlooking the fine-grained video reasoning that is crucial for analyzing specific processes and capturing detailed task execution within a surgical procedure. To bridge this gap, we propose SurgVidLM, the first video language model designed to address both full and fine-grained surgical video comprehension. To train our SurgVidLM, we construct the SVU-31K that is a large-scale dataset with over 31K video-instruction pairs, enabling both holistic understanding and detailed analysis of surgical procedures. Building on this resource, SurgVidLM incorporates a two-stage StageFocus mechanism: the first stage extracts global procedural context, while the second stage performs high-frequency local analysis guided by temporal cues. We also develop the Multi-frequency Fusion Attention to effectively integrate low- and high-frequency visual tokens, ensuring the preservation of critical task-specific details. Experimental results demonstrate that SurgVidLM significantly outperforms state-of-the-art Vid-LLMs of comparable parameter scale in both full and fine-grained video understanding tasks, showcasing its superior capability in capturing the context of complex robot-assisted surgeries. Our code and dataset will be publicly accessible soon.

Zhongye Liu, Hongbin Liu, Yuepeng Hu et al.

Visual hallucination (VH) occurs when a multimodal large language model (MLLM) generates responses with incorrect visual details for prompts. Existing methods for generating VH test cases primarily rely on human annotations, typically in the form of triples: (image, question, answer). In this paper, we introduce VHExpansion, the first automated method for expanding VH test cases for MLLMs. Given an initial VH test case, VHExpansion automatically expands it by perturbing the question and answer through negation as well as modifying the image using both common and adversarial perturbations. Additionally, we propose a new evaluation metric, symmetric accuracy, which measures the proportion of correctly answered VH test-case pairs. Each pair consists of a test case and its negated counterpart. Our theoretical analysis shows that symmetric accuracy is an unbiased evaluation metric that remains unaffected by the imbalance of VH testing cases with varying answers when an MLLM is randomly guessing the answers, whereas traditional accuracy is prone to such imbalance. We apply VHExpansion to expand three VH datasets annotated manually and use these expanded datasets to benchmark seven MLLMs. Our evaluation shows that VHExpansion effectively identifies more VH test cases. Moreover, symmetric accuracy, being unbiased, leads to different conclusions about the vulnerability of MLLMs to VH compared to traditional accuracy metric. Finally, we show that fine-tuning MLLMs on the expanded VH dataset generated by VHExpansion mitigates VH more effectively than fine-tuning on the original, manually annotated dataset. Our code is available at: https://github.com/lycheeefish/VHExpansion.

Qingyao Tian, Bingyu Yang, Huai Liao et al.

Vision-language models (VLMs) have recently shown remarkable performance in navigation and localization tasks by leveraging large-scale pretraining for semantic understanding. However, applying VLMs to 6-DoF endoscopic camera localization presents several challenges: 1) the lack of large-scale, high-quality, densely annotated, and localization-oriented vision-language datasets in real-world medical settings; 2) limited capability for fine-grained pose regression; and 3) high computational latency when extracting temporal features from past frames. To address these issues, we first construct BREATH dataset, the largest in-vivo endoscopic localization dataset to date, collected in the complex human airway. Building on this dataset, we propose BREATH-VL, a hybrid framework that integrates semantic cues from VLMs with geometric information from vision-based registration methods for accurate 6-DoF pose estimation. Our motivation lies in the complementary strengths of both approaches: VLMs offer generalizable semantic understanding, while registration methods provide precise geometric alignment. To further enhance the VLM's ability to capture temporal context, we introduce a lightweight context-learning mechanism that encodes motion history as linguistic prompts, enabling efficient temporal reasoning without expensive video-level computation. Extensive experiments demonstrate that the vision-language module delivers robust semantic localization in challenging surgical scenes. Building on this, our BREATH-VL outperforms state-of-the-art vision-only localization methods in both accuracy and generalization, reducing translational error by 25.5% compared with the best-performing baseline, while achieving competitive computational latency.

Yimeng Geng, Mingyang Zhao, Fan Xu et al.

Ultrasound deformable registration estimates spatial transformations between pairs of deformed ultrasound images, which is crucial for capturing biomechanical properties and enhancing diagnostic accuracy in diseases such as thyroid nodules and breast cancer. However, ultrasound deformable registration remains highly challenging, especially under large deformation. The inherently low contrast, heavy noise and ambiguous tissue boundaries in ultrasound images severely hinder reliable feature extraction and correspondence matching. Existing methods often suffer from poor anatomical alignment and lack physical interpretability. To address the problem, we propose PADReg, a physics-aware deformable registration framework guided by contact force. PADReg leverages synchronized contact force measured by robotic ultrasound systems as a physical prior to constrain the registration. Specifically, instead of directly predicting deformation fields, we first construct a pixel-wise stiffness map utilizing the multi-modal information from contact force and ultrasound images. The stiffness map is then combined with force data to estimate a dense deformation field, through a lightweight physics-aware module inspired by Hooke's law. This design enables PADReg to achieve physically plausible registration with better anatomical alignment than previous methods relying solely on image similarity. Experiments on in-vivo datasets demonstrate that it attains a HD95 of 12.90, which is 21.34\% better than state-of-the-art methods. The source code is available at https://github.com/evelynskip/PADReg.

Bingyu Yang, Qingyao Tian, Yimeng Geng et al.

Generalizable dense feature matching in endoscopic images is crucial for robot-assisted tasks, including 3D reconstruction, navigation, and surgical scene understanding. Yet, it remains a challenge due to difficult visual conditions (e.g., weak textures, large viewpoint variations) and a scarcity of annotated data. To address these challenges, we propose EndoMatcher, a generalizable endoscopic image matcher via large-scale, multi-domain data pre-training. To address difficult visual conditions, EndoMatcher employs a two-branch Vision Transformer to extract multi-scale features, enhanced by dual interaction blocks for robust correspondence learning. To overcome data scarcity and improve domain diversity, we construct Endo-Mix6, the first multi-domain dataset for endoscopic matching. Endo-Mix6 consists of approximately 1.2M real and synthetic image pairs across six domains, with correspondence labels generated using Structure-from-Motion and simulated transformations. The diversity and scale of Endo-Mix6 introduce new challenges in training stability due to significant variations in dataset sizes, distribution shifts, and error imbalance. To address them, a progressive multi-objective training strategy is employed to promote balanced learning and improve representation quality across domains. This enables EndoMatcher to generalize across unseen organs and imaging conditions in a zero-shot fashion. Extensive zero-shot matching experiments demonstrate that EndoMatcher increases the number of inlier matches by 140.69% and 201.43% on the Hamlyn and Bladder datasets over state-of-the-art methods, respectively, and improves the Matching Direction Prediction Accuracy (MDPA) by 9.40% on the Gastro-Matching dataset, achieving dense and accurate matching under challenging endoscopic conditions. The code is publicly available at https://github.com/Beryl2000/EndoMatcher.

Yiming Huang, Long Bai, Beilei Cui et al.

Accurate reconstruction of soft tissue is crucial for advancing automation in image-guided robotic surgery. The recent 3D Gaussian Splatting (3DGS) techniques and their variants, 4DGS, achieve high-quality renderings of dynamic surgical scenes in real-time. However, 3D-GS-based methods still struggle in scenarios with varying illumination, such as low light and over-exposure. Training 3D-GS in such extreme light conditions leads to severe optimization problems and devastating rendering quality. To address these challenges, we present Endo-4DGX, a novel reconstruction method with illumination-adaptive Gaussian Splatting designed specifically for endoscopic scenes with uneven lighting. By incorporating illumination embeddings, our method effectively models view-dependent brightness variations. We introduce a region-aware enhancement module to model the sub-area lightness at the Gaussian level and a spatial-aware adjustment module to learn the view-consistent brightness adjustment. With the illumination adaptive design, Endo-4DGX achieves superior rendering performance under both low-light and over-exposure conditions while maintaining geometric accuracy. Additionally, we employ an exposure control loss to restore the appearance from adverse exposure to the normal level for illumination-adaptive optimization. Experimental results demonstrate that Endo-4DGX significantly outperforms combinations of state-of-the-art reconstruction and restoration methods in challenging lighting environments, underscoring its potential to advance robot-assisted surgical applications. Our code is available at https://github.com/lastbasket/Endo-4DGX.

Hongbin Liu, Moyang Guo, Zhengyuan Jiang et al.

The increasing realism of synthetic speech, driven by advancements in text-to-speech models, raises ethical concerns regarding impersonation and disinformation. Audio watermarking offers a promising solution via embedding human-imperceptible watermarks into AI-generated audios. However, the robustness of audio watermarking against common/adversarial perturbations remains understudied. We present AudioMarkBench, the first systematic benchmark for evaluating the robustness of audio watermarking against watermark removal and watermark forgery. AudioMarkBench includes a new dataset created from Common-Voice across languages, biological sexes, and ages, 3 state-of-the-art watermarking methods, and 15 types of perturbations. We benchmark the robustness of these methods against the perturbations in no-box, black-box, and white-box settings. Our findings highlight the vulnerabilities of current watermarking techniques and emphasize the need for more robust and fair audio watermarking solutions. Our dataset and code are publicly available at https://github.com/moyangkuo/AudioMarkBench.

Kun Zhao, Yongkun Liu, Siyuan Hao et al.

Street view images classification aiming at urban land use analysis is difficult because the class labels (e.g., commercial area), are concepts with higher abstract level compared to the ones of general visual tasks (e.g., persons and cars). Therefore, classification models using only visual features often fail to achieve satisfactory performance. In this paper, a novel approach based on a "Detector-Encoder-Classifier" framework is proposed. Instead of using visual features of the whole image directly as common image-level models based on convolutional neural networks (CNNs) do, the proposed framework firstly obtains the bounding boxes of buildings in street view images from a detector. Their contextual information such as the co-occurrence patterns of building classes and their layout are then encoded into metadata by the proposed algorithm "CODING" (Context encOding of Detected buildINGs). Finally, these bounding box metadata are classified by a recurrent neural network (RNN). In addition, we made a dual-labeled dataset named "BEAUTY" (Building dEtection And Urban funcTional-zone portraYing) of 19,070 street view images and 38,857 buildings based on the existing BIC GSV [1]. The dataset can be used not only for street view image classification, but also for multi-class building detection. Experiments on "BEAUTY" show that the proposed approach achieves a 12.65% performance improvement on macro-precision and 12% on macro-recall over image-level CNN based models. Our code and dataset are available at https://github.com/kyle-one/Context-Encoding-of-Detected-Buildings/

Haoyan Gong, Hongbin Liu

Real-world License Plate Recognition (LPR) faces significant challenges from severe degradations such as motion blur, low resolution, and complex illumination. The prevailing "restoration-then-recognition" two-stage paradigm suffers from a fundamental flaw: the pixel-level optimization objectives of image restoration models are misaligned with the semantic goals of character recognition, leading to artifact interference and error accumulation. While Vision-Language Models (VLMs) have demonstrated powerful general capabilities, they lack explicit structural modeling for license plate character sequences (e.g., fixed length, specific order). To address this, we propose an end-to-end structure-aware multimodal reasoning framework based on Qwen3-VL. The core innovation lies in the Character-Aware Multimodal Reasoning Module (CMRM), which introduces a set of learnable Character Slot Queries. Through a cross-attention mechanism, these queries actively retrieve fine-grained evidence corresponding to character positions from visual features. Subsequently, we inject these character-aware representations back into the visual tokens via residual modulation, enabling the language model to perform autoregressive generation based on explicit structural priors. Furthermore, combined with the LoRA parameter-efficient fine-tuning strategy, the model achieves domain adaptation while retaining the generalization capabilities of the large model. Extensive experiments on both synthetic and real-world severely degraded datasets demonstrate that our method significantly outperforms existing restoration-recognition combinations and general VLMs, validating the superiority of incorporating structured reasoning into large models for low-quality text recognition tasks.

Ruochen Liu, Yi Tian, Jiahao Wang et al.

Cell detection in pathological images presents unique challenges due to densely packed objects, subtle inter-class differences, and severe background clutter. In this paper, we propose CellMamba, a lightweight and accurate one-stage detector tailored for fine-grained biomedical instance detection. Built upon a VSSD backbone, CellMamba integrates CellMamba Blocks, which couple either NC-Mamba or Multi-Head Self-Attention (MSA) with a novel Triple-Mapping Adaptive Coupling (TMAC) module. TMAC enhances spatial discriminability by splitting channels into two parallel branches, equipped with dual idiosyncratic and one consensus attention map, adaptively fused to preserve local sensitivity and global consistency. Furthermore, we design an Adaptive Mamba Head that fuses multi-scale features via learnable weights for robust detection under varying object sizes. Extensive experiments on two public datasets-CoNSeP and CytoDArk0-demonstrate that CellMamba outperforms both CNN-based, Transformer-based, and Mamba-based baselines in accuracy, while significantly reducing model size and inference latency. Our results validate CellMamba as an efficient and effective solution for high-resolution cell detection.

Guankun Wang, Long Bai, Wan Jun Nah et al.

Recent advancements in Surgical Visual Question Answering (Surgical-VQA) and related region grounding have shown great promise for robotic and medical applications, addressing the critical need for automated methods in personalized surgical mentorship. However, existing models primarily provide simple structured answers and struggle with complex scenarios due to their limited capability in recognizing long-range dependencies and aligning multimodal information. In this paper, we introduce Surgical-LVLM, a novel personalized large vision-language model tailored for complex surgical scenarios. Leveraging the pre-trained large vision-language model and specialized Visual Perception LoRA (VP-LoRA) blocks, our model excels in understanding complex visual-language tasks within surgical contexts. In addressing the visual grounding task, we propose the Token-Interaction (TIT) module, which strengthens the interaction between the grounding module and the language responses of the Large Visual Language Model (LVLM) after projecting them into the latent space. We demonstrate the effectiveness of Surgical-LVLM on several benchmarks, including EndoVis-17-VQLA, EndoVis-18-VQLA, and a newly introduced EndoVis Conversations dataset, which sets new performance standards. Our work contributes to advancing the field of automated surgical mentorship by providing a context-aware solution.

Henry W. Sprueill, Carl Edwards, Khushbu Agarwal et al.

The discovery of new catalysts is essential for the design of new and more efficient chemical processes in order to transition to a sustainable future. We introduce an AI-guided computational screening framework unifying linguistic reasoning with quantum-chemistry based feedback from 3D atomistic representations. Our approach formulates catalyst discovery as an uncertain environment where an agent actively searches for highly effective catalysts via the iterative combination of large language model (LLM)-derived hypotheses and atomistic graph neural network (GNN)-derived feedback. Identified catalysts in intermediate search steps undergo structural evaluation based on spatial orientation, reaction pathways, and stability. Scoring functions based on adsorption energies and reaction energy barriers steer the exploration in the LLM's knowledge space toward energetically favorable, high-efficiency catalysts. We introduce planning methods that automatically guide the exploration without human input, providing competitive performance against expert-enumerated chemical descriptor-based implementations. By integrating language-guided reasoning with computational chemistry feedback, our work pioneers AI-accelerated, trustworthy catalyst discovery.

Guankun Wang, Long Bai, Junyi Wang et al.

Recently, Multimodal Large Language Models (MLLMs) have demonstrated their immense potential in computer-aided diagnosis and decision-making. In the context of robotic-assisted surgery, MLLMs can serve as effective tools for surgical training and guidance. However, there is still a lack of MLLMs specialized for surgical scene understanding in clinical applications. In this work, we introduce EndoChat to address various dialogue paradigms and subtasks in surgical scene understanding that surgeons encounter. To train our EndoChat, we construct the Surg-396K dataset through a novel pipeline that systematically extracts surgical information and generates structured annotations based on collected large-scale endoscopic surgery datasets. Furthermore, we introduce a multi-scale visual token interaction mechanism and a visual contrast-based reasoning mechanism to enhance the model's representation learning and reasoning capabilities. Our model achieves state-of-the-art performance across five dialogue paradigms and eight surgical scene understanding tasks. Additionally, we conduct evaluations with professional surgeons, most of whom provide positive feedback on collaborating with EndoChat. Overall, these results demonstrate that our EndoChat has great potential to significantly advance training and automation in robotic-assisted surgery.