Cong Xu

Hao Zhang, Cong Xu, Shuaijie Zhang

With the rapid development of detectors, Bounding Box Regression (BBR) loss function has constantly updated and optimized. However, the existing IoU-based BBR still focus on accelerating convergence by adding new loss terms, ignoring the limitations of IoU loss term itself. Although theoretically IoU loss can effectively describe the state of bounding box regression,in practical applications, it cannot adjust itself according to different detectors and detection tasks, and does not have strong generalization. Based on the above, we first analyzed the BBR model and concluded that distinguishing different regression samples and using different scales of auxiliary bounding boxes to calculate losses can effectively accelerate the bounding box regression process. For high IoU samples, using smaller auxiliary bounding boxes to calculate losses can accelerate convergence, while larger auxiliary bounding boxes are suitable for low IoU samples. Then, we propose Inner-IoU loss, which calculates IoU loss through auxiliary bounding boxes. For different datasets and detectors, we introduce a scaling factor ratio to control the scale size of the auxiliary bounding boxes for calculating losses. Finally, integrate Inner-IoU into the existing IoU-based loss functions for simulation and comparative experiments. The experiment result demonstrate a further enhancement in detection performance with the utilization of the method proposed in this paper, verifying the effectiveness and generalization ability of Inner-IoU loss. Code is available at https://github.com/malagoutou/Inner-IoU.

Huaying Hao, Cong Xu, Dan Zhang et al.

Retinal Optical Coherence Tomography Angiography (OCTA) with high-resolution is important for the quantification and analysis of retinal vasculature. However, the resolution of OCTA images is inversely proportional to the field of view at the same sampling frequency, which is not conducive to clinicians for analyzing larger vascular areas. In this paper, we propose a novel Sparse-based domain Adaptation Super-Resolution network (SASR) for the reconstruction of realistic 6x6 mm2/low-resolution (LR) OCTA images to high-resolution (HR) representations. To be more specific, we first perform a simple degradation of the 3x3 mm2/high-resolution (HR) image to obtain the synthetic LR image. An efficient registration method is then employed to register the synthetic LR with its corresponding 3x3 mm2 image region within the 6x6 mm2 image to obtain the cropped realistic LR image. We then propose a multi-level super-resolution model for the fully-supervised reconstruction of the synthetic data, guiding the reconstruction of the realistic LR images through a generative-adversarial strategy that allows the synthetic and realistic LR images to be unified in the feature domain. Finally, a novel sparse edge-aware loss is designed to dynamically optimize the vessel edge structure. Extensive experiments on two OCTA sets have shown that our method performs better than state-of-the-art super-resolution reconstruction methods. In addition, we have investigated the performance of the reconstruction results on retina structure segmentations, which further validate the effectiveness of our approach.

Cong Xu, Pu Luo, Yumei Li et al.

This paper describes our approach for the 8th UG2+ Workshop (CVPR 2026) Track~2, which targets semantic segmentation of outdoor scenes degraded by five weather conditions: blur, darkness, snow, haze, and glare. A central challenge we observe is a severe generalization gap -- models that perform well on the validation set often collapse on the test set. For instance, SegFormer-B5 drops 16.1 mIoU points from validation to test, suggesting that model capacity alone is insufficient for robustness. We investigate whether a carefully designed training recipe, rather than architectural complexity, can address this gap. Starting from a pre-trained SegMAN-S backbone, we systematically study the effects of domain-adaptive fine-tuning, multi-source data mixing, scene-balanced sampling, and synthetic degradation augmentation. Our final system achieves 59.9\% mIoU on the official test set while maintaining a validation-test gap of only 6.5 points -- less than half that of larger models. We analyze negative results from architectural modifications, loss function variants, and model scaling to provide practical insights for weather-robust segmentation under limited data.

Siyu Wu, Cong Xu, Wei Zhang

Knowledge Tracing (KT) models students' knowledge states based on learning interactions to predict performance. While deep learning-based KT models have boosted predictive accuracy, most models rely on deterministic vector embeddings and opaque latent state transitions, limiting interpretability regarding how specific past behaviors influence predictions. To address this limitation, we propose Probabilistic Logical Knowledge Tracing (PLKT), an interpretable KT framework that formulates prediction as a goal-conditioned evidence reasoning process over historical learning behaviors. Instead of representing knowledge states as deterministic vector embeddings, PLKT employs robust Beta-distributed probabilistic embeddings to represent student knowledge states. This probabilistic foundation allows us to model the uncertainty of historical behaviors and perform explicit logical operations (e.g., conjunction), constructing transparent reasoning paths that reveal how specific past interactions contribute to the prediction. Extensive experiments show that PLKT outperforms state-of-the-art KT methods while achieving superior interpretability. Our code is available at https://anonymous.4open.science/r/PLKT-D3CE/.

Yufeng Wang, Cong Xu, Min Yang et al.

Although Physics-Informed Neural Networks (PINNs) have been successfully applied in a wide variety of science and engineering fields, they can fail to accurately predict the underlying solution in slightly challenging convection-diffusion-reaction problems. In this paper, we investigate the reason of this failure from a domain distribution perspective, and identify that learning multi-scale fields simultaneously makes the network unable to advance its training and easily get stuck in poor local minima. We show that the widespread experience of sampling more collocation points in high-loss layer regions hardly help optimize and may even worsen the results. These findings motivate the development of a novel curriculum learning method that encourages neural networks to prioritize learning on easier non-layer regions while downplaying learning on harder layer regions. The proposed method helps PINNs automatically adjust the learning emphasis and thereby facilitate the optimization procedure. Numerical results on typical benchmark equations show that the proposed curriculum learning approach mitigates the failure modes of PINNs and can produce accurate results for very sharp boundary and interior layers. Our work reveals that for equations whose solutions have large scale differences, paying less attention to high-loss regions can be an effective strategy for learning them accurately.

Giacomo Pedretti, John Moon, Pedro Bruel et al.

Structured, or tabular, data is the most common format in data science. While deep learning models have proven formidable in learning from unstructured data such as images or speech, they are less accurate than simpler approaches when learning from tabular data. In contrast, modern tree-based Machine Learning (ML) models shine in extracting relevant information from structured data. An essential requirement in data science is to reduce model inference latency in cases where, for example, models are used in a closed loop with simulation to accelerate scientific discovery. However, the hardware acceleration community has mostly focused on deep neural networks and largely ignored other forms of machine learning. Previous work has described the use of an analog content addressable memory (CAM) component for efficiently mapping random forests. In this work, we develop an analog-digital architecture that implements a novel increased precision analog CAM and a programmable chip for inference of state-of-the-art tree-based ML models, such as XGBoost, CatBoost, and others. Thanks to hardware-aware training, X-TIME reaches state-of-the-art accuracy and 119x higher throughput at 9740x lower latency with >150x improved energy efficiency compared with a state-of-the-art GPU for models with up to 4096 trees and depth of 8, with a 19W peak power consumption.

Cong Xu, Jun Wang, Jianyong Wang et al.

Embedding plays a key role in modern recommender systems because they are virtual representations of real-world entities and the foundation for subsequent decision-making models. In this paper, we propose a novel embedding update mechanism, Structure-aware Embedding Evolution (SEvo for short), to encourage related nodes to evolve similarly at each step. Unlike GNN (Graph Neural Network) that typically serves as an intermediate module, SEvo is able to directly inject graph structural information into embedding with minimal computational overhead during training. The convergence properties of SEvo along with its potential variants are theoretically analyzed to justify the validity of the designs. Moreover, SEvo can be seamlessly integrated into existing optimizers for state-of-the-art performance. Particularly SEvo-enhanced AdamW with moment estimate correction demonstrates consistent improvements across a spectrum of models and datasets, suggesting a novel technical route to effectively utilize graph structural information beyond explicit GNN modules.

Runze Zhang, Guoguang Du, Xiaochuan Li et al.

Spatio-temporal consistency is a critical research topic in video generation. A qualified generated video segment must ensure plot plausibility and coherence while maintaining visual consistency of objects and scenes across varying viewpoints. Prior research, especially in open-source projects, primarily focuses on either temporal or spatial consistency, or their basic combination, such as appending a description of a camera movement after a prompt without constraining the outcomes of this movement. However, camera movement may introduce new objects to the scene or eliminate existing ones, thereby overlaying and affecting the preceding narrative. Especially in videos with numerous camera movements, the interplay between multiple plots becomes increasingly complex. This paper introduces and examines integral spatio-temporal consistency, considering the synergy between plot progression and camera techniques, and the long-term impact of prior content on subsequent generation. Our research encompasses dataset construction through to the development of the model. Initially, we constructed a DropletVideo-10M dataset, which comprises 10 million videos featuring dynamic camera motion and object actions. Each video is annotated with an average caption of 206 words, detailing various camera movements and plot developments. Following this, we developed and trained the DropletVideo model, which excels in preserving spatio-temporal coherence during video generation. The DropletVideo dataset and model are accessible at https://dropletx.github.io.

Yunsheng Pang, Zijian Liu, Yudong Li et al.

Slate recommendation, which presents users with a ranked item list in a single display, is ubiquitous across mainstream online platforms. Recent advances in generative models have shown significant potential for this task via autoregressive modeling of discrete semantic ID sequences. However, existing methods suffer from three key limitations: entangled item tokenization, inefficient sequential decoding, and the absence of holistic slate planning. These issues often result in substantial inference overhead and inadequate alignment with diverse user preferences and practical business requirements, hindering the industrial deployment of generative slate recommendation systems. In this paper, we propose HiGR, an efficient generative slate recommendation framework that integrates hierarchical planning with listwise preference alignment. First, we design an auto-encoder incorporating residual quantization and contrastive constraints, which tokenizes items into semantically structured IDs to enable controllable generation. Second, HiGR decouples the generation process into two stages: a list-level planning stage to capture global slate intent, and an item-level decoding stage to select specific items, effectively reducing the search space and enabling efficient generation. Third, we introduce a multi-objective and listwise preference alignment mechanism that enhances slate quality by leveraging implicit user feedback. Extensive experiments have validated the effectiveness of our HiGR method. Notably, it outperforms state-of-the-art baselines by over 10\% in offline recommendation quality while achieving a $5\times$ inference speedup. Furthermore, we have deployed HiGR on a commercial platform under Tencent (serving hundreds of millions of users), and online A/B tests show that it increases average watch time and average video plays by 1.22\% and 1.73\%, respectively.

Cong Xu, Wei Zhang, Jun Wang et al.

The adversarial robustness of a neural network mainly relies on two factors: model capacity and anti-perturbation ability. In this paper, we study the anti-perturbation ability of the network from the feature maps of convolutional layers. Our theoretical analysis discovers that larger convolutional feature maps before average pooling can contribute to better resistance to perturbations, but the conclusion is not true for max pooling. It brings new inspiration to the design of robust neural networks and urges us to apply these findings to improve existing architectures. The proposed modifications are very simple and only require upsampling the inputs or slightly modifying the stride configurations of downsampling operators. We verify our approaches on several benchmark neural network architectures, including AlexNet, VGG, RestNet18, and PreActResNet18. Non-trivial improvements in terms of both natural accuracy and adversarial robustness can be achieved under various attack and defense mechanisms. The code is available at \url{https://github.com/MTandHJ/rcm}.

Yufeng Wang, Dan Li, Cong Xu et al.

Missing data imputation is a fundamental problem in data analysis, and many studies have been conducted to improve its performance by exploring model structures and learning procedures. However, data augmentation, as a simple yet effective method, has not received enough attention in this area. In this paper, we propose a novel data augmentation method called Missingness Augmentation (MisA) for generative imputation models. Our approach dynamically produces incomplete samples at each epoch by utilizing the generator's output, constraining the augmented samples using a simple reconstruction loss, and combining this loss with the original loss to form the final optimization objective. As a general augmentation technique, MisA can be easily integrated into generative imputation frameworks, providing a simple yet effective way to enhance their performance. Experimental results demonstrate that MisA significantly improves the performance of many recently proposed generative imputation models on a variety of tabular and image datasets. The code is available at \url{https://github.com/WYu-Feng/Missingness-Augmentation}.

Cong Xu, Xiang Li, Min Yang

Neural networks are susceptible to artificially designed adversarial perturbations. Recent efforts have shown that imposing certain modifications on classification layer can improve the robustness of the neural networks. In this paper, we explicitly construct a dense orthogonal weight matrix whose entries have the same magnitude, thereby leading to a novel robust classifier. The proposed classifier avoids the undesired structural redundancy issue in previous work. Applying this classifier in standard training on clean data is sufficient to ensure the high accuracy and good robustness of the model. Moreover, when extra adversarial samples are used, better robustness can be further obtained with the help of a special worst-case loss. Experimental results show that our method is efficient and competitive to many state-of-the-art defensive approaches. Our code is available at \url{https://github.com/MTandHJ/roboc}.

Wei Wen, Cong Xu, Feng Yan et al.

High network communication cost for synchronizing gradients and parameters is the well-known bottleneck of distributed training. In this work, we propose TernGrad that uses ternary gradients to accelerate distributed deep learning in data parallelism. Our approach requires only three numerical levels {-1,0,1}, which can aggressively reduce the communication time. We mathematically prove the convergence of TernGrad under the assumption of a bound on gradients. Guided by the bound, we propose layer-wise ternarizing and gradient clipping to improve its convergence. Our experiments show that applying TernGrad on AlexNet does not incur any accuracy loss and can even improve accuracy. The accuracy loss of GoogLeNet induced by TernGrad is less than 2% on average. Finally, a performance model is proposed to study the scalability of TernGrad. Experiments show significant speed gains for various deep neural networks. Our source code is available.

Wei Wen, Cong Xu, Chunpeng Wu et al.

Very large-scale Deep Neural Networks (DNNs) have achieved remarkable successes in a large variety of computer vision tasks. However, the high computation intensity of DNNs makes it challenging to deploy these models on resource-limited systems. Some studies used low-rank approaches that approximate the filters by low-rank basis to accelerate the testing. Those works directly decomposed the pre-trained DNNs by Low-Rank Approximations (LRA). How to train DNNs toward lower-rank space for more efficient DNNs, however, remains as an open area. To solve the issue, in this work, we propose Force Regularization, which uses attractive forces to enforce filters so as to coordinate more weight information into lower-rank space. We mathematically and empirically verify that after applying our technique, standard LRA methods can reconstruct filters using much lower basis and thus result in faster DNNs. The effectiveness of our approach is comprehensively evaluated in ResNets, AlexNet, and GoogLeNet. In AlexNet, for example, Force Regularization gains 2x speedup on modern GPU without accuracy loss and 4.05x speedup on CPU by paying small accuracy degradation. Moreover, Force Regularization better initializes the low-rank DNNs such that the fine-tuning can converge faster toward higher accuracy. The obtained lower-rank DNNs can be further sparsified, proving that Force Regularization can be integrated with state-of-the-art sparsity-based acceleration methods. Source code is available in https://github.com/wenwei202/caffe

Peng Liu, Cong Xu, Jiawei Zhu et al.

As a key stage of Recommender Systems (RSs), Multi-Task Fusion (MTF) is responsible for merging multiple scores output by Multi-Task Learning (MTL) into a single score, finally determining the recommendation results. Recently, Reinforcement Learning (RL) has been applied to MTF to maximize long-term user satisfaction within a recommendation session. However, due to limitations in modeling paradigm, all existing RL algorithms for MTF can only utilize user features and statistical features as the state to generate actions at the user level, but unable to leverage item features and other valuable features, which leads to suboptimal performance. Overcoming this problem requires a breakthrough in the existing modeling paradigm, yet, to date, no prior work has addressed it. To tackle this challenge, we propose EnhancedRL, an innovative RL algorithm. Unlike existing RL-MTF methods, EnhancedRL takes the enhanced state as input, incorporating not only user features but also item features and other valuable information. Furthermore, it introduces a tailored actor-critic framework - including redesigned actor and critics and a novel learning procedure - to optimize long-term rewards at the user-item pair level within a recommendation session. Extensive offline and online experiments are conducted in an industrial RS and the results demonstrate that EnhancedRL outperforms other methods remarkably, achieving a +3.84% increase in user valid consumption and a +0.58% increase in user duration time. To the best of our knowledge, EnhancedRL is the first work to address this challenge, and it has been fully deployed in a large-scale RS since September 14, 2023, yielding significant improvements.

Yunhang He, Cong Xu, Zhangchi Zhu et al.

Graph filter design is central to spectral collaborative filtering, yet most existing methods rely on manually tuned hyperparameters rather than fully learnable filters. We show that this challenge stems from a bias in traditional recommendation objectives, which induces a spectral phenomenon termed low-frequency explosion, thereby fundamentally hindering the effective learning of graph filters. To overcome this limitation, we propose a novel adaptive spectral graph collaborative filtering framework (ASPIRE) based on a bi-level optimization objective. Guided by our theoretical analysis, we disentangle the filter learning objective, which in turn leads to excellent recommendation performance, spectral adaptivity, and training stability in practice. Extensive experiments show our learned filters match the performance of carefully engineered task-specific designs. Furthermore, ASPIRE is equally effective in LLM-powered collaborative filtering. Our findings demonstrate that graph filter learning is viable and generalizable, paving the way for more expressive graph neural networks in collaborative filtering.

Peng Liu, Cong Xu, Ming Zhao et al.

As the last pivotal stage of Recommender System (RS), Multi-Task Fusion (MTF) is responsible for combining multiple scores outputted by Multi-Task Learning (MTL) model into a final score to maximize user satisfaction. Recently, to optimize long-term user satisfaction, Reinforcement Learning (RL) is used for MTF in RSs. However, the existing offline RL algorithms used for MTF have the following severe problems: a) To avoid Out-of-Distribution (OOD), their constraints are overly strict, which seriously damage performance; b) They are unaware of the exploration policy used to collect training data, only suboptimal policy can be learned; c) Their exploration policies are inefficient and hurt user experience. To solve the above problems, we propose an innovative method called UnifiedRL tailored for MTF in large-scale RSs. UnifiedRL seamlessly integrates offline RL model with its custom exploration policy to relax overly strict constraints, which is different from existing RL-MTF methods and significantly improves performance. In addition, compared to existing exploration policies, UnifiedRL's custom exploration policy is highly efficient, enabling frequent online exploration and offline training iterations, which further improves performance. Extensive offline and online experiments are conducted in a large-scale RS. The results demonstrate that UnifiedRL outperforms other existing MTF methods remarkably, achieving a +4.64% increase in user valid consumption and a +1.74% increase in user duration time. To the best of our knowledge, UnifiedRL is the first RL algorithm tailored for MTF in RSs and has been successfully deployed in multiple large-scale RSs since June 2023, yielding significant benefits.

Peng Xu, Shengwu Xiong, Jiajun Zhang et al.

This paper reviews the MARS2 2025 Challenge on Multimodal Reasoning. We aim to bring together different approaches in multimodal machine learning and LLMs via a large benchmark. We hope it better allows researchers to follow the state-of-the-art in this very dynamic area. Meanwhile, a growing number of testbeds have boosted the evolution of general-purpose large language models. Thus, this year's MARS2 focuses on real-world and specialized scenarios to broaden the multimodal reasoning applications of MLLMs. Our organizing team released two tailored datasets Lens and AdsQA as test sets, which support general reasoning in 12 daily scenarios and domain-specific reasoning in advertisement videos, respectively. We evaluated 40+ baselines that include both generalist MLLMs and task-specific models, and opened up three competition tracks, i.e., Visual Grounding in Real-world Scenarios (VG-RS), Visual Question Answering with Spatial Awareness (VQA-SA), and Visual Reasoning in Creative Advertisement Videos (VR-Ads). Finally, 76 teams from the renowned academic and industrial institutions have registered and 40+ valid submissions (out of 1200+) have been included in our ranking lists. Our datasets, code sets (40+ baselines and 15+ participants' methods), and rankings are publicly available on the MARS2 workshop website and our GitHub organization page https://github.com/mars2workshop/, where our updates and announcements of upcoming events will be continuously provided.

Shaowen Wang, Anan Liu, Jian Xiao et al.

Modern recommendation systems frequently employ online learning to dynamically update their models with freshly collected data. The most commonly used optimizer for updating neural networks in these contexts is the Adam optimizer, which integrates momentum ($m_t$) and adaptive learning rate ($v_t$). However, the volatile nature of online learning data, characterized by its frequent distribution shifts and presence of noise, poses significant challenges to Adam's standard optimization process: (1) Adam may use outdated momentum and the average of squared gradients, resulting in slower adaptation to distribution changes, and (2) Adam's performance is adversely affected by data noise. To mitigate these issues, we introduce CAdam, a confidence-based optimization strategy that assesses the consistency between the momentum and the gradient for each parameter dimension before deciding on updates. If momentum and gradient are in sync, CAdam proceeds with parameter updates according to Adam's original formulation; if not, it temporarily withholds updates and monitors potential shifts in data distribution in subsequent iterations. This method allows CAdam to distinguish between the true distributional shifts and mere noise, and to adapt more quickly to new data distributions. In various settings with distribution shift or noise, our experiments demonstrate that CAdam surpasses other well-known optimizers, including the original Adam. Furthermore, in large-scale A/B testing within a live recommendation system, CAdam significantly enhances model performance compared to Adam, leading to substantial increases in the system's gross merchandise volume (GMV).

Wenquan Dong, Edward T. A. Mitchard, Yuwei Chen et al.

Large-scale high spatial resolution aboveground biomass (AGB) maps play a crucial role in determining forest carbon stocks and how they are changing, which is instrumental in understanding the global carbon cycle, and implementing policy to mitigate climate change. The advent of the new space-borne LiDAR sensor, NASA's GEDI instrument, provides unparalleled possibilities for the accurate and unbiased estimation of forest AGB at high resolution, particularly in dense and tall forests, where Synthetic Aperture Radar (SAR) and passive optical data exhibit saturation. However, GEDI is a sampling instrument, collecting dispersed footprints, and its data must be combined with that from other continuous cover satellites to create high-resolution maps, using local machine learning methods. In this study, we developed local models to estimate forest AGB from GEDI L2A data, as the models used to create GEDI L4 AGB data incorporated minimal field data from China. We then applied LightGBM and random forest regression to generate wall-to-wall AGB maps at 25 m resolution, using extensive GEDI footprints as well as Sentinel-1 data, ALOS-2 PALSAR-2 and Sentinel-2 optical data. Through a 5-fold cross-validation, LightGBM demonstrated a slightly better performance than Random Forest across two contrasting regions. However, in both regions, the computation speed of LightGBM is substantially faster than that of the random forest model, requiring roughly one-third of the time to compute on the same hardware. Through the validation against field data, the 25 m resolution AGB maps generated using the local models developed in this study exhibited higher accuracy compared to the GEDI L4B AGB data. We found in both regions an increase in error as slope increased. The trained models were tested on nearby but different regions and exhibited good performance.

Peng Liu, Nian Wang, Cong Xu et al.

Recommender System (RS) provides personalized recommendation service based on user interest. However, lots of users' interests are sparse due to lacking consumption behaviors, making it challenging to provide accurate recommendations for them, which is widespread in large-scale RSs. In particular, efficiently solving this problem in the ranking stage of RS is an even greater challenge, which requires an end-to-end and real-time approach. To solve this problem, we propose an innovative method called Dynamic User Interest Augmentation (DUIA). DUIA enhances user interest including user profile and user history behavior sequences by generating enhancement vectors and personalized enhancement vectors through dynamic stream clustering of similar users and relevant items from multiple perspectives. To realize stream clustering, we specially design an algorithm called Gradient-based Hierarchical Clustering Algorithm (GHCA) for DUIA, which performs clustering via gradient descent and stores the cluster centers in memory networks. Extensive offline and online experiments demonstrate that DUIA not only significantly improves model performance for users with sparse interests but also delivers notable gains for other users. As an end-to-end method, DUIA can be easily integrated with existing models. Furthermore, DUIA is also used for long-tail items and cold-start problem, which also yields excellent improvements. Since 2022, DUIA has been successfully deployed in multiple industrial RSs in Tencent and was made public in May 2024. Moreover, the thoughts behind DUIA, dynamic stream clustering and similarity-based enhancement, have inspired relevant works and have also been applied in other stages of RS.

Nathan C. L. Kong, Dae Lee, Huyen Do et al.

Electrocardiograms (ECGs) and photoplethysmograms (PPGs) are generally used to monitor an individual's cardiovascular health. In clinical settings, ECGs and fingertip PPGs are the main signals used for assessing cardiovascular health, but the equipment necessary for their collection precludes their use in daily monitoring. Although PPGs obtained from wrist-worn devices are susceptible to noise due to motion, they have been widely used to continuously monitor cardiovascular health because of their convenience. Therefore, we would like to combine the ease with which PPGs can be collected with the information that ECGs provide about cardiovascular health by developing models to synthesize ECG signals from paired PPG signals. We tackled this problem using generative adversarial networks (GANs) and found that models trained using the original GAN formulations can be successfully used to synthesize ECG signals from which heart rate can be extracted using standard signal processing pipelines. Incorporating a frequency-domain constraint to model training improved the stability of model performance and also the performance on heart rate estimation.

Andrei Dumitriu, Florin Miron, Florin Tatui et al.

This report presents an overview of the AIM 2025 RipSeg Challenge, a competition designed to advance techniques for automatic rip current segmentation in still images. Rip currents are dangerous, fast-moving flows that pose a major risk to beach safety worldwide, making accurate visual detection an important and underexplored research task. The challenge builds on RipVIS, the largest available rip current dataset, and focuses on single-class instance segmentation, where precise delineation is critical to fully capture the extent of rip currents. The dataset spans diverse locations, rip current types, and camera orientations, providing a realistic and challenging benchmark. In total, $75$ participants registered for this first edition, resulting in $5$ valid test submissions. Teams were evaluated on a composite score combining $F_1$, $F_2$, $AP_{50}$, and $AP_{[50:95]}$, ensuring robust and application-relevant rankings. The top-performing methods leveraged deep learning architectures, domain adaptation techniques, pretrained models, and domain generalization strategies to improve performance under diverse conditions. This report outlines the dataset details, competition framework, evaluation metrics, and final results, providing insights into the current state of rip current segmentation. We conclude with a discussion of key challenges, lessons learned from the submissions, and future directions for expanding RipSeg.

Yihua Wang, Qi Jia, Cong Xu et al.

Multimodal sarcasm detection is a complex task that requires distinguishing subtle complementary signals across modalities while filtering out irrelevant information. Many advanced methods rely on learning shortcuts from datasets rather than extracting intended sarcasm-related features. However, our experiments show that shortcut learning impairs the model's generalization in real-world scenarios. Furthermore, we reveal the weaknesses of current modality fusion strategies for multimodal sarcasm detection through systematic experiments, highlighting the necessity of focusing on effective modality fusion for complex emotion recognition. To address these challenges, we construct MUStARD++$^{R}$ by removing shortcut signals from MUStARD++. Then, a Multimodal Conditional Information Bottleneck (MCIB) model is introduced to enable efficient multimodal fusion for sarcasm detection. Experimental results show that the MCIB achieves the best performance without relying on shortcut learning.

Minhao Wang, Yunhang He, Cong Xu et al.

Recommender systems in concert with Large Language Models (LLMs) present promising avenues for generating semantically-informed recommendations. However, LLM-based recommenders exhibit a tendency to overemphasize semantic correlations within users' interaction history. When taking pretrained collaborative ID embeddings as input, LLM-based recommenders progressively weaken the inherent collaborative signals as the embeddings propagate through LLM backbones layer by layer, as opposed to traditional Transformer-based sequential models in which collaborative signals are typically preserved or even enhanced for state-of-the-art performance. To address this limitation, we introduce FreLLM4Rec, an approach designed to balance semantic and collaborative information from a spectral perspective. Item embeddings that incorporate both semantic and collaborative information are first purified using a Global Graph Low-Pass Filter (G-LPF) to preliminarily remove irrelevant high-frequency noise. Temporal Frequency Modulation (TFM) then actively preserves collaborative signal layer by layer. Note that the collaborative preservation capability of TFM is theoretically guaranteed by establishing a connection between the optimal but hard-to-implement local graph fourier filters and the suboptimal yet computationally efficient frequency-domain filters. Extensive experiments on four benchmark datasets demonstrate that FreLLM4Rec successfully mitigates collaborative signal attenuation and achieves competitive performance, with improvements of up to 8.00\% in NDCG@10 over the best baseline. Our findings provide insights into how LLMs process collaborative information and offer a principled approach for improving LLM-based recommendation systems.

Cong Xu, Wenbin Liang, Mo Yu et al.

The rapid scaling of models has led to prohibitively high training and fine-tuning costs. A major factor accounting for memory consumption is the widespread use of stateful optimizers (e.g., Adam), which maintain auxiliary information of even 2x the model size in order to achieve optimal convergence. We therefore present SOLO in this work to spawn a novel type of optimizer that requires an extremely light memory footprint. While previous efforts have achieved certain success in 8-bit or 4-bit cases, SOLO enables Adam-style optimizers to maintain quantized states with precision as low as 3 bits, or even 2 bits. This immense progress is due to the identification and resolution of two key challenges: the signal swamping problem in unsigned quantization that results in unchanged state dynamics, and the increased gradient variance in signed quantization that leads to incorrect descent directions. The theoretical analysis suggests a tailored logarithmic quantization for the former and a precision-specific momentum hyperparameter for the latter. SOLO can thus be seamlessly applied to Adam-style optimizers, leading to substantial memory savings with minimal accuracy loss.

Cong Xu, Gayathri Saranathan, Mahammad Parwez Alam et al.

Evaluating LLMs and text-to-image models is a computationally intensive task often overlooked. Efficient evaluation is crucial for understanding the diverse capabilities of these models and enabling comparisons across a growing number of new models and benchmarks. To address this, we introduce SubLIME, a data-efficient evaluation framework that employs adaptive sampling techniques, such as clustering and quality-based methods, to create representative subsets of benchmarks. Our approach ensures statistically aligned model rankings compared to full datasets, evidenced by high Pearson correlation coefficients. Empirical analysis across six NLP benchmarks reveals that: (1) quality-based sampling consistently achieves strong correlations (0.85 to 0.95) with full datasets at a 10\% sampling rate such as Quality SE and Quality CPD (2) clustering methods excel in specific benchmarks such as MMLU (3) no single method universally outperforms others across all metrics. Extending this framework, we leverage the HEIM leaderboard to cover 25 text-to-image models on 17 different benchmarks. SubLIME dynamically selects the optimal technique for each benchmark, significantly reducing evaluation costs while preserving ranking integrity and score distribution. Notably, a minimal sampling rate of 1% proves effective for benchmarks like MMLU. Additionally, we demonstrate that employing difficulty-based sampling to target more challenging benchmark segments enhances model differentiation with broader score distributions. We also combine semantic search, tool use, and GPT-4 review to identify redundancy across benchmarks within specific LLM categories, such as coding benchmarks. This allows us to further reduce the number of samples needed to maintain targeted rank preservation. Overall, SubLIME offers a versatile and cost-effective solution for the robust evaluation of LLMs and text-to-image models.

Qi Jia, Baoyu Fan, Cong Xu et al.

Existing video multi-modal sentiment analysis mainly focuses on the sentiment expression of people within the video, yet often neglects the induced sentiment of viewers while watching the videos. Induced sentiment of viewers is essential for inferring the public response to videos, has broad application in analyzing public societal sentiment, effectiveness of advertising and other areas. The micro videos and the related comments provide a rich application scenario for viewers induced sentiment analysis. In light of this, we introduces a novel research task, Multi-modal Sentiment Analysis for Comment Response of Video Induced(MSA-CRVI), aims to inferring opinions and emotions according to the comments response to micro video. Meanwhile, we manually annotate a dataset named Comment Sentiment toward to Micro Video (CSMV) to support this research. It is the largest video multi-modal sentiment dataset in terms of scale and video duration to our knowledge, containing 107,267 comments and 8,210 micro videos with a video duration of 68.83 hours. To infer the induced sentiment of comment should leverage the video content, so we propose the Video Content-aware Comment Sentiment Analysis (VC-CSA) method as baseline to address the challenges inherent in this new task. Extensive experiments demonstrate that our method is showing significant improvements over other established baselines.

Yufeng Wang, Dan Li, Cong Xu et al.

Deep image inpainting research mainly focuses on constructing various neural network architectures or imposing novel optimization objectives. However, on the one hand, building a state-of-the-art deep inpainting model is an extremely complex task, and on the other hand, the resulting performance gains are sometimes very limited. We believe that besides the frameworks of inpainting models, lightweight traditional image processing techniques, which are often overlooked, can actually be helpful to these deep models. In this paper, we enhance the deep image inpainting models with the help of classical image complexity metrics. A knowledge-assisted index composed of missingness complexity and forward loss is presented to guide the batch selection in the training procedure. This index helps find samples that are more conducive to optimization in each iteration and ultimately boost the overall inpainting performance. The proposed approach is simple and can be plugged into many deep inpainting models by changing only a few lines of code. We experimentally demonstrate the improvements for several recently developed image inpainting models on various datasets.

Cong Xu, Dan Li, Min Yang

Recently proposed adversarial self-supervised learning methods usually require big batches and long training epochs to extract robust features, which will bring heavy computational overhead on platforms with limited resources. In order to help the network learn more powerful feature representations in smaller batches and fewer epochs, this paper proposes a novel adversarial momentum contrastive learning method, which introduces two memory banks corresponding to clean samples and adversarial samples, respectively. These memory banks can be dynamically incorporated into the training process to track invariant features among historical mini-batches. Compared with the previous adversarial pre-training model, our method achieves superior performance with smaller batch size and less training epochs. In addition, the model outperforms some state-of-the-art supervised defensive methods on multiple benchmark datasets after being fine-tuned on downstream classification tasks.

Cong Xu, Dan Li, Min Yang

It is well-known that deep neural networks are vulnerable to adversarial attacks. Recent studies show that well-designed classification parts can lead to better robustness. However, there is still much space for improvement along this line. In this paper, we first prove that, from a geometric point of view, the robustness of a neural network is equivalent to some angular margin condition of the classifier weights. We then explain why ReLU type function is not a good choice for activation under this framework. These findings reveal the limitations of the existing approaches and lead us to develop a novel light-weight-penalized defensive method, which is simple and has a good scalability. Empirical results on multiple benchmark datasets demonstrate that our method can effectively improve the robustness of the network without requiring too much additional computation, while maintaining a high classification precision for clean data.

Cong Xu, Min Yang, Jin Zhang

The implementation of conventional sparse principal component analysis (SPCA) on high-dimensional data sets has become a time consuming work. In this paper, a series of subspace projections are constructed efficiently by using Household QR factorization. With the aid of these subspace projections, a fast deflation method, called SPCA-SP, is developed for SPCA. This method keeps a good tradeoff between various criteria, including sparsity, orthogonality, explained variance, balance of sparsity, and computational cost. Comparative experiments on the benchmark data sets confirm the effectiveness of the proposed method.

Wei Wen, Yandan Wang, Feng Yan et al.

In Deep Learning, Stochastic Gradient Descent (SGD) is usually selected as a training method because of its efficiency; however, recently, a problem in SGD gains research interest: sharp minima in Deep Neural Networks (DNNs) have poor generalization; especially, large-batch SGD tends to converge to sharp minima. It becomes an open question whether escaping sharp minima can improve the generalization. To answer this question, we propose SmoothOut framework to smooth out sharp minima in DNNs and thereby improve generalization. In a nutshell, SmoothOut perturbs multiple copies of the DNN by noise injection and averages these copies. Injecting noises to SGD is widely used in the literature, but SmoothOut differs in lots of ways: (1) a de-noising process is applied before parameter updating; (2) noise strength is adapted to filter norm; (3) an alternative interpretation on the advantage of noise injection, from the perspective of sharpness and generalization; (4) usage of uniform noise instead of Gaussian noise. We prove that SmoothOut can eliminate sharp minima. Training multiple DNN copies is inefficient, we further propose an unbiased stochastic SmoothOut which only introduces the overhead of noise injecting and de-noising per batch. An adaptive variant of SmoothOut, AdaSmoothOut, is also proposed to improve generalization. In a variety of experiments, SmoothOut and AdaSmoothOut consistently improve generalization in both small-batch and large-batch training on the top of state-of-the-art solutions.