Jianfei Yang

Jianfei Yang, Xinyan Chen, Dazhuo Wang et al. · berkeley

WiFi sensing has been evolving rapidly in recent years. Empowered by propagation models and deep learning methods, many challenging applications are realized such as WiFi-based human activity recognition and gesture recognition. However, in contrast to deep learning for visual recognition and natural language processing, no sufficiently comprehensive public benchmark exists. In this paper, we review the recent progress on deep learning enabled WiFi sensing, and then propose a benchmark, SenseFi, to study the effectiveness of various deep learning models for WiFi sensing. These advanced models are compared in terms of distinct sensing tasks, WiFi platforms, recognition accuracy, model size, computational complexity, feature transferability, and adaptability of unsupervised learning. It is also regarded as a tutorial for deep learning based WiFi sensing, starting from CSI hardware platform to sensing algorithms. The extensive experiments provide us with experiences in deep model design, learning strategy skills and training techniques for real-world applications. To the best of our knowledge, this is the first benchmark with an open-source library for deep learning in WiFi sensing research. The benchmark codes are available at https://github.com/xyanchen/WiFi-CSI-Sensing-Benchmark.

Jianfei Yang, Xinyan Chen, Han Zou et al. · berkeley

WiFi technology has been applied to various places due to the increasing requirement of high-speed Internet access. Recently, besides network services, WiFi sensing is appealing in smart homes since it is device-free, cost-effective and privacy-preserving. Though numerous WiFi sensing methods have been developed, most of them only consider single smart home scenario. Without the connection of powerful cloud server and massive users, large-scale WiFi sensing is still difficult. In this paper, we firstly analyze and summarize these obstacles, and propose an efficient large-scale WiFi sensing framework, namely EfficientFi. The EfficientFi works with edge computing at WiFi APs and cloud computing at center servers. It consists of a novel deep neural network that can compress fine-grained WiFi Channel State Information (CSI) at edge, restore CSI at cloud, and perform sensing tasks simultaneously. A quantized auto-encoder and a joint classifier are designed to achieve these goals in an end-to-end fashion. To the best of our knowledge, the EfficientFi is the first IoT-cloud-enabled WiFi sensing framework that significantly reduces communication overhead while realizing sensing tasks accurately. We utilized human activity recognition and identification via WiFi sensing as two case studies, and conduct extensive experiments to evaluate the EfficientFi. The results show that it compresses CSI data from 1.368Mb/s to 0.768Kb/s with extremely low error of data reconstruction and achieves over 98% accuracy for human activity recognition.

Yucheng Wang, Yuecong Xu, Jianfei Yang et al.

Multivariate Time-Series (MTS) data is crucial in various application fields. With its sequential and multi-source (multiple sensors) properties, MTS data inherently exhibits Spatial-Temporal (ST) dependencies, involving temporal correlations between timestamps and spatial correlations between sensors in each timestamp. To effectively leverage this information, Graph Neural Network-based methods (GNNs) have been widely adopted. However, existing approaches separately capture spatial dependency and temporal dependency and fail to capture the correlations between Different sEnsors at Different Timestamps (DEDT). Overlooking such correlations hinders the comprehensive modelling of ST dependencies within MTS data, thus restricting existing GNNs from learning effective representations. To address this limitation, we propose a novel method called Fully-Connected Spatial-Temporal Graph Neural Network (FC-STGNN), including two key components namely FC graph construction and FC graph convolution. For graph construction, we design a decay graph to connect sensors across all timestamps based on their temporal distances, enabling us to fully model the ST dependencies by considering the correlations between DEDT. Further, we devise FC graph convolution with a moving-pooling GNN layer to effectively capture the ST dependencies for learning effective representations. Extensive experiments show the effectiveness of FC-STGNN on multiple MTS datasets compared to SOTA methods. The code is available at https://github.com/Frank-Wang-oss/FCSTGNN.

Yikai Tang, Haoran Geng, Jindou Jia et al.

Imitation learning has emerged as a crucial approach for acquiring visuomotor skills from demonstrations, where designing effective observation encoders is essential for policy generalization. However, existing methods tend to struggle once test-time conditions differ from the demonstrations, such as changes in lighting, texture, viewpoint, object placement, or object identity. To address this challenge, we propose DIffusion POlicy with compLementarity Encoders (DIPOLE), a visuomotor policy that learns to fuse complementary modalities through a training-time mechanism rather than a specialized fusion architecture. A modality-wise dropout masks one branch at each training step, encouraging each modality to remain individually informative. A lightweight cross-attention layer then exchanges complementary cues between the two. This design endows DIPOLE with five core strengths: stable high performance across diverse tasks, robustness to visual changes, spatial generalization at sub-centimeter precision, emergent capability beyond either modality, and zero-shot transfer to unseen objects. Across 18 simulated and 4 real-world tasks, DIPOLE outperforms six baselines by 39.1% on average, with gains of 41.5% under unseen visual distractors and 15.2% under randomized object placement.

Lang Deng, Jianfei Yang, Shenghai Yuan et al. · berkeley

As an important biomarker for human identification, human gait can be collected at a distance by passive sensors without subject cooperation, which plays an essential role in crime prevention, security detection and other human identification applications. At present, most research works are based on cameras and computer vision techniques to perform gait recognition. However, vision-based methods are not reliable when confronting poor illuminations, leading to degrading performances. In this paper, we propose a novel multimodal gait recognition method, namely GaitFi, which leverages WiFi signals and videos for human identification. In GaitFi, Channel State Information (CSI) that reflects the multi-path propagation of WiFi is collected to capture human gaits, while videos are captured by cameras. To learn robust gait information, we propose a Lightweight Residual Convolution Network (LRCN) as the backbone network, and further propose the two-stream GaitFi by integrating WiFi and vision features for the gait retrieval task. The GaitFi is trained by the triplet loss and classification loss on different levels of features. Extensive experiments are conducted in the real world, which demonstrates that the GaitFi outperforms state-of-the-art gait recognition methods based on single WiFi or camera, achieving 94.2% for human identification tasks of 12 subjects.

Jianfei Yang, Xinyan Chen, Han Zou et al. · berkeley

WiFi sensing technology has shown superiority in smart homes among various sensors for its cost-effective and privacy-preserving merits. It is empowered by Channel State Information (CSI) extracted from WiFi signals and advanced machine learning models to analyze motion patterns in CSI. Many learning-based models have been proposed for kinds of applications, but they severely suffer from environmental dependency. Though domain adaptation methods have been proposed to tackle this issue, it is not practical to collect high-quality, well-segmented and balanced CSI samples in a new environment for adaptation algorithms, but randomly-captured CSI samples can be easily collected. {\color{black}In this paper, we firstly explore how to learn a robust model from these low-quality CSI samples, and propose AutoFi, an annotation-efficient WiFi sensing model based on a novel geometric self-supervised learning algorithm.} The AutoFi fully utilizes unlabeled low-quality CSI samples that are captured randomly, and then transfers the knowledge to specific tasks defined by users, which is the first work to achieve cross-task transfer in WiFi sensing. The AutoFi is implemented on a pair of Atheros WiFi APs for evaluation. The AutoFi transfers knowledge from randomly collected CSI samples into human gait recognition and achieves state-of-the-art performance. Furthermore, we simulate cross-task transfer using public datasets to further demonstrate its capacity for cross-task learning. For the UT-HAR and Widar datasets, the AutoFi achieves satisfactory results on activity recognition and gesture recognition without any prior training. We believe that the AutoFi takes a huge step toward automatic WiFi sensing without any developer engagement.

Yucheng Wang, Yuecong Xu, Jianfei Yang et al.

Contrastive learning, as a self-supervised learning paradigm, becomes popular for Multivariate Time-Series (MTS) classification. It ensures the consistency across different views of unlabeled samples and then learns effective representations for these samples. Existing contrastive learning methods mainly focus on achieving temporal consistency with temporal augmentation and contrasting techniques, aiming to preserve temporal patterns against perturbations for MTS data. However, they overlook spatial consistency that requires the stability of individual sensors and their correlations. As MTS data typically originate from multiple sensors, ensuring spatial consistency becomes essential for the overall performance of contrastive learning on MTS data. Thus, we propose Graph-Aware Contrasting for spatial consistency across MTS data. Specifically, we propose graph augmentations including node and edge augmentations to preserve the stability of sensors and their correlations, followed by graph contrasting with both node- and graph-level contrasting to extract robust sensor- and global-level features. We further introduce multi-window temporal contrasting to ensure temporal consistency in the data for each sensor. Extensive experiments demonstrate that our proposed method achieves state-of-the-art performance on various MTS classification tasks. The code is available at https://github.com/Frank-Wang-oss/TS-GAC.

Jianfei Yang, Yunjiao Zhou, He Huang et al. · berkeley

Avatar refers to a representative of a physical user in the virtual world that can engage in different activities and interact with other objects in metaverse. Simulating the avatar requires accurate human pose estimation. Though camera-based solutions yield remarkable performance, they encounter the privacy issue and degraded performance caused by varying illumination, especially in smart home. In this paper, we propose a WiFi-based IoT-enabled human pose estimation scheme for metaverse avatar simulation, namely MetaFi. Specifically, a deep neural network is designed with customized convolutional layers and residual blocks to map the channel state information to human pose landmarks. It is enforced to learn the annotations from the accurate computer vision model, thus achieving cross-modal supervision. WiFi is ubiquitous and robust to illumination, making it a feasible solution for avatar applications in smart home. The experiments are conducted in the real world, and the results show that the MetaFi achieves very high performance with a PCK@50 of 95.23%.

Jianfei Yang, Han Zou, Lihua Xie · berkeley

Deep neural networks have empowered accurate device-free human activity recognition, which has wide applications. Deep models can extract robust features from various sensors and generalize well even in challenging situations such as data-insufficient cases. However, these systems could be vulnerable to input perturbations, i.e. adversarial attacks. We empirically demonstrate that both black-box Gaussian attacks and modern adversarial white-box attacks can render their accuracies to plummet. In this paper, we firstly point out that such phenomenon can bring severe safety hazards to device-free sensing systems, and then propose a novel learning framework, SecureSense, to defend common attacks. SecureSense aims to achieve consistent predictions regardless of whether there exists an attack on its input or not, alleviating the negative effect of distribution perturbation caused by adversarial attacks. Extensive experiments demonstrate that our proposed method can significantly enhance the model robustness of existing deep models, overcoming possible attacks. The results validate that our method works well on wireless human activity recognition and person identification systems. To the best of our knowledge, this is the first work to investigate adversarial attacks and further develop a novel defense framework for wireless human activity recognition in mobile computing research.

Yuecong Xu, Haozhi Cao, Zhenghua Chen et al.

Video analysis tasks such as action recognition have received increasing research interest with growing applications in fields such as smart healthcare, thanks to the introduction of large-scale datasets and deep learning-based representations. However, video models trained on existing datasets suffer from significant performance degradation when deployed directly to real-world applications due to domain shifts between the training public video datasets (source video domains) and real-world videos (target video domains). Further, with the high cost of video annotation, it is more practical to use unlabeled videos for training. To tackle performance degradation and address concerns in high video annotation cost uniformly, the video unsupervised domain adaptation (VUDA) is introduced to adapt video models from the labeled source domain to the unlabeled target domain by alleviating video domain shift, improving the generalizability and portability of video models. This paper surveys recent progress in VUDA with deep learning. We begin with the motivation of VUDA, followed by its definition, and recent progress of methods for both closed-set VUDA and VUDA under different scenarios, and current benchmark datasets for VUDA research. Eventually, future directions are provided to promote further VUDA research. The repository of this survey is provided at https://github.com/xuyu0010/awesome-video-domain-adaptation.

Yunjiao Zhou, Jianfei Yang, Han Zou et al. · berkeley

Recent achievements in language models have showcased their extraordinary capabilities in bridging visual information with semantic language understanding. This leads us to a novel question: can language models connect textual semantics with IoT sensory signals to perform recognition tasks, e.g., Human Activity Recognition (HAR)? If so, an intelligent HAR system with human-like cognition can be built, capable of adapting to new environments and unseen categories. This paper explores its feasibility with an innovative approach, IoT-sEnsors-language alignmEnt pre-Training (TENT), which jointly aligns textual embeddings with IoT sensor signals, including camera video, LiDAR, and mmWave. Through the IoT-language contrastive learning, we derive a unified semantic feature space that aligns multi-modal features with language embeddings, so that the IoT data corresponds to specific words that describe the IoT data. To enhance the connection between textual categories and their IoT data, we propose supplementary descriptions and learnable prompts that bring more semantic information into the joint feature space. TENT can not only recognize actions that have been seen but also ``guess'' the unseen action by the closest textual words from the feature space. We demonstrate TENT achieves state-of-the-art performance on zero-shot HAR tasks using different modalities, improving the best vision-language models by over 12%.

Haozhi Cao, Yuecong Xu, Jianfei Yang et al.

Multi-modal unsupervised domain adaptation (MM-UDA) for 3D semantic segmentation is a practical solution to embed semantic understanding in autonomous systems without expensive point-wise annotations. While previous MM-UDA methods can achieve overall improvement, they suffer from significant class-imbalanced performance, restricting their adoption in real applications. This imbalanced performance is mainly caused by: 1) self-training with imbalanced data and 2) the lack of pixel-wise 2D supervision signals. In this work, we propose Multi-modal Prior Aided (MoPA) domain adaptation to improve the performance of rare objects. Specifically, we develop Valid Ground-based Insertion (VGI) to rectify the imbalance supervision signals by inserting prior rare objects collected from the wild while avoiding introducing artificial artifacts that lead to trivial solutions. Meanwhile, our SAM consistency loss leverages the 2D prior semantic masks from SAM as pixel-wise supervision signals to encourage consistent predictions for each object in the semantic mask. The knowledge learned from modal-specific prior is then shared across modalities to achieve better rare object segmentation. Extensive experiments show that our method achieves state-of-the-art performance on the challenging MM-UDA benchmark. Code will be available at https://github.com/AronCao49/MoPA.

Dazhuo Wang, Jianfei Yang, Wei Cui et al.

WiFi-based smart human sensing technology enabled by Channel State Information (CSI) has received great attention in recent years. However, CSI-based sensing systems suffer from performance degradation when deployed in different environments. Existing works solve this problem by domain adaptation using massive unlabeled high-quality data from the new environment, which is usually unavailable in practice. In this paper, we propose a novel augmented environment-invariant robust WiFi gesture recognition system named AirFi that deals with the issue of environment dependency from a new perspective. The AirFi is a novel domain generalization framework that learns the critical part of CSI regardless of different environments and generalizes the model to unseen scenarios, which does not require collecting any data for adaptation to the new environment. AirFi extracts the common features from several training environment settings and minimizes the distribution differences among them. The feature is further augmented to be more robust to environments. Moreover, the system can be further improved by few-shot learning techniques. Compared to state-of-the-art methods, AirFi is able to work in different environment settings without acquiring any CSI data from the new environment. The experimental results demonstrate that our system remains robust in the new environment and outperforms the compared systems.

Yuecong Xu, Jianfei Yang, Haozhi Cao et al.

Video-based Unsupervised Domain Adaptation (VUDA) methods improve the robustness of video models, enabling them to be applied to action recognition tasks across different environments. However, these methods require constant access to source data during the adaptation process. Yet in many real-world applications, subjects and scenes in the source video domain should be irrelevant to those in the target video domain. With the increasing emphasis on data privacy, such methods that require source data access would raise serious privacy issues. Therefore, to cope with such concern, a more practical domain adaptation scenario is formulated as the Source-Free Video-based Domain Adaptation (SFVDA). Though there are a few methods for Source-Free Domain Adaptation (SFDA) on image data, these methods yield degenerating performance in SFVDA due to the multi-modality nature of videos, with the existence of additional temporal features. In this paper, we propose a novel Attentive Temporal Consistent Network (ATCoN) to address SFVDA by learning temporal consistency, guaranteed by two novel consistency objectives, namely feature consistency and source prediction consistency, performed across local temporal features. ATCoN further constructs effective overall temporal features by attending to local temporal features based on prediction confidence. Empirical results demonstrate the state-of-the-art performance of ATCoN across various cross-domain action recognition benchmarks.

Jianfei Yang, Xiangyu Peng, Kai Wang et al.

Domain Adaptation of Black-box Predictors (DABP) aims to learn a model on an unlabeled target domain supervised by a black-box predictor trained on a source domain. It does not require access to both the source-domain data and the predictor parameters, thus addressing the data privacy and portability issues of standard domain adaptation. Existing DABP approaches mostly rely on model distillation from the black-box predictor, \emph{i.e.}, training the model with its noisy target-domain predictions, which however inevitably introduces the confirmation bias accumulated from the prediction noises. To mitigate such bias, we propose a new method, named BETA, to incorporate knowledge distillation and noisy label learning into one coherent framework. This is enabled by a new divide-to-adapt strategy. BETA divides the target domain into an easy-to-adapt subdomain with less noise and a hard-to-adapt subdomain. Then it deploys mutually-teaching twin networks to filter the predictor errors for each other and improve them progressively, from the easy to hard subdomains. As such, BETA effectively purifies the noisy labels and reduces error accumulation. We theoretically show that the target error of BETA is minimized by decreasing the noise ratio of the subdomains. Extensive experiments demonstrate BETA outperforms existing methods on all DABP benchmarks, and is even comparable with the standard domain adaptation methods that use the source-domain data.

Kai Wang, Xiangyu Peng, Shuo Yang et al.

Learning with noisy labels has aroused much research interest since data annotations, especially for large-scale datasets, may be inevitably imperfect. Recent approaches resort to a semi-supervised learning problem by dividing training samples into clean and noisy sets. This paradigm, however, is prone to significant degeneration under heavy label noise, as the number of clean samples is too small for conventional methods to behave well. In this paper, we introduce a novel framework, termed as LC-Booster, to explicitly tackle learning under extreme noise. The core idea of LC-Booster is to incorporate label correction into the sample selection, so that more purified samples, through the reliable label correction, can be utilized for training, thereby alleviating the confirmation bias. Experiments show that LC-Booster advances state-of-the-art results on several noisy-label benchmarks, including CIFAR-10, CIFAR-100, Clothing1M and WebVision. Remarkably, under the extreme 90\% noise ratio, LC-Booster achieves 92.9\% and 48.4\% accuracy on CIFAR-10 and CIFAR-100, surpassing state-of-the-art methods by a large margin.

Haozhi Cao, Yuecong Xu, Jianfei Yang et al.

Continual Test-Time Adaptation (CTTA) generalizes conventional Test-Time Adaptation (TTA) by assuming that the target domain is dynamic over time rather than stationary. In this paper, we explore Multi-Modal Continual Test-Time Adaptation (MM-CTTA) as a new extension of CTTA for 3D semantic segmentation. The key to MM-CTTA is to adaptively attend to the reliable modality while avoiding catastrophic forgetting during continual domain shifts, which is out of the capability of previous TTA or CTTA methods. To fulfill this gap, we propose an MM-CTTA method called Continual Cross-Modal Adaptive Clustering (CoMAC) that addresses this task from two perspectives. On one hand, we propose an adaptive dual-stage mechanism to generate reliable cross-modal predictions by attending to the reliable modality based on the class-wise feature-centroid distance in the latent space. On the other hand, to perform test-time adaptation without catastrophic forgetting, we design class-wise momentum queues that capture confident target features for adaptation while stochastically restoring pseudo-source features to revisit source knowledge. We further introduce two new benchmarks to facilitate the exploration of MM-CTTA in the future. Our experimental results show that our method achieves state-of-the-art performance on both benchmarks.

Yuecong Xu, Jianfei Yang, Yunjiao Zhou et al.

For video models to be transferred and applied seamlessly across video tasks in varied environments, Video Unsupervised Domain Adaptation (VUDA) has been introduced to improve the robustness and transferability of video models. However, current VUDA methods rely on a vast amount of high-quality unlabeled target data, which may not be available in real-world cases. We thus consider a more realistic \textit{Few-Shot Video-based Domain Adaptation} (FSVDA) scenario where we adapt video models with only a few target video samples. While a few methods have touched upon Few-Shot Domain Adaptation (FSDA) in images and in FSVDA, they rely primarily on spatial augmentation for target domain expansion with alignment performed statistically at the instance level. However, videos contain more knowledge in terms of rich temporal and semantic information, which should be fully considered while augmenting target domains and performing alignment in FSVDA. We propose a novel SSA2lign to address FSVDA at the snippet level, where the target domain is expanded through a simple snippet-level augmentation followed by the attentive alignment of snippets both semantically and statistically, where semantic alignment of snippets is conducted through multiple perspectives. Empirical results demonstrate state-of-the-art performance of SSA2lign across multiple cross-domain action recognition benchmarks.

Yuecong Xu, Jianfei Yang, Haozhi Cao et al.

To enable video models to be applied seamlessly across video tasks in different environments, various Video Unsupervised Domain Adaptation (VUDA) methods have been proposed to improve the robustness and transferability of video models. Despite improvements made in model robustness, these VUDA methods require access to both source data and source model parameters for adaptation, raising serious data privacy and model portability issues. To cope with the above concerns, this paper firstly formulates Black-box Video Domain Adaptation (BVDA) as a more realistic yet challenging scenario where the source video model is provided only as a black-box predictor. While a few methods for Black-box Domain Adaptation (BDA) are proposed in image domain, these methods cannot apply to video domain since video modality has more complicated temporal features that are harder to align. To address BVDA, we propose a novel Endo and eXo-TEmporal Regularized Network (EXTERN) by applying mask-to-mix strategies and video-tailored regularizations: endo-temporal regularization and exo-temporal regularization, performed across both clip and temporal features, while distilling knowledge from the predictions obtained from the black-box predictor. Empirical results demonstrate the state-of-the-art performance of EXTERN across various cross-domain closed-set and partial-set action recognition benchmarks, which even surpassed most existing video domain adaptation methods with source data accessibility.

Jin Hao, Jiaxiang Liu, Jin Li et al.

A critical step in virtual dental treatment planning is to accurately delineate all tooth-bone structures from CBCT with high fidelity and accurate anatomical information. Previous studies have established several methods for CBCT segmentation using deep learning. However, the inherent resolution discrepancy of CBCT and the loss of occlusal and dentition information largely limited its clinical applicability. Here, we present a Deep Dental Multimodal Analysis (DDMA) framework consisting of a CBCT segmentation model, an intraoral scan (IOS) segmentation model (the most accurate digital dental model), and a fusion model to generate 3D fused crown-root-bone structures with high fidelity and accurate occlusal and dentition information. Our model was trained with a large-scale dataset with 503 CBCT and 28,559 IOS meshes manually annotated by experienced human experts. For CBCT segmentation, we use a five-fold cross validation test, each with 50 CBCT, and our model achieves an average Dice coefficient and IoU of 93.99% and 88.68%, respectively, significantly outperforming the baselines. For IOS segmentations, our model achieves an mIoU of 93.07% and 95.70% on the maxillary and mandible on a test set of 200 IOS meshes, which are 1.77% and 3.52% higher than the state-of-art method. Our DDMA framework takes about 20 to 25 minutes to generate the fused 3D mesh model following the sequential processing order, compared to over 5 hours by human experts. Notably, our framework has been incorporated into a software by a clear aligner manufacturer, and real-world clinical cases demonstrate that our model can visualize crown-root-bone structures during the entire orthodontic treatment and can predict risks like dehiscence and fenestration. These findings demonstrate the potential of multi-modal deep learning to improve the quality of digital dental models and help dentists make better clinical decisions.

Yunjiao Zhou, Jianfei Yang, He Huang et al.

WiFi-based pose estimation is a technology with great potential for the development of smart homes and metaverse avatar generation. However, current WiFi-based pose estimation methods are predominantly evaluated under controlled laboratory conditions with sophisticated vision models to acquire accurately labeled data. Furthermore, WiFi CSI is highly sensitive to environmental variables, and direct application of a pre-trained model to a new environment may yield suboptimal results due to domain shift. In this paper, we proposes a domain adaptation algorithm, AdaPose, designed specifically for weakly-supervised WiFi-based pose estimation. The proposed method aims to identify consistent human poses that are highly resistant to environmental dynamics. To achieve this goal, we introduce a Mapping Consistency Loss that aligns the domain discrepancy of source and target domains based on inner consistency between input and output at the mapping level. We conduct extensive experiments on domain adaptation in two different scenes using our self-collected pose estimation dataset containing WiFi CSI frames. The results demonstrate the effectiveness and robustness of AdaPose in eliminating domain shift, thereby facilitating the widespread application of WiFi-based pose estimation in smart cities.

Ruizhe Chen, Jianfei Yang, Huimin Xiong et al.

Recent discoveries have revealed that deep neural networks might behave in a biased manner in many real-world scenarios. For instance, deep networks trained on a large-scale face recognition dataset CelebA tend to predict blonde hair for females and black hair for males. Such biases not only jeopardize the robustness of models but also perpetuate and amplify social biases, which is especially concerning for automated decision-making processes in healthcare, recruitment, etc., as they could exacerbate unfair economic and social inequalities among different groups. Existing debiasing methods suffer from high costs in bias labeling or model re-training, while also exhibiting a deficiency in terms of elucidating the origins of biases within the model. To this respect, we propose a fast model debiasing framework (FMD) which offers an efficient approach to identify, evaluate and remove biases inherent in trained models. The FMD identifies biased attributes through an explicit counterfactual concept and quantifies the influence of data samples with influence functions. Moreover, we design a machine unlearning-based strategy to efficiently and effectively remove the bias in a trained model with a small counterfactual dataset. Experiments on the Colored MNIST, CelebA, and Adult Income datasets along with experiments with large language models demonstrate that our method achieves superior or competing accuracies compared with state-of-the-art methods while attaining significantly fewer biases and requiring much less debiasing cost. Notably, our method requires only a small external dataset and updating a minimal amount of model parameters, without the requirement of access to training data that may be too large or unavailable in practice.

Jindou Jia, Tuo An, Yuxuan Hu et al.

Imitation learning has become a cornerstone for solving complex robotic manipulation tasks. In particular, multimodality, which enables robots to capture diverse yet valid behavioral patterns, has driven the rapid emergence of generative policies as a dominant paradigm in robot learning. However, achieving such multimodality typically relies on stochastic noise initialization and iterative denoising procedures, resulting in substantial training complexity and low inference efficiency. Meanwhile, not all phases of a robotic task inherently require behavioral diversity. Motivated by this insight, we propose the Modality-Adaptive Robot Sampling (MARS) policy, which adaptively invokes tailored stochasticity only when it is truly beneficial, while reverting to an efficient deterministic learning during single-modal phases. In other words, the proper amount of noise is injected only at the proper time. By selectively activating multimodal generation, MARS policy bridges the gap between the multimodal capability of generative policies and the superior training and inference efficiency of deterministic models. Empirical studies across 8 simulated and 4 real-world tasks demonstrate that MARS exhibits robust multimodal expressivity and high efficiency, with a 16.67% success rate improvement and an 83.20% inference latency reduction in real-world tests. Counterintuitively, MARS also outpaces deterministic policies in training efficiency on near-deterministic tasks by more effectively modeling nuanced action diversity.

Geng Li, Guohao Chen, Ting Chen et al.

Vision-language models (VLMs) rely on long visual token sequences for visual understanding, making the prefill stage expensive in both computation and memory. Most existing pruning methods follow an absolute-ranking paradigm, assigning importance scores to visual tokens and retaining a fixed top-K subset. In this work, we argue that this paradigm is fundamentally brittle: attention sinks distort token importance rankings, while image redundancy and query-dependent visual evidence make fixed token budgets unreliable across inputs. We propose OccamToken, a training-free framework that replaces absolute token ranking with register-anchored relative evidence testing. Instead of asking which tokens are globally important, OccamToken evaluates whether a visual token provides information beyond a register-based reference. Our key insight is that register tokens naturally absorb low-information attention patterns, making them a stable reference for identifying genuinely informative visual evidence. Based on this principle, OccamToken performs both image-adaptive redundancy pruning and query-adaptive relevance pruning through dynamic thresholds derived from register attention. Across LLaVA-NeXT, LLaVA-v1.5, and Qwen3-VL, OccamToken consistently improves the accuracy-efficiency trade-off without additional training. Notably, on LLaVA-NeXT, it reduces 2,880 visual tokens to approximately 40 while preserving over 93% of the original accuracy, enabling stable visual token compression even in the extreme 1.4% retention regime.

Kuangji Zuo, Gen Li, Bofan Lyu et al.

Vision-Language-Action (VLA) models have recently shown strong potential for robot learning by following language instructions. However, in practice, language alone is often insufficient to precisely convey human intent. It is difficult to describe which exact object to interact with among similar candidates, where to act on the object, or how the target may change during execution. To address this limitation, we propose Gaze2Act, a novel VLA framework that leverages human gaze as a dynamic and intuitive intent signal for complex interactive manipulation. Gaze2Act first bridges the ego-exo view gap by mapping first-person gaze into the robot's perspective through cross-view semantic matching, producing both an object mask and a gaze point for coarse-to-fine target specification. These cues are then integrated into the policy through perception-level prompting and action-level conditioning, allowing the robot to attend to relevant regions and execute precise interactions under dynamic intent. In a systematic evaluation across seven task categories and 16 real-robot tasks on a Unitree G1 humanoid, Gaze2Act achieves state-of-the-art performance in both intent accuracy and task success rate. It notably outperforms baselines in object disambiguation, fine-grained interaction, and dynamic intent steering. These results demonstrate that human gaze provides a natural, low-burden, and highly expressive modality for human-in-the-loop VLA control.

Jindou Jia, Gen Li, Xiangyu Chen et al.

Diffusion-based policies have recently achieved remarkable success in robotics by formulating action prediction as a conditional denoising process. However, the standard practice of sampling from random Gaussian noise often requires multiple iterative steps to produce clean actions, leading to high inference latency that incurs a major bottleneck for real-time control. In this paper, we challenge the necessity of uninformed noise sampling and propose Action-to-Action flow matching (A2A), a novel policy paradigm that shifts from random sampling to initialization informed by the previous proprioceptive action. Unlike existing methods that treat proprioceptive action feedback as static conditions, A2A leverages historical proprioceptive sequences, embedding them into a high-dimensional latent space as the starting point for action generation. This design bypasses costly iterative denoising while effectively capturing the robot's physical dynamics and temporal continuity. Extensive experiments demonstrate that A2A exhibits high training efficiency, fast inference speed, and improved generalization. Notably, A2A enables high-quality action generation in as few as a single inference step, and exhibits superior robustness to visual perturbations and enhanced generalization to unseen configurations. Lastly, we also extend A2A to video generation, demonstrating its broader versatility in temporal modeling. Project site: https://lorenzo-0-0.github.io/A2A_Flow_Matching.

Xiyu Wang, Yuecong Xu, Kezhi Mao et al.

Assuming the source label space subsumes the target one, Partial Video Domain Adaptation (PVDA) is a more general and practical scenario for cross-domain video classification problems. The key challenge of PVDA is to mitigate the negative transfer caused by the source-only outlier classes. To tackle this challenge, a crucial step is to aggregate target predictions to assign class weights by up-weighing target classes and down-weighing outlier classes. However, the incorrect predictions of class weights can mislead the network and lead to negative transfer. Previous works improve the class weight accuracy by utilizing temporal features and attention mechanisms, but these methods may fall short when trying to generate accurate class weight when domain shifts are significant, as in most real-world scenarios. To deal with these challenges, we propose the Multi-modality Cluster-calibrated partial Adversarial Network (MCAN). MCAN enhances video feature extraction with multi-modal features from multiple temporal scales to form more robust overall features. It utilizes a novel class weight calibration method to alleviate the negative transfer caused by incorrect class weights. The calibration method tries to identify and weigh correct and incorrect predictions using distributional information implied by unsupervised clustering. Extensive experiments are conducted on prevailing PVDA benchmarks, and the proposed MCAN achieves significant improvements when compared to state-of-the-art PVDA methods.

Yucheng Wang, Yuecong Xu, Jianfei Yang et al.

Unsupervised Domain Adaptation (UDA) methods have been successful in reducing label dependency by minimizing the domain discrepancy between a labeled source domain and an unlabeled target domain. However, these methods face challenges when dealing with Multivariate Time-Series (MTS) data. MTS data typically consist of multiple sensors, each with its own unique distribution. This characteristic makes it hard to adapt existing UDA methods, which mainly focus on aligning global features while overlooking the distribution discrepancies at the sensor level, to reduce domain discrepancies for MTS data. To address this issue, a practical domain adaptation scenario is formulated as Multivariate Time-Series Unsupervised Domain Adaptation (MTS-UDA). In this paper, we propose SEnsor Alignment (SEA) for MTS-UDA, aiming to reduce domain discrepancy at both the local and global sensor levels. At the local sensor level, we design endo-feature alignment, which aligns sensor features and their correlations across domains. To reduce domain discrepancy at the global sensor level, we design exo-feature alignment that enforces restrictions on global sensor features. We further extend SEA to SEA++ by enhancing the endo-feature alignment. Particularly, we incorporate multi-graph-based high-order alignment for both sensor features and their correlations. Extensive empirical results have demonstrated the state-of-the-art performance of our SEA and SEA++ on public MTS datasets for MTS-UDA.

Xiyu Wang, Yuecong Xu, Jianfei Yang et al.

Continuous Video Domain Adaptation (CVDA) is a scenario where a source model is required to adapt to a series of individually available changing target domains continuously without source data or target supervision. It has wide applications, such as robotic vision and autonomous driving. The main underlying challenge of CVDA is to learn helpful information only from the unsupervised target data while avoiding forgetting previously learned knowledge catastrophically, which is out of the capability of previous Video-based Unsupervised Domain Adaptation methods. Therefore, we propose a Confidence-Attentive network with geneRalization enhanced self-knowledge disTillation (CART) to address the challenge in CVDA. Firstly, to learn from unsupervised domains, we propose to learn from pseudo labels. However, in continuous adaptation, prediction errors can accumulate rapidly in pseudo labels, and CART effectively tackles this problem with two key modules. Specifically, The first module generates refined pseudo labels using model predictions and deploys a novel attentive learning strategy. The second module compares the outputs of augmented data from the current model to the outputs of weakly augmented data from the source model, forming a novel consistency regularization on the model to alleviate the accumulation of prediction errors. Extensive experiments suggest that the CVDA performance of CART outperforms existing methods by a considerable margin.

Junfei Wang, He Huang, Jingze Feng et al.

There is a significant demand for indoor localization technology in smart buildings, and the most promising solution in this field is using RF sensors and fingerprinting-based methods that employ machine learning models trained on crowd-sourced user data gathered from IoT devices. However, this raises security and privacy issues in practice. Some researchers propose to use federated learning to partially overcome privacy problems, but there still remain security concerns, e.g., single-point failure and malicious attacks. In this paper, we propose a framework named DFLoc to achieve precise 3D localization tasks while considering the following two security concerns. Particularly, we design a specialized blockchain to decentralize the framework by distributing the tasks such as model distribution and aggregation which are handled by a central server to all clients in most previous works, to address the issue of the single-point failure for a reliable and accurate indoor localization system. Moreover, we introduce an updated model verification mechanism within the blockchain to alleviate the concern of malicious node attacks. Experimental results substantiate the framework's capacity to deliver accurate 3D location predictions and its superior resistance to the impacts of single-point failure and malicious attacks when compared to conventional centralized federated learning systems.

Zhongyuang Liu, Min He, Shaonan Yu et al.

Language-guided embodied navigation requires an agent to interpret object-referential instructions, search across multiple rooms, localize the referenced target, and execute reliable motion toward it. Existing systems remain limited in real indoor environments because narrow field-of-view sensing exposes only a partial local scene at each step, often forcing repeated rotations, delaying target discovery, and producing fragmented spatial understanding; meanwhile, directly prompting LLMs with dense 3D maps or exhaustive object lists quickly exceeds the context budget. We present OmniVLN, a zero-shot visual-language navigation framework that couples omnidirectional 3D perception with token-efficient hierarchical reasoning for both aerial and ground robots. OmniVLN fuses a rotating LiDAR and panoramic vision into a hardware-agnostic mapping stack, incrementally constructs a five-layer Dynamic Scene Graph (DSG) from mesh geometry to room- and building-level structure, and stabilizes high-level topology through persistent-homology-based room partitioning and hybrid geometric/VLM relation verification. For navigation, the global DSG is transformed into an agent-centric 3D octant representation with multi-resolution spatial attention prompting, enabling the LLM to progressively filter candidate rooms, infer egocentric orientation, localize target objects, and emit executable navigation primitives while preserving fine local detail and compact long-range memory. Experiments show that the proposed hierarchical interface improves spatial referring accuracy from 77.27\% to 93.18\%, reduces cumulative prompt tokens by up to 61.7\% in cluttered multi-room settings, and improves navigation success by up to 11.68\% over a flat-list baseline. We will release the code and an omnidirectional multimodal dataset to support reproducible research.

Ruizhe Chen, Yichen Li, Jianfei Yang et al.

Generating fair and accurate predictions plays a pivotal role in deploying large language models (LLMs) in the real world. However, existing debiasing methods inevitably generate unfair or incorrect predictions as they are designed and evaluated to achieve parity across different social groups but leave aside individual commonsense facts, resulting in modified knowledge that elicits unreasonable or undesired predictions. In this paper, we first establish a new bias mitigation benchmark, BiaScope, which systematically assesses performance by leveraging newly constructed datasets and metrics on knowledge retention and generalization. Then, we propose a novel debiasing approach, Fairness Stamp (FAST), which enables fine-grained calibration of individual social biases. FAST identifies the decisive layer responsible for storing social biases and then calibrates its outputs by integrating a small modular network, considering both bias mitigation and knowledge-preserving demands. Comprehensive experiments demonstrate that FAST surpasses state-of-the-art baselines with superior debiasing performance while not compromising the overall model capability for knowledge retention and downstream predictions. This highlights the potential of fine-grained debiasing strategies to achieve fairness in LLMs.

Bohan Hou, Jiuning Gu, Jiayan Guo et al.

Existing benchmarks for multimodal agentic search evaluate multimodal search and visual browsing, but visual evidence is either confined to the input or treated as an answer endpoint rather than part of an interleaved search trajectory. We introduce \textbf{InterLV-Search}, a benchmark for Interleaved Language-Vision Agentic Search, in which textual and visual evidence is repeatedly used to condition later search. It contains 2,061 examples across three levels: active visual evidence seeking, controlled offline interleaved multimodal search, and open-web interleaved multimodal search. Beyond existing benchmarks, it also includes multimodal multi-branch samples that involve comparison between multiple entities during the evidence search. We construct Level 1 and Level 2 with automated pipelines and Level 3 with a machine-led, human-supervised open-web pipeline. We further provide InterLV-Agent for standardized tool use, trajectory logging, and evaluation. Experiments on proprietary and open-source multimodal agents show that current systems remain far from solving interleaved multimodal search, with the best model below 50% overall accuracy, highlighting challenges in visual evidence seeking, search control, and multimodal evidence integration. We release the benchmark data and evaluation code at https://github.com/hbhalpha/InterLV-Search-Bench

Shenghai Yuan, Yizhuo Yang, Thien Hoang Nguyen et al.

In response to the evolving challenges posed by small unmanned aerial vehicles (UAVs), which possess the potential to transport harmful payloads or independently cause damage, we introduce MMAUD: a comprehensive Multi-Modal Anti-UAV Dataset. MMAUD addresses a critical gap in contemporary threat detection methodologies by focusing on drone detection, UAV-type classification, and trajectory estimation. MMAUD stands out by combining diverse sensory inputs, including stereo vision, various Lidars, Radars, and audio arrays. It offers a unique overhead aerial detection vital for addressing real-world scenarios with higher fidelity than datasets captured on specific vantage points using thermal and RGB. Additionally, MMAUD provides accurate Leica-generated ground truth data, enhancing credibility and enabling confident refinement of algorithms and models, which has never been seen in other datasets. Most existing works do not disclose their datasets, making MMAUD an invaluable resource for developing accurate and efficient solutions. Our proposed modalities are cost-effective and highly adaptable, allowing users to experiment and implement new UAV threat detection tools. Our dataset closely simulates real-world scenarios by incorporating ambient heavy machinery sounds. This approach enhances the dataset's applicability, capturing the exact challenges faced during proximate vehicular operations. It is expected that MMAUD can play a pivotal role in advancing UAV threat detection, classification, trajectory estimation capabilities, and beyond. Our dataset, codes, and designs will be available in https://github.com/ntu-aris/MMAUD.

Chenxi Jiang, Chuhao Zhou, Jianfei Yang

Robot task planning decomposes human instructions into executable action sequences that enable robots to complete a series of complex tasks. Although recent large language model (LLM)-based task planners achieve amazing performance, they assume that human instructions are clear and straightforward. However, real-world users are not experts, and their instructions to robots often contain significant vagueness. Linguists suggest that such vagueness frequently arises from referring expressions (REs), whose meanings depend heavily on dialogue context and environment. This vagueness is even more prevalent among the elderly and children, who are the groups that robots should serve more. This paper studies how such vagueness in REs within human instructions affects LLM-based robot task planning and how to overcome this issue. To this end, we propose the first robot task planning benchmark that systematically models vague REs grounded in pragmatic theory (REI-Bench), where we discover that the vagueness of REs can severely degrade robot planning performance, leading to success rate drops of up to 36.9%. We also observe that most failure cases stem from missing objects in planners. To mitigate the REs issue, we propose a simple yet effective approach: task-oriented context cognition, which generates clear instructions for robots, achieving state-of-the-art performance compared to aware prompts, chains of thought, and in-context learning. By tackling the overlooked issue of vagueness, this work contributes to the research community by advancing real-world task planning and making robots more accessible to non-expert users, e.g., the elderly and children.

Hanfang Liang, Yizhuo Yang, Jinming Hu et al.

Compact UAV systems, while advancing delivery and surveillance, pose significant security challenges due to their small size, which hinders detection by traditional methods. This paper presents a cost-effective, unsupervised UAV detection method using spatial-temporal sequence processing to fuse multiple LiDAR scans for accurate UAV tracking in real-world scenarios. Our approach segments point clouds into foreground and background, analyzes spatial-temporal data, and employs a scoring mechanism to enhance detection accuracy. Tested on a public dataset, our solution placed 4th in the CVPR 2024 UG2+ Challenge, demonstrating its practical effectiveness. We plan to open-source all designs, code, and sample data for the research community github.com/lianghanfang/UnLiDAR-UAV-Est.

Xiangyu Peng, Kai Wang, Jianfei Yang et al.

With the advance of deep learning, much progress has been made in building powerful artificial intelligence (AI) systems for automatic Chest X-ray (CXR) analysis. Most existing AI models are trained to be a binary classifier with the aim of distinguishing positive and negative cases. However, a large gap exists between the simple binary setting and complicated real-world medical scenarios. In this work, we reinvestigate the problem of automatic radiology diagnosis. We first observe that there is considerable diversity among cases within the positive class, which means simply classifying them as positive loses many important details. This motivates us to build AI models that can communicate fine-grained knowledge from medical images like human experts. To this end, we first propose a new benchmark on fine granularity learning from medical images. Specifically, we devise a division rule based on medical knowledge to divide positive cases into two subcategories, namely atypical positive and typical positive. Then, we propose a new metric termed AUC$^\text{FG}$ on the two subcategories for evaluation of the ability to separate them apart. With the proposed benchmark, we encourage the community to develop AI diagnosis systems that could better learn fine granularity from medical images. Last, we propose a simple risk modulation approach to this problem by only using coarse labels in training. Empirical results show that despite its simplicity, the proposed method achieves superior performance and thus serves as a strong baseline.

Tao Wu, Chuhao Zhou, Guangyu Zhao et al.

Embodied Question Answering (EQA) requires an agent to interpret language, perceive its environment, and navigate within 3D scenes to produce responses. Existing EQA benchmarks assume that every question must be answered, but embodied agents should know when they do not have sufficient information to answer. In this work, we focus on a minimal requirement for EQA agents, abstention: knowing when to withhold an answer. From an initial study of 500 human queries, we find that 32.4% contain missing or underspecified context. Drawing on this initial study and cognitive theories of human communication errors, we derive five representative categories requiring abstention: actionability limitation, referential underspecification, preference dependence, information unavailability, and false presupposition. We augment OpenEQA by having annotators transform well-posed questions into ambiguous variants outlined by these categories. The resulting dataset, AbstainEQA, comprises 1,636 annotated abstention cases paired with 1,636 original OpenEQA instances for balanced evaluation. Evaluating on AbstainEQA, we find that even the best frontier model only attains 42.79% abstention recall, while humans achieve 91.17%. We also find that scaling, prompting, and reasoning only yield marginal gains, and that fine-tuned models overfit to textual cues. Together, these results position abstention as a fundamental prerequisite for reliable interaction in embodied settings and as a necessary basis for effective clarification.

Tianchen Deng, Zhenxiang Xiong, Nailin Wang et al.

Visual Geometry Grounded Transformers (VGGT) have set new benchmarks in high-fidelity 3D scene reconstruction. However, as the sequence length increases, these models suffer from catastrophic geometric forgetting and accumulation drift, primarily due to the quadratic complexity of global attention which necessitates truncated temporal windows. To overcome the resulting geometric drift, we present Mamba-VGGT, an enhanced VGGT framework capable of persistent long-range reasoning. Our key contribution is a Sliding Window Mamba (SWM) memory module that maintains an explicit external memory token across temporal windows. This module leverages selective state-space modeling to distill and propagate global geometric priors, effectively bypassing the memory constraints of traditional transformers. To integrate these long-term temporal cues without disrupting the highly optimized spatial features of the pre-trained VGGT, we propose a Zero-Init Spatial Memory Injector. Utilizing zero-convolutional layers, this injector adaptively fuses persistent memory into the patch token stream, ensuring structural stability and seamless feature alignment. Extensive experiments demonstrate that our approach significantly outperforms existing VGGT-based methods in maintaining spatial consistency and reducing trajectory accumulation errors. Our work provides a scalable, linear-complexity solution for geometry-grounded world modeling in extensive 3D environments.

Jiaqi Bai, Jindou Jia, Yuxuan Hu et al.

Generative models such as diffusion and flow matching have become dominant paradigms for visuomotor policy learning, yet their reliance on iterative denoising incurs high inference latency incompatible with real-time robotic control. We present Fast Legendre-polynomial Action policy via Sparse History-anchored flow (FLASH Policy), which replaces discrete action-chunk generation with continuous Legendre polynomial trajectory representation. Specifically, by fitting expert demonstrations under sparse temporal sampling, FLASH enables a single inference to cover a significantly extended action horizon. To further accelerate generation, FLASH initiates the flow matching process from history polynomial coefficients rather than uninformative Gaussian noise, shortening the transport distance and enabling accurate single-step inference. Moreover, analytic polynomial differentiation directly provides desired velocity feed-forward signals to the torque controller without numerical approximation. Extensive experiments on five simulated and two real-world manipulation tasks demonstrate that FLASH achieves state-of-the-art success rates ($\ge 92\%$ across all tasks), a per-episode inference time of $31.40\,ms$ (up to $175\times$ faster than diffusion policies and $18\times$ faster than prior flow matching policies), up to $4\times$ faster training convergence than ACT, and $5\times$ to $7\times$ reduction in controller tracking error compared to discrete-action baselines.

Guohao Chen, Shuaicheng Niu, Geng Li et al.

Test-time model evolution offers a promising way for deployed models to improve from unlabeled test-time experience, yet most existing methods depend on backpropagation (BP), which incurs substantial memory overhead and makes them difficult to deploy on edge devices, quantized models, specialized accelerators, or black-box models. In this work, we study test-time model evolution under a strict two-forward budget, a setting that pushes adaptation toward highly efficient real-world deployment. We reveal three key obstacles in zeroth-order test-time optimization: susceptibility to shortcut solutions, uncontrolled weight drift, and ineffective update direction estimation. To overcome them, we propose EVA-0, a minimal zeroth-order adaptation framework that: 1) keeps the loss scale-invariant to prevent shortcut solutions; 2) devises an anchor-guided optimization strategy to alleviate weight drift; 3) uses sample-wise symmetric two-sided perturbation for update direction estimation and inference. EVA-0 requires no BP and performs both inference and adaptation within only two forward passes per sample. Results on ImageNet-C & ViT-Base show that EVA-0 outperforms both BP-based DeYO and BP-free FOA, while achieving a 14x speed-up over FOA. Code will be released.

Tuo An, Jindou Jia, Gen Li et al.

World models aim to improve robotic decision making by predicting the consequences of actions. However, in practice, their predictions often become unreliable once the robot encounters states outside the training distribution, limiting their effectiveness at deployment. We observe that execution itself provides a natural but underutilized signal: after each action, the robot directly observes the true next state, revealing the mismatch between predicted and actual outcomes. Building on this insight, we propose feedback world model, a new paradigm that closes the loop between prediction and observation at inference time. Instead of treating the world model as a static open-loop predictor, our method maintains a lightweight feedback state that is updated online to iteratively correct future predictions, compensating for model errors using real-time observations without additional training data or parameter updates. We show that this process can be interpreted as a latent-space observer and admits convergence guarantees under mild conditions. We further introduce action-aware guidance to better translate corrected predictions into control by emphasizing action-controllable components while suppressing irrelevant variations. Experiments on LIBERO-Plus, Robomimic, and real-world manipulation tasks demonstrate that our method substantially improves both prediction accuracy and policy performance under distribution shift. In particular, it reduces world model prediction error by up to 76.4% and improves out-of-distribution (OOD) success rate by 30%. These results show that incorporating real-time feedback at inference time provides a simple yet powerful alternative to static world modeling.

Jiandong Shao, Yao Lu, Jianfei Yang · mila

Large Language Models (LLMs) exhibit impressive performance on complex reasoning tasks, yet they frequently fail on basic numerical problems, producing incorrect outputs. Inspired by Benford's Law -- a statistical pattern where lower digits occur more frequently as leading digits -- we hypothesize that the long-tailed digit distributions in web-collected corpora may be learned by LLMs during pretraining, leading to biased numerical generation. To investigate the hypothesis, we first examine whether digits frequencies in pretraining corpus (OLMo2) follows Benford's law. We then construct an evaluation benchmark with uniformly distributed ground-truth digits across seven numerical reasoning tasks. Our evaluation results demonstrate that leading open-source LLMs show a consistent pattern of digit bias that resembles Benford's law. Through logit-lens tracing and neuron-level dissection, we identify that this bias arises predominantly from a small subset of highly digit-selective feed-forward network (FFN) neurons in the deeper layers. Finally, we demonstrate that pruning these neurons mitigates imbalanced overgeneration and partially corrects erroneous outputs, providing causal evidence that fine-grained pretraining digit bias can propagate into model behavior. Our findings reveal a fundamental connection between corpus-level statistics and symbolic failure modes in LLMs, offering a new lens for diagnosing and mitigating hallucinations in numerical tasks.

Haichao Liu, Yuheng Zhou, Zhenyu Wu et al.

Embodied Artificial Intelligence (EAI) is rapidly developing, gradually subverting previous autonomous systems' paradigms from isolated perception to integrated, continuous action. This transition is highly significant for industrial robotic manipulation, promising to free human workers from repetitive, dangerous daily labor. To benchmark and advance this capability, we introduce the Robotic Collaborative Assembly Assistance (RoCo) Challenge with a dataset towards simulation and real-world assembly manipulation. Set against the backdrop of human-centered manufacturing, this challenge focuses on a high-precision planetary gearbox assembly task, a demanding yet highly representative operation in modern industry. Built upon a self-developed data collection, training, and evaluation system in Isaac Sim, and utilizing a dual-arm robot for real-world deployment, the challenge operates in two phases. The Simulation Round defines fine-grained task phases for step-wise scoring to handle the long-horizon nature of the assembly. The Real-World Round mirrors this evaluation with physical gearbox components and high-quality teleoperated datasets. The core tasks require assembling an epicyclic gearbox from scratch, including mounting three planet gears, a sun gear, and a ring gear. Attracting over 60 teams and 170+ participants from more than 10 countries, the challenge yielded highly effective solutions, most notably ARC-VLA and RoboCola. Results demonstrate that a dual-model framework for long-horizon multi-task learning is highly effective, and the strategic utilization of recovery-from-failure curriculum data is a critical insight for successful deployment. This report outlines the competition setup, evaluation approach, key findings, and future directions for industrial EAI. Our dataset, CAD files, code, and evaluation results can be found at: https://rocochallenge.github.io/RoCo2026/.

Fuyu Dong, Ke Li, Di Wang et al.

Change detection visual question answering (CDVQA) requires answering text queries by reasoning about semantic changes in bi-temporal remote sensing images. A straightforward approach is to boost CDVQA performance with generic vision-language models via supervised fine-tuning (SFT). Despite recent progress, we observe that a significant portion of failures do not stem from clearly incorrect predictions, but from decision ambiguity, where the model assigns similar confidence to the correct answer and strong distractors. To formalize this challenge, we define Decision-Ambiguous Samples (DAS) as instances with a small probability margin between the ground-truth answer and the most competitive alternative. We argue that explicitly optimizing DAS is crucial for improving the discriminability and robustness of CDVQA models. To this end, we propose DARFT, a Decision-Ambiguity-guided Reinforcement Fine-Tuning framework that first mines DAS using an SFT-trained reference policy and then applies group-relative policy optimization on the mined subset. By leveraging multi-sample decoding and intra-group relative advantages, DARFT suppresses strong distractors and sharpens decision boundaries without additional supervision. Extensive experiments demonstrate consistent gains over SFT baselines, particularly under few-shot settings.

Xinying Guo, Chenxi Jiang, Hyun Bin Kim et al.

Robotic manipulation often requires memory: occlusion and state changes can make decision-time observations perceptually aliased, making action selection non-Markovian at the observation level because the same observation may arise from different interaction histories. Most embodied agents implement memory via semantically compressed traces and similarity-based retrieval, which discards disambiguating fine-grained perceptual cues and can return perceptually similar but decision-irrelevant episodes. Inspired by human episodic memory, we propose Chameleon, which writes geometry-grounded multimodal tokens to preserve disambiguating context and produces goal-directed recall through a differentiable memory stack. We also introduce Camo-Dataset, a real-robot UR5e dataset spanning episodic recall, spatial tracking, and sequential manipulation under perceptual aliasing. Across tasks, Chameleon consistently improves decision reliability and long-horizon control over strong baselines in perceptually confusable settings.

Xiaojiang Peng, Kai Wang, Zhaoyang Zeng et al.

Deep networks achieve excellent results on large-scale clean data but degrade significantly when learning from noisy labels. To suppressing the impact of mislabeled data, this paper proposes a conceptually simple yet efficient training block, termed as Attentive Feature Mixup (AFM), which allows paying more attention to clean samples and less to mislabeled ones via sample interactions in small groups. Specifically, this plug-and-play AFM first leverages a \textit{group-to-attend} module to construct groups and assign attention weights for group-wise samples, and then uses a \textit{mixup} module with the attention weights to interpolate massive noisy-suppressed samples. The AFM has several appealing benefits for noise-robust deep learning. (i) It does not rely on any assumptions and extra clean subset. (ii) With massive interpolations, the ratio of useless samples is reduced dramatically compared to the original noisy ratio. (iii) \pxj{It jointly optimizes the interpolation weights with classifiers, suppressing the influence of mislabeled data via low attention weights. (iv) It partially inherits the vicinal risk minimization of mixup to alleviate over-fitting while improves it by sampling fewer feature-target vectors around mislabeled data from the mixup vicinal distribution.} Extensive experiments demonstrate that AFM yields state-of-the-art results on two challenging real-world noisy datasets: Food101N and Clothing1M. The code will be available at https://github.com/kaiwang960112/AFM.

Kai Wang, Xiaojiang Peng, Jianfei Yang et al.

Annotating a qualitative large-scale facial expression dataset is extremely difficult due to the uncertainties caused by ambiguous facial expressions, low-quality facial images, and the subjectiveness of annotators. These uncertainties lead to a key challenge of large-scale Facial Expression Recognition (FER) in deep learning era. To address this problem, this paper proposes a simple yet efficient Self-Cure Network (SCN) which suppresses the uncertainties efficiently and prevents deep networks from over-fitting uncertain facial images. Specifically, SCN suppresses the uncertainty from two different aspects: 1) a self-attention mechanism over mini-batch to weight each training sample with a ranking regularization, and 2) a careful relabeling mechanism to modify the labels of these samples in the lowest-ranked group. Experiments on synthetic FER datasets and our collected WebEmotion dataset validate the effectiveness of our method. Results on public benchmarks demonstrate that our SCN outperforms current state-of-the-art methods with \textbf{88.14}\% on RAF-DB, \textbf{60.23}\% on AffectNet, and \textbf{89.35}\% on FERPlus. The code will be available at \href{https://github.com/kaiwang960112/Self-Cure-Network}{https://github.com/kaiwang960112/Self-Cure-Network}.

Bohan Hou, Gen Li, Jindou Jia et al.

World models, which are predictive representations of how environments evolve under actions, have become a central component of robot learning. They support policy learning, planning, simulation, evaluation, data generation, and have advanced rapidly with the rise of foundation models and large-scale video generation. However, the literature remains fragmented across architectures, functional roles, and embodied application domains. To address this gap, we present a comprehensive review of world models from a robot-learning perspective. We examine how world models are coupled with robot policies, how they serve as learned simulators for reinforcement learning and evaluation, and how robotic video world models have progressed from imagination-based generation to controllable, structured, and foundation-scale formulations. We further connect these ideas to navigation and autonomous driving, and summarize representative datasets, benchmarks, and evaluation protocols. Overall, this survey systematically reviews the rapidly growing literature on world models for robot learning, clarifies key paradigms and applications, and highlights major challenges and future directions for predictive modeling in embodied agents. To facilitate continued access to newly emerging works, benchmarks, and resources, we will maintain and regularly update the accompanying GitHub repository alongside this survey.

Xichen Yuan, Zhe Li, Bofan Lyu et al.

While generative models have become effective at producing human-like motions from text, transferring these motions to humanoid robots for physical execution remains challenging. Existing pipelines are often limited by retargeting, where kinematic quality is undermined by physical infeasibility, contact-transition errors, and the high cost of real-world dynamical data. We present a unified latent-driven framework that bridges natural language and whole-body humanoid locomotion through a retarget-free, physics-optimized pipeline. Rather than treating generation and control as separate stages, our key insight is to couple them bidirectionally under physical constraints.We introduce a Physical Plausibility Optimization (PP-Opt) module as the coupling interface. In the forward direction, PP-Opt refines a teacher-student distillation policy with a plausibility-centric reward to suppress artifacts such as floating, skating, and penetration. In the backward direction, it converts reward-optimized simulation rollouts into high-quality explicit motion data, which is used to fine-tune the motion generator toward a more physically plausible latent distribution. This bidirectional design forms a self-improving cycle: the generator learns a physically grounded latent space, while the controller learns to execute latent-conditioned behaviors with dynamical integrity.Extensive experiments on the Unitree G1 humanoid show that our bidirectional optimization improves tracking accuracy and success rates. Across IsaacLab and MuJoCo, the implicit latent-driven pipeline consistently outperforms conventional explicit retargeting baselines in both precision and stability. By coupling diffusion-based motion generation with physical plausibility optimization, our framework provides a practical path toward deployable text-guided humanoid intelligence.