Yifan Zhang

Yifan Zhang, Qijian Zhang, Zhiyu Zhu et al.

The inherent ambiguity in ground-truth annotations of 3D bounding boxes, caused by occlusions, signal missing, or manual annotation errors, can confuse deep 3D object detectors during training, thus deteriorating detection accuracy. However, existing methods overlook such issues to some extent and treat the labels as deterministic. In this paper, we formulate the label uncertainty problem as the diversity of potentially plausible bounding boxes of objects. Then, we propose GLENet, a generative framework adapted from conditional variational autoencoders, to model the one-to-many relationship between a typical 3D object and its potential ground-truth bounding boxes with latent variables. The label uncertainty generated by GLENet is a plug-and-play module and can be conveniently integrated into existing deep 3D detectors to build probabilistic detectors and supervise the learning of the localization uncertainty. Besides, we propose an uncertainty-aware quality estimator architecture in probabilistic detectors to guide the training of the IoU-branch with predicted localization uncertainty. We incorporate the proposed methods into various popular base 3D detectors and demonstrate significant and consistent performance gains on both KITTI and Waymo benchmark datasets. Especially, the proposed GLENet-VR outperforms all published LiDAR-based approaches by a large margin and achieves the top rank among single-modal methods on the challenging KITTI test set. The source code and pre-trained models are publicly available at \url{https://github.com/Eaphan/GLENet}.

Yifan Zhang, Daquan Zhou, Bryan Hooi et al.

The power of DNNs relies heavily on the quantity and quality of training data. However, collecting and annotating data on a large scale is often expensive and time-consuming. To address this issue, we explore a new task, termed dataset expansion, aimed at expanding a ready-to-use small dataset by automatically creating new labeled samples. To this end, we present a Guided Imagination Framework (GIF) that leverages cutting-edge generative models like DALL-E2 and Stable Diffusion (SD) to "imagine" and create informative new data from the input seed data. Specifically, GIF conducts data imagination by optimizing the latent features of the seed data in the semantically meaningful space of the prior model, resulting in the creation of photo-realistic images with new content. To guide the imagination towards creating informative samples for model training, we introduce two key criteria, i.e., class-maintained information boosting and sample diversity promotion. These criteria are verified to be essential for effective dataset expansion: GIF-SD obtains 13.5% higher model accuracy on natural image datasets than unguided expansion with SD. With these essential criteria, GIF successfully expands small datasets in various scenarios, boosting model accuracy by 36.9% on average over six natural image datasets and by 13.5% on average over three medical datasets. The source code is available at https://github.com/Vanint/DatasetExpansion.

Yifan Zhang, Zhiyu Zhu, Junhui Hou et al.

The Detection Transformer (DETR) has revolutionized the design of CNN-based object detection systems, showcasing impressive performance. However, its potential in the domain of multi-frame 3D object detection remains largely unexplored. In this paper, we present STEMD, a novel end-to-end framework that enhances the DETR-like paradigm for multi-frame 3D object detection by addressing three key aspects specifically tailored for this task. First, to model the inter-object spatial interaction and complex temporal dependencies, we introduce the spatial-temporal graph attention network, which represents queries as nodes in a graph and enables effective modeling of object interactions within a social context. To solve the problem of missing hard cases in the proposed output of the encoder in the current frame, we incorporate the output of the previous frame to initialize the query input of the decoder. Finally, it poses a challenge for the network to distinguish between the positive query and other highly similar queries that are not the best match. And similar queries are insufficiently suppressed and turn into redundant prediction boxes. To address this issue, our proposed IoU regularization term encourages similar queries to be distinct during the refinement. Through extensive experiments, we demonstrate the effectiveness of our approach in handling challenging scenarios, while incurring only a minor additional computational overhead. The code is publicly available at https://github.com/Eaphan/STEMD.

Zhiquan Tan, Yifan Zhang, Jingqin Yang et al.

Contrastive learning is a powerful self-supervised learning method, but we have a limited theoretical understanding of how it works and why it works. In this paper, we prove that contrastive learning with the standard InfoNCE loss is equivalent to spectral clustering on the similarity graph. Using this equivalence as the building block, we extend our analysis to the CLIP model and rigorously characterize how similar multi-modal objects are embedded together. Motivated by our theoretical insights, we introduce the Kernel-InfoNCE loss, incorporating mixtures of kernel functions that outperform the standard Gaussian kernel on several vision datasets. The code is available at https://github.com/yifanzhang-pro/Kernel-InfoNCE.

Weihan Cao, Yifan Zhang, Jianfei Gao et al.

Knowledge distillation(KD) is a widely-used technique to train compact models in object detection. However, there is still a lack of study on how to distill between heterogeneous detectors. In this paper, we empirically find that better FPN features from a heterogeneous teacher detector can help the student although their detection heads and label assignments are different. However, directly aligning the feature maps to distill detectors suffers from two problems. First, the difference in feature magnitude between the teacher and the student could enforce overly strict constraints on the student. Second, the FPN stages and channels with large feature magnitude from the teacher model could dominate the gradient of distillation loss, which will overwhelm the effects of other features in KD and introduce much noise. To address the above issues, we propose to imitate features with Pearson Correlation Coefficient to focus on the relational information from the teacher and relax constraints on the magnitude of the features. Our method consistently outperforms the existing detection KD methods and works for both homogeneous and heterogeneous student-teacher pairs. Furthermore, it converges faster. With a powerful MaskRCNN-Swin detector as the teacher, ResNet-50 based RetinaNet and FCOS achieve 41.5% and 43.9% mAP on COCO2017, which are 4.1\% and 4.8\% higher than the baseline, respectively.

Yaqi Liu, Chao Xia, Song Xiao et al.

Copy-move forgery detection aims at detecting duplicated regions in a suspected forged image, and deep learning based copy-move forgery detection methods are in the ascendant. These deep learning based methods heavily rely on synthetic training data, and the performance will degrade when facing new tasks. In this paper, we propose a Transformer-style copy-move forgery detection network named as CMFDFormer, and provide a novel PCSD (Pooled Cube and Strip Distillation) continual learning framework to help CMFDFormer handle new tasks. CMFDFormer consists of a MiT (Mix Transformer) backbone network and a PHD (Pluggable Hybrid Decoder) mask prediction network. The MiT backbone network is a Transformer-style network which is adopted on the basis of comprehensive analyses with CNN-style and MLP-style backbones. The PHD network is constructed based on self-correlation computation, hierarchical feature integration, a multi-scale cycle fully-connected block and a mask reconstruction block. The PHD network is applicable to feature extractors of different styles for hierarchical multi-scale information extraction, achieving comparable performance. Last but not least, we propose a PCSD continual learning framework to improve the forgery detectability and avoid catastrophic forgetting when handling new tasks. Our continual learning framework restricts intermediate features from the PHD network, and takes advantage of both cube pooling and strip pooling. Extensive experiments on publicly available datasets demonstrate the good performance of CMFDFormer and the effectiveness of the PCSD continual learning framework.

Yifan Zhang, Liang Hu, Haofeng Sun et al.

Despite recent progress in multimodal agentic systems, existing approaches often treat image manipulation and web search as disjoint capabilities, rely heavily on costly reinforcement learning, and lack planning grounded in real tool-execution traces. To address these limitations, we present Skywork-R1V4, a 30B (A3B) parameter multimodal agentic model that unifies multimodal planning, active image manipulation ("thinking with images"), deep multimodal search, and, most critically, interleaved reasoning that dynamically alternates between visual operations and external knowledge retrieval. Trained solely via supervised fine-tuning on fewer than 30,000 high-quality, planning-execution-consistent trajectories and validated through stepwise consistency filtering, Skywork-R1V4 achieves state-of-the-art results across perception and multimodal search benchmarks: it scores 66.1 on MMSearch and 67.2 on FVQA, surpassing Gemini 2.5 Flash on all 11 metrics. Skywork-R1V4 exhibits emergent long-horizon reasoning at inference time, successfully orchestrating more than 10 tool calls to solve complex, multi-step tasks. Our results demonstrate that sophisticated agentic multimodal intelligence can be achieved through carefully curated supervised learning alone, without any reliance on reinforcement learning.

Yifan Zhang, Wenyu Du, Dongming Jin et al.

Chain-of-thought (CoT) significantly enhances the performance of large language models (LLMs) across a wide range of tasks, and prior research shows that CoT can theoretically increase expressiveness. However, there is limited mechanistic understanding of the algorithms that Transformer+CoT can learn. Our key contributions are: (1) We evaluate the state tracking capabilities of Transformer+CoT and its variants, confirming the effectiveness of CoT. (2) Next, we identify the circuit (a subset of model components, responsible for tracking the world state), indicating that late-layer MLP neurons play a key role. We propose two metrics, compression and distinction, and show that the neuron sets for each state achieve nearly 100% accuracy, providing evidence of an implicit finite state automaton (FSA) embedded within the model. (3) Additionally, we explore three challenging settings: skipping intermediate steps, introducing data noises, and testing length generalization. Our results demonstrate that Transformer+CoT learns robust algorithms (FSAs), highlighting its resilience in challenging scenarios. Our code is available at https://github.com/IvanChangPKU/FSA.

Yifan Zhang

The quadratic complexity of dot-product attention introduced in Transformer remains a fundamental bottleneck impeding the progress of foundation models toward unbounded context lengths. Addressing this challenge, we introduce the Deep Tensor Network, a new architectural framework that fundamentally reformulates attention by unifying the expressive power of tensor algebra with neural network design. Our approach moves beyond both conventional dot-product attention and subsequent linear-time approximations to capture higher-order statistical dependencies. We introduce two core operators derived from this framework: \emph{Tensor Attention}, which models complex token-mixing via data-dependent polynomial kernels, and Tensor Interaction, a novel mechanism for adaptive channel-mixing. We demonstrate that these operators are powered by second-order summaries that entirely bypass the formation of $n \times n$ matrices, enabling a causality-preserving streaming implementation with $O(d^2)$ per-token updates and $O(d^2)$ state. This efficiency rivals that of modern State Space Models while retaining an attention-like formulation. The Deep Tensor Network thus provides a principled and powerful new class of building blocks for next-generation sequence models, bridging the gap between scalable computation and rich, expressive interaction modeling.

Bo Pang, Yizhuo Li, Yifan Zhang et al.

Multi-object tracking is a fundamental vision problem that has been studied for a long time. As deep learning brings excellent performances to object detection algorithms, Tracking by Detection (TBD) has become the mainstream tracking framework. Despite the success of TBD, this two-step method is too complicated to train in an end-to-end manner and induces many challenges as well, such as insufficient exploration of video spatial-temporal information, vulnerability when facing object occlusion, and excessive reliance on detection results. To address these challenges, we propose a concise end-to-end model TubeTK which only needs one step training by introducing the ``bounding-tube" to indicate temporal-spatial locations of objects in a short video clip. TubeTK provides a novel direction of multi-object tracking, and we demonstrate its potential to solve the above challenges without bells and whistles. We analyze the performance of TubeTK on several MOT benchmarks and provide empirical evidence to show that TubeTK has the ability to overcome occlusions to some extent without any ancillary technologies like Re-ID. Compared with other methods that adopt private detection results, our one-stage end-to-end model achieves state-of-the-art performances even if it adopts no ready-made detection results. We hope that the proposed TubeTK model can serve as a simple but strong alternative for video-based MOT task. The code and models are available at https://github.com/BoPang1996/TubeTK.

Shaoqi Dong, Chaoyou Fu, Haihan Gao et al.

Vision-Language Action (VLA) models significantly advance robotic manipulation by leveraging the strong perception capabilities of pretrained vision-language models (VLMs). By integrating action modules into these pretrained models, VLA methods exhibit improved generalization. However, training them from scratch is costly. In this work, we propose a simple yet effective distillation-based framework that equips VLMs with action-execution capability by transferring knowledge from pretrained small action models. Our architecture retains the original VLM structure, adding only an action token and a state encoder to incorporate physical inputs. To distill action knowledge, we adopt a two-stage training strategy. First, we perform lightweight alignment by mapping VLM hidden states into the action space of the small action model, enabling effective reuse of its pretrained action decoder and avoiding expensive pretraining. Second, we selectively fine-tune the language model, state encoder, and action modules, enabling the system to integrate multimodal inputs with precise action generation. Specifically, the action token provides the VLM with a direct handle for predicting future actions, while the state encoder allows the model to incorporate robot dynamics not captured by vision alone. This design yields substantial efficiency gains over training large VLA models from scratch. Compared with previous state-of-the-art methods, our method achieves 97.3% average success rate on LIBERO (11.8% improvement) and 93.5% on LIBERO-LONG (24.5% improvement). In real-world experiments across five manipulation tasks, our method consistently outperforms the teacher model, achieving 82.0% success rate (17% improvement), which demonstrate that action distillation effectively enables VLMs to generate precise actions while substantially reducing training costs.

Shengda Gu, Kai Li, Junliang Xing et al.

Combinatorial optimization problems are notoriously challenging due to their discrete structure and exponentially large solution space. Recent advances in deep reinforcement learning (DRL) have enabled the learning heuristics directly from data. However, DRL methods often suffer from limited exploration and susceptibility to local optima. On the other hand, evolutionary algorithms such as Genetic Algorithms (GAs) exhibit strong global exploration capabilities but are typically sample inefficient and computationally intensive. In this work, we propose the Evolutionary Augmentation Mechanism (EAM), a general and plug-and-play framework that synergizes the learning efficiency of DRL with the global search power of GAs. EAM operates by generating solutions from a learned policy and refining them through domain-specific genetic operations such as crossover and mutation. These evolved solutions are then selectively reinjected into the policy training loop, thereby enhancing exploration and accelerating convergence. We further provide a theoretical analysis that establishes an upper bound on the KL divergence between the evolved solution distribution and the policy distribution, ensuring stable and effective policy updates. EAM is model-agnostic and can be seamlessly integrated with state-of-the-art DRL solvers such as the Attention Model, POMO, and SymNCO. Extensive results on benchmark problems (e.g., TSP, CVRP, PCTSP, and OP) demonstrate that EAM significantly improves both solution quality and training efficiency over competitive baselines.

Won Seok Jang, Hieu Tran, Manav Mistry et al.

Patients must possess the knowledge necessary to actively participate in their care. We present NoteAid-Chatbot, a conversational AI that promotes patient understanding via a novel 'learning as conversation' framework, built on a multi-agent large language model (LLM) and reinforcement learning (RL) setup without human-labeled data. NoteAid-Chatbot was built on a lightweight LLaMA 3.2 3B model trained in two stages: initial supervised fine-tuning on conversational data synthetically generated using medical conversation strategies, followed by RL with rewards derived from patient understanding assessments in simulated hospital discharge scenarios. Our evaluation, which includes comprehensive human-aligned assessments and case studies, demonstrates that NoteAid-Chatbot exhibits key emergent behaviors critical for patient education, such as clarity, relevance, and structured dialogue, even though it received no explicit supervision for these attributes. Our results show that even simple Proximal Policy Optimization (PPO)-based reward modeling can successfully train lightweight, domain-specific chatbots to handle multi-turn interactions, incorporate diverse educational strategies, and meet nuanced communication objectives. Our Turing test demonstrates that NoteAid-Chatbot surpasses non-expert human. Although our current focus is on healthcare, the framework we present illustrates the feasibility and promise of applying low-cost, PPO-based RL to realistic, open-ended conversational domains, broadening the applicability of RL-based alignment methods.

Lei Shi, Yifan Zhang, Jian Cheng et al.

Existing methods for skeleton-based action recognition mainly focus on improving the recognition accuracy, whereas the efficiency of the model is rarely considered. Recently, there are some works trying to speed up the skeleton modeling by designing light-weight modules. However, in addition to the model size, the amount of the data involved in the calculation is also an important factor for the running speed, especially for the skeleton data where most of the joints are redundant or non-informative to identify a specific skeleton. Besides, previous works usually employ one fix-sized model for all the samples regardless of the difficulty of recognition, which wastes computations for easy samples. To address these limitations, a novel approach, called AdaSGN, is proposed in this paper, which can reduce the computational cost of the inference process by adaptively controlling the input number of the joints of the skeleton on-the-fly. Moreover, it can also adaptively select the optimal model size for each sample to achieve a better trade-off between accuracy and efficiency. We conduct extensive experiments on three challenging datasets, namely, NTU-60, NTU-120 and SHREC, to verify the superiority of the proposed approach, where AdaSGN achieves comparable or even higher performance with much lower GFLOPs compared with the baseline method.

Congqi Cao, Yue Lu, Yifan Zhang et al.

Contextual information plays an important role in action recognition. Local operations have difficulty to model the relation between two elements with a long-distance interval. However, directly modeling the contextual information between any two points brings huge cost in computation and memory, especially for action recognition, where there is an additional temporal dimension. Inspired from 2D criss-cross attention used in segmentation task, we propose a recurrent 3D criss-cross attention (RCCA-3D) module to model the dense long-range spatiotemporal contextual information in video for action recognition. The global context is factorized into sparse relation maps. We model the relationship between points in the same line along the direction of horizon, vertical and depth at each time, which forms a 3D criss-cross structure, and duplicate the same operation with recurrent mechanism to transmit the relation between points in a line to a plane finally to the whole spatiotemporal space. Compared with the non-local method, the proposed RCCA-3D module reduces the number of parameters and FLOPs by 25% and 30% for video context modeling. We evaluate the performance of RCCA-3D with two latest action recognition networks on three datasets and make a thorough analysis of the architecture, obtaining the optimal way to factorize and fuse the relation maps. Comparisons with other state-of-the-art methods demonstrate the effectiveness and efficiency of our model.

Lei Shi, Yifan Zhang, Jian Cheng et al.

Dynamic skeletal data, represented as the 2D/3D coordinates of human joints, has been widely studied for human action recognition due to its high-level semantic information and environmental robustness. However, previous methods heavily rely on designing hand-crafted traversal rules or graph topologies to draw dependencies between the joints, which are limited in performance and generalizability. In this work, we present a novel decoupled spatial-temporal attention network(DSTA-Net) for skeleton-based action recognition. It involves solely the attention blocks, allowing for modeling spatial-temporal dependencies between joints without the requirement of knowing their positions or mutual connections. Specifically, to meet the specific requirements of the skeletal data, three techniques are proposed for building attention blocks, namely, spatial-temporal attention decoupling, decoupled position encoding and spatial global regularization. Besides, from the data aspect, we introduce a skeletal data decoupling technique to emphasize the specific characteristics of space/time and different motion scales, resulting in a more comprehensive understanding of the human actions.To test the effectiveness of the proposed method, extensive experiments are conducted on four challenging datasets for skeleton-based gesture and action recognition, namely, SHREC, DHG, NTU-60 and NTU-120, where DSTA-Net achieves state-of-the-art performance on all of them.

Lei Shi, Yifan Zhang, Jian Cheng et al.

Skeleton data, which consists of only the 2D/3D coordinates of the human joints, has been widely studied for human action recognition. Existing methods take the semantics as prior knowledge to group human joints and draw correlations according to their spatial locations, which we call the semantic perspective for skeleton modeling. In this paper, in contrast to previous approaches, we propose to model skeletons from a novel spatial perspective, from which the model takes the spatial location as prior knowledge to group human joints and mines the discriminative patterns of local areas in a hierarchical manner. The two perspectives are orthogonal and complementary to each other; and by fusing them in a unified framework, our method achieves a more comprehensive understanding of the skeleton data. Besides, we customized two networks for the two perspectives. From the semantic perspective, we propose a Transformer-like network that is expert in modeling joint correlations, and present three effective techniques to adapt it for skeleton data. From the spatial perspective, we transform the skeleton data into the sparse format for efficient feature extraction and present two types of sparse convolutional networks for sparse skeleton modeling. Extensive experiments are conducted on three challenging datasets for skeleton-based human action/gesture recognition, namely, NTU-60, NTU-120 and SHREC, where our method achieves state-of-the-art performance.

Lei Shi, Yifan Zhang, Jian Cheng et al.

Pose-based action recognition has drawn considerable attention recently. Existing methods exploit the joint positions to extract the body-part features from the activation map of the convolutional networks to assist human action recognition. However, these features are simply concatenated or max-pooled in previous works. The structured correlations among the body parts, which are essential for understanding complex human actions, are not fully exploited. To address the problem, we propose a pose-based graph convolutional network (PGCN), which encodes the body-part features into a human-based spatiotemporal graph, and explicitly models their correlations with a novel light-weight adaptive graph convolutional module to produce a highly discriminative representation for human action recognition. Besides, we discover that the backbone network tends to identify patterns from the most discriminative areas of the input regardless of the others. Thus the features pooled by the joint positions from other areas are less informative, which consequently hampers the performance of the followed aggregation process for recognizing actions. To alleviate this issue, we introduce a simple intermediate dense supervision mechanism for the backbone network, which adequately addresses the problem with no extra computation cost during inference. We evaluate the proposed approach on three popular benchmarks for pose-based action recognition tasks, i.e., Sub-JHMDB, PennAction and NTU-RGBD, where our approach significantly outperforms state-of-the-arts without the bells and whistles.

Lei Shi, Yifan Zhang, Jian Cheng et al.

In skeleton-based action recognition, graph convolutional networks (GCNs), which model the human body skeletons as spatiotemporal graphs, have achieved remarkable performance. However, in existing GCN-based methods, the topology of the graph is set manually, and it is fixed over all layers and input samples. This may not be optimal for the hierarchical GCN and diverse samples in action recognition tasks. In addition, the second-order information (the lengths and directions of bones) of the skeleton data, which is naturally more informative and discriminative for action recognition, is rarely investigated in existing methods. In this work, we propose a novel two-stream adaptive graph convolutional network (2s-AGCN) for skeleton-based action recognition. The topology of the graph in our model can be either uniformly or individually learned by the BP algorithm in an end-to-end manner. This data-driven method increases the flexibility of the model for graph construction and brings more generality to adapt to various data samples. Moreover, a two-stream framework is proposed to model both the first-order and the second-order information simultaneously, which shows notable improvement for the recognition accuracy. Extensive experiments on the two large-scale datasets, NTU-RGBD and Kinetics-Skeleton, demonstrate that the performance of our model exceeds the state-of-the-art with a significant margin.

Congqi Cao, Yifan Zhang, Chunjie Zhang et al.

Three dimensional convolutional neural networks (3D CNNs) have been established as a powerful tool to simultaneously learn features from both spatial and temporal dimensions, which is suitable to be applied to video-based action recognition. In this work, we propose not to directly use the activations of fully-connected layers of a 3D CNN as the video feature, but to use selective convolutional layer activations to form a discriminative descriptor for video. It pools the feature on the convolutional layers under the guidance of body joint positions. Two schemes of mapping body joints into convolutional feature maps for pooling are discussed. The body joint positions can be obtained from any off-the-shelf skeleton estimation algorithm. The helpfulness of the body joint guided feature pooling with inaccurate skeleton estimation is systematically evaluated. To make it end-to-end and do not rely on any sophisticated body joint detection algorithm, we further propose a two-stream bilinear model which can learn the guidance from the body joints and capture the spatio-temporal features simultaneously. In this model, the body joint guided feature pooling is conveniently formulated as a bilinear product operation. Experimental results on three real-world datasets demonstrate the effectiveness of body joint guided pooling which achieves promising performance.