Yiyi Zhang

Ying Zheng, Lei Yao, Yuejiao Su et al.

Embodied learning for object-centric robotic manipulation is a rapidly developing and challenging area in embodied AI. It is crucial for advancing next-generation intelligent robots and has garnered significant interest recently. Unlike data-driven machine learning methods, embodied learning focuses on robot learning through physical interaction with the environment and perceptual feedback, making it especially suitable for robotic manipulation. In this paper, we provide a comprehensive survey of the latest advancements in this field and categorize the existing work into three main branches: 1) Embodied perceptual learning, which aims to predict object pose and affordance through various data representations; 2) Embodied policy learning, which focuses on generating optimal robotic decisions using methods such as reinforcement learning and imitation learning; 3) Embodied task-oriented learning, designed to optimize the robot's performance based on the characteristics of different tasks in object grasping and manipulation. In addition, we offer an overview and discussion of public datasets, evaluation metrics, representative applications, current challenges, and potential future research directions. A project associated with this survey has been established at https://github.com/RayYoh/OCRM_survey.

Ying Zheng, Yiyi Zhang, Yi Wang et al.

Source-Free Domain Adaptation (SFDA) adapts pre-trained models to unlabeled target domains without requiring access to source data. Although state-of-the-art methods leveraging local neighborhood structures show promise for SFDA, they tend to over-rely on prediction similarity among neighbors. This over-reliance accelerates the forgetting of source knowledge and increases susceptibility to local noise overfitting. To address these issues, we introduce ProCal, a probability calibration method that dynamically calibrates neighborhood-based predictions through a dual-model collaborative prediction mechanism. ProCal integrates the source model's initial predictions with the current model's online outputs to effectively calibrate neighbor probabilities. This strategy not only mitigates the interference of local noise but also preserves the discriminative information from the source model, thereby achieving a balance between knowledge retention and domain adaptation. Furthermore, we design a joint optimization objective that combines a soft supervision loss with a diversity loss to guide the target model. Our theoretical analysis shows that ProCal converges to an equilibrium where source knowledge and target information are effectively fused, reducing both knowledge forgetting and overfitting. We validate the effectiveness of our approach through extensive experiments on 31 cross-domain tasks across four public datasets. Our code is available at: https://github.com/zhengyinghit/ProCal.

Yujie Gao, Yao Xiao, Xiangnan Zhu et al.

Reconstructing a complete 3D head from a single portrait remains challenging because existing methods still face a sharp quality-speed trade-off: high-fidelity pipelines often rely on multi-stage processing and per-subject optimization, while fast feed-forward models struggle with complete geometry and fine appearance details. To bridge this gap, we propose Any3DAvatar, a fast and high-quality method for single-image 3D Gaussian head avatar generation, whose fastest setting reconstructs a full head in under one second while preserving high-fidelity geometry and texture. First, we build AnyHead, a unified data suite that combines identity diversity, dense multi-view supervision, and realistic accessories, filling the main gaps of existing head data in coverage, full-head geometry, and complex appearance. Second, rather than sampling unstructured noise, we initialize from a Plücker-aware structured 3D Gaussian scaffold and perform one-step conditional denoising, formulating full-head reconstruction into a single forward pass while retaining high fidelity. Third, we introduce auxiliary view-conditioned appearance supervision on the same latent tokens alongside 3D Gaussian reconstruction, improving novel-view texture details at zero extra inference cost. Experiments show that Any3DAvatar outperforms prior single-image full-head reconstruction methods in rendering fidelity while remaining substantially faster.

Owen Oertell, Jonathan D. Chang, Yiyi Zhang et al.

Reinforcement learning (RL) has improved guided image generation with diffusion models by directly optimizing rewards that capture image quality, aesthetics, and instruction following capabilities. However, the resulting generative policies inherit the same iterative sampling process of diffusion models that causes slow generation. To overcome this limitation, consistency models proposed learning a new class of generative models that directly map noise to data, resulting in a model that can generate an image in as few as one sampling iteration. In this work, to optimize text-to-image generative models for task specific rewards and enable fast training and inference, we propose a framework for fine-tuning consistency models via RL. Our framework, called Reinforcement Learning for Consistency Model (RLCM), frames the iterative inference process of a consistency model as an RL procedure. Comparing to RL finetuned diffusion models, RLCM trains significantly faster, improves the quality of the generation measured under the reward objectives, and speeds up the inference procedure by generating high quality images with as few as two inference steps. Experimentally, we show that RLCM can adapt text-to-image consistency models to objectives that are challenging to express with prompting, such as image compressibility, and those derived from human feedback, such as aesthetic quality. Our code is available at https://rlcm.owenoertell.com.

Ying Zheng, Yiyi Zhang, Yi Wang et al.

Conventional deep network training generally optimizes all samples under a largely uniform learning paradigm, without explicitly modeling the heterogeneous competition among them. Such an oversimplified treatment can lead to several well-known issues, including bias under class imbalance, insufficient learning of hard samples, and the erroneous reinforcement of noisy samples. In this work, we present \textit{Natural Selection} (NS), a novel evolution-inspired optimization method that explicitly incorporates competitive interactions into deep network training. Unlike conventional sample reweighting strategies that rely mainly on predefined heuristics or static criteria, NS estimates the competitive status of each sample in a group-wise context and uses it to adaptively regulate its training contribution. Specifically, NS first assembles multiple samples into a composite image and rescales it to the original input size for model inference. Based on the resulting predictions, a natural selection score is computed for each sample to characterize its relative competitive variation within the constructed group. These scores are then used to dynamically reweight the sample-wise loss, thereby introducing an explicit competition-driven mechanism into the optimization process. In this way, NS provides a simple yet effective means of moving beyond uniform sample treatment and enables more adaptive and balanced model optimization. Extensive experiments on 12 public datasets across four image classification tasks demonstrate the effectiveness of the proposed method. Moreover, NS is compatible with diverse network architectures and does not depend on task-specific assumptions, indicating its strong generality and practical potential. The code will be made publicly available.

Siyu Sun, Jing Ren, Zhaohe Liao et al.

Adapting decoder-only multimodal large language models (MLLMs) for unified multimodal retrieval faces two structural gaps. First, existing methods rely on implicit pooling, which overloads the hidden state of a standard vocabulary token (e.g., <EOS>) as the sequence-level representation, a mechanism never designed for information aggregation. Second, contrastive fine-tuning specifies what the embedding should match but provides no token-level guidance on how information should be compressed into it. We address both gaps with two complementary components. Architecturally, we introduce Bottleneck Tokens (BToks), a small set of learnable tokens that serve as a fixed-capacity explicit pooling mechanism. For training, we propose Generative Information Condensation: a next-token prediction objective coupled with a Condensation Mask that severs the direct attention path from target tokens to query tokens. All predictive signals are thereby forced through the BToks, converting the generative loss into dense, token-level supervision for semantic compression. At inference time, only the input and BToks are processed in a single forward pass with negligible overhead over conventional last-token pooling. On MMEB-V2 (78 datasets, 3 modalities, 9 meta-tasks), our approach achieves state-of-the-art among 2B-scale methods under comparable data conditions, attaining an Overall score of 59.0 (+3.6 over VLM2Vec-V2) with substantial gains on semantically demanding tasks (e.g., +12.6 on Video-QA).

Ka Nam Canaan Law, Mingshuo Yu, Lianglei Zhang et al.

The quality control of printed circuit boards (PCBs) is paramount in advancing electronic device technology. While numerous machine learning methodologies have been utilized to augment defect detection efficiency and accuracy, previous studies have predominantly focused on optimizing individual models for specific defect types, often overlooking the potential synergies between different approaches. This paper introduces a comprehensive inspection framework leveraging an ensemble learning strategy to address this gap. Initially, we utilize four distinct PCB defect detection models utilizing state-of-the-art methods: EfficientDet, MobileNet SSDv2, Faster RCNN, and YOLOv5. Each method is capable of identifying PCB defects independently. Subsequently, we integrate these models into an ensemble learning framework to enhance detection performance. A comparative analysis reveals that our ensemble learning framework significantly outperforms individual methods, achieving a 95% accuracy in detecting diverse PCB defects. These findings underscore the efficacy of our proposed ensemble learning framework in enhancing PCB quality control processes.

Xinhao Hu, Yiyi Zhang, Liqing Zhang et al.

Pedestrian motion, due to its causal nature, is strongly influenced by domain gaps arising from discrepancies between training and testing data distributions. Focusing on 3D human pose estimation, this work presents a controllable human pose generation framework that synthesizes diverse video data by systematically varying poses, backgrounds, and camera viewpoints. This generative augmentation enriches training datasets, enhances model generalization, and alleviates the limitations of existing methods in handling domain discrepancies. By leveraging both indoor/real-world and outdoor/virtual datasets, we perform cross-domain data fusion and controllable video generation to construct enriched training data, tailored to realistic deployment settings. Extensive experiments show that the augmented datasets significantly improve model performance on unseen scenarios and datasets, validating the effectiveness of the proposed approach.

Zengqi Zhao, Weidi Xia, Peter Wei et al.

We present DOCFORGE-BENCH, the first unified zero-shot benchmark for document forgery detection, evaluating 14 methods across eight datasets spanning text tampering, receipt forgery, and identity document manipulation. Unlike fine-tuning-oriented evaluations such as ForensicHub [Du et al., 2025], DOCFORGE-BENCH applies all methods with their published pretrained weights and no domain adaptation -- a deliberate design choice that reflects the realistic deployment scenario where practitioners lack labeled document training data. Our central finding is a pervasive calibration failure invisible under single-threshold protocols: methods achieve moderate Pixel-AUC (>=0.76) yet near-zero Pixel-F1. This AUC-F1 gap is not a discrimination failure but a score-distribution shift: tampered regions occupy only 0.27-4.17% of pixels in document images -- an order of magnitude less than in natural image benchmarks -- making the standard tau=0.5 threshold catastrophically miscalibrated. Oracle-F1 is 2-10x higher than fixed-threshold Pixel-F1, confirming that calibration, not representation, is the bottleneck. A controlled calibration experiment validates this: adapting a single threshold on N=10 domain images recovers 39-55% of the Oracle-F1 gap, demonstrating that threshold adaptation -- not retraining -- is the key missing step for practical deployment. Overall, no evaluated method works reliably out-of-the-box on diverse document types, underscoring that document forgery detection remains an unsolved problem. We further note that all eight datasets predate the era of generative AI editing; benchmarks covering diffusion- and LLM-based document forgeries represent a critical open gap on the modern attack surface.

Nicolas Espinosa-Dice, Yiyi Zhang, Yiding Chen et al.

Diffusion and flow models have emerged as powerful generative approaches capable of modeling diverse and multimodal behavior. However, applying these models to offline reinforcement learning (RL) remains challenging due to the iterative nature of their noise sampling processes, making policy optimization difficult. In this paper, we introduce Scalable Offline Reinforcement Learning (SORL), a new offline RL algorithm that leverages shortcut models - a novel class of generative models - to scale both training and inference. SORL's policy can capture complex data distributions and can be trained simply and efficiently in a one-stage training procedure. At test time, SORL introduces both sequential and parallel inference scaling by using the learned Q-function as a verifier. We demonstrate that SORL achieves strong performance across a range of offline RL tasks and exhibits positive scaling behavior with increased test-time compute. We release the code at nico-espinosadice.github.io/projects/sorl.

Yiding Chen, Yiyi Zhang, Owen Oertell et al.

Diffusion models accomplish remarkable success in data generation tasks across various domains. However, the iterative sampling process is computationally expensive. Consistency models are proposed to learn consistency functions to map from noise to data directly, which allows one-step fast data generation and multistep sampling to improve sample quality. In this paper, we study the convergence of consistency models when the self-consistency property holds approximately under the training distribution. Our analysis requires only mild data assumption and applies to a family of forward processes. When the target data distribution has bounded support or has tails that decay sufficiently fast, we show that the samples generated by the consistency model are close to the target distribution in Wasserstein distance; when the target distribution satisfies some smoothness assumption, we show that with an additional perturbation step for smoothing, the generated samples are close to the target distribution in total variation distance. We provide two case studies with commonly chosen forward processes to demonstrate the benefit of multistep sampling.

Ying Zheng, Yiyi Zhang, Yi Wang et al.

Source-free domain adaptation in visual emotion recognition (SFDA-VER) is a highly challenging task that requires adapting VER models to the target domain without relying on source data, which is of great significance for data privacy protection. However, due to the unignorable disparities between visual emotion data and traditional image classification data, existing SFDA methods perform poorly on this task. In this paper, we investigate the SFDA-VER task from a fuzzy perspective and identify two key issues: fuzzy emotion labels and fuzzy pseudo-labels. These issues arise from the inherent uncertainty of emotion annotations and the potential mispredictions in pseudo-labels. To address these issues, we propose a novel fuzzy-aware loss (FAL) to enable the VER model to better learn and adapt to new domains under fuzzy labels. Specifically, FAL modifies the standard cross entropy loss and focuses on adjusting the losses of non-predicted categories, which prevents a large number of uncertain or incorrect predictions from overwhelming the VER model during adaptation. In addition, we provide a theoretical analysis of FAL and prove its robustness in handling the noise in generated pseudo-labels. Extensive experiments on 26 domain adaptation sub-tasks across three benchmark datasets demonstrate the effectiveness of our method.

Yiyi Zhang, Yuchen Yuan, Ying Zheng et al.

Surgical triplet recognition, which involves identifying instrument, verb, target, and their combinations, is a complex surgical scene understanding challenge plagued by long-tailed data distribution. The mainstream multi-task learning paradigm benefiting from cross-task collaborative promotion has shown promising performance in identifying triples, but two key challenges remain: 1) inter-task optimization conflicts caused by entangling task-generic and task-specific representations; 2) intra-task optimization conflicts due to class-imbalanced training data. To overcome these difficulties, we propose the MLLM-Engaged Joint Optimization (MEJO) framework that empowers both inter- and intra-task optimization for surgical triplet recognition. For inter-task optimization, we introduce the Shared-Specific-Disentangled (S$^2$D) learning scheme that decomposes representations into task-shared and task-specific components. To enhance task-shared representations, we construct a Multimodal Large Language Model (MLLM) powered probabilistic prompt pool to dynamically augment visual features with expert-level semantic cues. Additionally, comprehensive task-specific cues are modeled via distinct task prompts covering the temporal-spatial dimensions, effectively mitigating inter-task ambiguities. To tackle intra-task optimization conflicts, we develop a Coordinated Gradient Learning (CGL) strategy, which dissects and rebalances the positive-negative gradients originating from head and tail classes for more coordinated learning behaviors. Extensive experiments on the CholecT45 and CholecT50 datasets demonstrate the superiority of our proposed framework, validating its effectiveness in handling optimization conflicts.

Liang Xu, Chengqun Yang, Zili Lin et al.

Learning action models from real-world human-centric interaction datasets is important towards building general-purpose intelligent assistants with efficiency. However, most existing datasets only offer specialist interaction category and ignore that AI assistants perceive and act based on first-person acquisition. We urge that both the generalist interaction knowledge and egocentric modality are indispensable. In this paper, we embed the manual-assisted task into a vision-language-action framework, where the assistant provides services to the instructor following egocentric vision and commands. With our hybrid RGB-MoCap system, pairs of assistants and instructors engage with multiple objects and the scene following GPT-generated scripts. Under this setting, we accomplish InterVLA, the first large-scale human-object-human interaction dataset with 11.4 hours and 1.2M frames of multimodal data, spanning 2 egocentric and 5 exocentric videos, accurate human/object motions and verbal commands. Furthermore, we establish novel benchmarks on egocentric human motion estimation, interaction synthesis, and interaction prediction with comprehensive analysis. We believe that our InterVLA testbed and the benchmarks will foster future works on building AI agents in the physical world.

Jianfu Zhang, Yujie Gao, Jiahui Zhan et al.

In this work, we introduce a novel high-fidelity 3D head reconstruction method from a single portrait image, regardless of perspective, expression, or accessories. Despite significant efforts in adapting 2D generative models for novel view synthesis and 3D optimization, most methods struggle to produce high-quality 3D portraits. The lack of crucial information, such as identity, expression, hair, and accessories, limits these approaches in generating realistic 3D head models. To address these challenges, we construct a new high-quality dataset containing 227 sequences of digital human portraits captured from 96 different perspectives, totalling 21,792 frames, featuring diverse expressions and accessories. To further improve performance, we integrate identity and expression information into the multi-view diffusion process to enhance facial consistency across views. Specifically, we apply identity- and expression-aware guidance and supervision to extract accurate facial representations, which guide the model and enforce objective functions to ensure high identity and expression consistency during generation. Finally, we generate an orbital video around the portrait consisting of 96 multi-view frames, which can be used for 3D portrait model reconstruction. Our method demonstrates robust performance across challenging scenarios, including side-face angles and complex accessories

Yiyi Zhang, Xingyu Chen, Kexin Chen et al.

Recent years have witnessed extensive efforts to enhance Large Language Models (LLMs) across various domains, alongside growing attention to their ethical implications. However, a critical challenge remains largely overlooked: LLMs must balance between rejecting harmful requests for safety and accommodating legitimate ones for utility. This paper presents a Direct Preference Optimization (DPO) based alignment framework that achieves better overall performance by addressing this ethical-utility trade-off, using chemical domain applications as a proof-of-concept. Our alignment pipeline starts with a GPT-assisted three-phase data generation scheme, in which we create LibraChemQA, a chemical question-answering dataset comprising 31.6k triplet instances. By incorporating an innovative balanced seed in the data generation process, our framework systematically considers both legitimate and illegitimate requests. The framework also introduces a rephrasing mechanism for efficient data augmentation that enhances the model's chemical comprehension. We further develop a novel hybrid evaluation scheme with LLM judges for precise assessment of both safety and utility. Experimental results demonstrate our model's substantial improvements in overall performance where both safety and utility are considered - the resulting model outperforms leading LLMs including Claude-3, GPT-4o, and LLaMA-3 by margins of 13.44%, 7.16%, and 7.10% respectively on our released benchmark. At the end of this paper, we analyze experimental results obtained from testing DeepSeek-R1 on our benchmark and reveal the critical ethical concerns raised by this highly acclaimed model. We highlight that the long Chain-of-Thought (CoT) reasoning process employed by DeepSeek-R1, as well as other LLMs distilled from it, introduces significant ethical vulnerabilities when exposed to users.

Yiyi Zhang, Zhiwen Ying, Ying Zheng et al.

Plant leaf identification is crucial for biodiversity protection and conservation and has gradually attracted the attention of academia in recent years. Due to the high similarity among different varieties, leaf cultivar recognition is also considered to be an ultra-fine-grained visual classification (UFGVC) task, which is facing a huge challenge. In practice, an instance may be related to multiple varieties to varying degrees, especially in the UFGVC datasets. However, deep learning methods trained on one-hot labels fail to reflect patterns shared across categories and thus perform poorly on this task. To address this issue, we generate soft targets integrated with inter-class similarity information. Specifically, we continuously update the prototypical features for each category and then capture the similarity scores between instances and prototypes accordingly. Original one-hot labels and the similarity scores are incorporated to yield enhanced labels. Prototype-enhanced soft labels not only contain original one-hot label information, but also introduce rich inter-category semantic association information, thus providing more effective supervision for deep model training. Extensive experimental results on public datasets show that our method can significantly improve the performance on the UFGVC task of leaf cultivar identification.

Yiyi Zhang, Ying Zheng, Xiaogang Xu et al.

Cross-domain few-shot learning (CDFSL) remains a largely unsolved problem in the area of computer vision, while self-supervised learning presents a promising solution. Both learning methods attempt to alleviate the dependency of deep networks on the requirement of large-scale labeled data. Although self-supervised methods have recently advanced dramatically, their utility on CDFSL is relatively unexplored. In this paper, we investigate the role of self-supervised representation learning in the context of CDFSL via a thorough evaluation of existing methods. It comes as a surprise that even with shallow architectures or small training datasets, self-supervised methods can perform favorably compared to the existing SOTA methods. Nevertheless, no single self-supervised approach dominates all datasets indicating that existing self-supervised methods are not universally applicable. In addition, we find that representations extracted from self-supervised methods exhibit stronger robustness than the supervised method. Intriguingly, whether self-supervised representations perform well on the source domain has little correlation with their applicability on the target domain. As part of our study, we conduct an objective measurement of the performance for six kinds of representative classifiers. The results suggest Prototypical Classifier as the standard evaluation recipe for CDFSL.

Ziyu Wang, Ke Chen, Junyan Jiang et al.

Music arrangement generation is a subtask of automatic music generation, which involves reconstructing and re-conceptualizing a piece with new compositional techniques. Such a generation process inevitably requires reference from the original melody, chord progression, or other structural information. Despite some promising models for arrangement, they lack more refined data to achieve better evaluations and more practical results. In this paper, we propose POP909, a dataset which contains multiple versions of the piano arrangements of 909 popular songs created by professional musicians. The main body of the dataset contains the vocal melody, the lead instrument melody, and the piano accompaniment for each song in MIDI format, which are aligned to the original audio files. Furthermore, we provide the annotations of tempo, beat, key, and chords, where the tempo curves are hand-labeled and others are done by MIR algorithms. Finally, we conduct several baseline experiments with this dataset using standard deep music generation algorithms.

Ziyu Wang, Yiyi Zhang, Yixiao Zhang et al.

The dominant approach for music representation learning involves the deep unsupervised model family variational autoencoder (VAE). However, most, if not all, viable attempts on this problem have largely been limited to monophonic music. Normally composed of richer modality and more complex musical structures, the polyphonic counterpart has yet to be addressed in the context of music representation learning. In this work, we propose the PianoTree VAE, a novel tree-structure extension upon VAE aiming to fit the polyphonic music learning. The experiments prove the validity of the PianoTree VAE via (i)-semantically meaningful latent code for polyphonic segments; (ii)-more satisfiable reconstruction aside of decent geometry learned in the latent space; (iii)-this model's benefits to the variety of the downstream music generation.

Yiyi Zhang, Li Niu, Ziqi Pan et al.

Static image action recognition, which aims to recognize action based on a single image, usually relies on expensive human labeling effort such as adequate labeled action images and large-scale labeled image dataset. In contrast, abundant unlabeled videos can be economically obtained. Therefore, several works have explored using unlabeled videos to facilitate image action recognition, which can be categorized into the following two groups: (a) enhance visual representations of action images with a designed proxy task on unlabeled videos, which falls into the scope of self-supervised learning; (b) generate auxiliary representations for action images with the generator learned from unlabeled videos. In this paper, we integrate the above two strategies in a unified framework, which consists of Visual Representation Enhancement (VRE) module and Motion Representation Augmentation (MRA) module. Specifically, the VRE module includes a proxy task which imposes pseudo motion label constraint and temporal coherence constraint on unlabeled videos, while the MRA module could predict the motion information of a static action image by exploiting unlabeled videos. We demonstrate the superiority of our framework based on four benchmark human action datasets with limited labeled data.

Ruihan Yang, Tianyao Chen, Yiyi Zhang et al.

Variational Autoencoders(VAEs) have already achieved great results on image generation and recently made promising progress on music generation. However, the generation process is still quite difficult to control in the sense that the learned latent representations lack meaningful music semantics. It would be much more useful if people can modify certain music features, such as rhythm and pitch contour, via latent representations to test different composition ideas. In this paper, we propose a new method to inspect the pitch and rhythm interpretations of the latent representations and we name it disentanglement by augmentation. Based on the interpretable representations, an intuitive graphical user interface is designed for users to better direct the music creation process by manipulating the pitch contours and rhythmic complexity.

Tao Bi, Yiyi Zhang, Chongyang Wang et al.

This position paper takes the first step to attempt to present the initial characterization of HCI research in China. We discuss the current streams and methodologies of Chinese HCI research based on two well-known HCI theories: Micro/Marco-HCI and the Three Paradigms of HCI. We evaluate the discussion with a survey of Chinese publications at CHI 2019, which shows HCI research in China has less attention to Macro-HCI topics and the third paradigms of HCI (Phenomenologically situated Interaction). We then propose future HCI research directions such as paying more attention to Macro-HCI topics and third paradigm of HCI, combining research methodologies from multiple HCI paradigms, including emergent users who have less access to technology, and addressing the cultural dimensions in order to provide better technical solutions and support.