Chunhui Zhang

Yijun Tian, Kaiwen Dong, Chunhui Zhang et al.

Generative self-supervised learning (SSL), especially masked autoencoders, has become one of the most exciting learning paradigms and has shown great potential in handling graph data. However, real-world graphs are always heterogeneous, which poses three critical challenges that existing methods ignore: 1) how to capture complex graph structure? 2) how to incorporate various node attributes? and 3) how to encode different node positions? In light of this, we study the problem of generative SSL on heterogeneous graphs and propose HGMAE, a novel heterogeneous graph masked autoencoder model to address these challenges. HGMAE captures comprehensive graph information via two innovative masking techniques and three unique training strategies. In particular, we first develop metapath masking and adaptive attribute masking with dynamic mask rate to enable effective and stable learning on heterogeneous graphs. We then design several training strategies including metapath-based edge reconstruction to adopt complex structural information, target attribute restoration to incorporate various node attributes, and positional feature prediction to encode node positional information. Extensive experiments demonstrate that HGMAE outperforms both contrastive and generative state-of-the-art baselines on several tasks across multiple datasets. Codes are available at https://github.com/meettyj/HGMAE.

Chunhui Zhang, Xin Sun, Yiqian Yang et al.

Current mainstream vision-language (VL) tracking framework consists of three parts, \ie a visual feature extractor, a language feature extractor, and a fusion model. To pursue better performance, a natural modus operandi for VL tracking is employing customized and heavier unimodal encoders, and multi-modal fusion models. Albeit effective, existing VL trackers separate feature extraction and feature integration, resulting in extracted features that lack semantic guidance and have limited target-aware capability in complex scenarios, \eg similar distractors and extreme illumination. In this work, inspired by the recent success of exploring foundation models with unified architecture for both natural language and computer vision tasks, we propose an All-in-One framework, which learns joint feature extraction and interaction by adopting a unified transformer backbone. Specifically, we mix raw vision and language signals to generate language-injected vision tokens, which we then concatenate before feeding into the unified backbone architecture. This approach achieves feature integration in a unified backbone, removing the need for carefully-designed fusion modules and resulting in a more effective and efficient VL tracking framework. To further improve the learning efficiency, we introduce a multi-modal alignment module based on cross-modal and intra-modal contrastive objectives, providing more reasonable representations for the unified All-in-One transformer backbone. Extensive experiments on five benchmarks, \ie OTB99-L, TNL2K, LaSOT, LaSOT$_{\rm Ext}$ and WebUAV-3M, demonstrate the superiority of the proposed tracker against existing state-of-the-arts on VL tracking. Codes will be made publicly available at https://github.com/983632847/All-in-One.

Chunhui Zhang, Yixiong Chen, Li Liu et al.

The self-supervised ultrasound (US) video model pretraining can use a small amount of labeled data to achieve one of the most promising results on US diagnosis. However, it does not take full advantage of multi-level knowledge for learning deep neural networks (DNNs), and thus is difficult to learn transferable feature representations. This work proposes a hierarchical contrastive learning (HiCo) method to improve the transferability for the US video model pretraining. HiCo introduces both peer-level semantic alignment and cross-level semantic alignment to facilitate the interaction between different semantic levels, which can effectively accelerate the convergence speed, leading to better generalization and adaptation of the learned model. Additionally, a softened objective function is implemented by smoothing the hard labels, which can alleviate the negative effect caused by local similarities of images between different classes. Experiments with HiCo on five datasets demonstrate its favorable results over state-of-the-art approaches. The source code of this work is publicly available at https://github.com/983632847/HiCo.

Yixiong Chen, Chunhui Zhang, Chris H. Q. Ding et al.

Well-annotated medical datasets enable deep neural networks (DNNs) to gain strong power in extracting lesion-related features. Building such large and well-designed medical datasets is costly due to the need for high-level expertise. Model pre-training based on ImageNet is a common practice to gain better generalization when the data amount is limited. However, it suffers from the domain gap between natural and medical images. In this work, we pre-train DNNs on ultrasound (US) domains instead of ImageNet to reduce the domain gap in medical US applications. To learn US image representations based on unlabeled US videos, we propose a novel meta-learning-based contrastive learning method, namely Meta Ultrasound Contrastive Learning (Meta-USCL). To tackle the key challenge of obtaining semantically consistent sample pairs for contrastive learning, we present a positive pair generation module along with an automatic sample weighting module based on meta-learning. Experimental results on multiple computer-aided diagnosis (CAD) problems, including pneumonia detection, breast cancer classification, and breast tumor segmentation, show that the proposed self-supervised method reaches state-of-the-art (SOTA). The codes are available at https://github.com/Schuture/Meta-USCL.

Yiren Jian, Tingkai Liu, Yunzhe Tao et al.

In this paper, we introduce $\text{EVL}_{\text{Gen}}$, a streamlined framework designed for the pre-training of visually conditioned language generation models with high computational demands, utilizing frozen pre-trained large language models (LLMs). The conventional approach in vision-language pre-training (VLP) typically involves a two-stage optimization process: an initial resource-intensive phase dedicated to general-purpose vision-language representation learning, focused on extracting and consolidating relevant visual features. This is followed by a subsequent phase that emphasizes end-to-end alignment between visual and linguistic modalities. Our novel one-stage, single-loss framework bypasses the computationally demanding first training stage by gradually merging similar visual tokens during training, while avoiding model collapse caused by single-stage training of BLIP-2 type models. The gradual merging process effectively condenses visual information while preserving semantic richness, resulting in rapid convergence without compromising performance. Our experimental findings demonstrate that our approach accelerates the training of vision-language models by a factor of 5 without a noticeable impact on overall performance. Furthermore, we illustrate that our models significantly narrow the performance gap to current vision-language models using only 1/10 of the data. Finally, we showcase how our image-text models can seamlessly adapt to video-conditioned language generation tasks through novel soft attentive temporal token contextualizing modules. Code is available at \url{https://github.com/yiren-jian/EVLGen}.

Zheyuan Liu, Guangyao Dou, Yijun Tian et al.

Machine Unlearning (MU) algorithms have become increasingly critical due to the imperative adherence to data privacy regulations. The primary objective of MU is to erase the influence of specific data samples on a given model without the need to retrain it from scratch. Accordingly, existing methods focus on maximizing user privacy protection. However, there are different degrees of privacy regulations for each real-world web-based application. Exploring the full spectrum of trade-offs between privacy, model utility, and runtime efficiency is critical for practical unlearning scenarios. Furthermore, designing the MU algorithm with simple control of the aforementioned trade-off is desirable but challenging due to the inherent complex interaction. To address the challenges, we present Controllable Machine Unlearning (ConMU), a novel framework designed to facilitate the calibration of MU. The ConMU framework contains three integral modules: an important data selection module that reconciles the runtime efficiency and model generalization, a progressive Gaussian mechanism module that balances privacy and model generalization, and an unlearning proxy that controls the trade-offs between privacy and runtime efficiency. Comprehensive experiments on various benchmark datasets have demonstrated the robust adaptability of our control mechanism and its superiority over established unlearning methods. ConMU explores the full spectrum of the Privacy-Utility-Efficiency trade-off and allows practitioners to account for different real-world regulations. Source code available at: https://github.com/guangyaodou/ConMU.

Shiming Ge, Zhao Luo, Chunhui Zhang et al.

Deep trackers have proven success in visual tracking. Typically, these trackers employ optimally pre-trained deep networks to represent all diverse objects with multi-channel features from some fixed layers. The deep networks employed are usually trained to extract rich knowledge from massive data used in object classification and so they are capable to represent generic objects very well. However, these networks are too complex to represent a specific moving object, leading to poor generalization as well as high computational and memory costs. This paper presents a novel and general framework termed channel distillation to facilitate deep trackers. To validate the effectiveness of channel distillation, we take discriminative correlation filter (DCF) and ECO for example. We demonstrate that an integrated formulation can turn feature compression, response map generation, and model update into a unified energy minimization problem to adaptively select informative feature channels that improve the efficacy of tracking moving objects on the fly. Channel distillation can accurately extract good channels, alleviating the influence of noisy channels and generally reducing the number of channels, as well as adaptively generalizing to different channels and networks. The resulting deep tracker is accurate, fast, and has low memory requirements. Extensive experimental evaluations on popular benchmarks clearly demonstrate the effectiveness and generalizability of our framework.

Zhichun Guo, Chunhui Zhang, Yujie Fan et al.

Graph neural networks (GNNs) have shown remarkable performance on diverse graph mining tasks. Although different GNNs can be unified as the same message passing framework, they learn complementary knowledge from the same graph. Knowledge distillation (KD) is developed to combine the diverse knowledge from multiple models. It transfers knowledge from high-capacity teachers to a lightweight student. However, to avoid oversmoothing, GNNs are often shallow, which deviates from the setting of KD. In this context, we revisit KD by separating its benefits from model compression and emphasizing its power of transferring knowledge. To this end, we need to tackle two challenges: how to transfer knowledge from compact teachers to a student with the same capacity; and, how to exploit student GNN's own strength to learn knowledge. In this paper, we propose a novel adaptive KD framework, called BGNN, which sequentially transfers knowledge from multiple GNNs into a student GNN. We also introduce an adaptive temperature module and a weight boosting module. These modules guide the student to the appropriate knowledge for effective learning. Extensive experiments have demonstrated the effectiveness of BGNN. In particular, we achieve up to 3.05% improvement for node classification and 6.35% improvement for graph classification over vanilla GNNs.

Zhongying Deng, Cheng Tang, Ziyan Huang et al. · pku

Foundation models have demonstrated remarkable success across diverse domains and tasks, primarily due to the thrive of large-scale, diverse, and high-quality datasets. However, in the field of medical imaging, the curation and assembling of such medical datasets are highly challenging due to the reliance on clinical expertise and strict ethical and privacy constraints, resulting in a scarcity of large-scale unified medical datasets and hindering the development of powerful medical foundation models. In this work, we present the largest survey to date of medical image datasets, covering over 1,000 open-access datasets with a systematic catalog of their modalities, tasks, anatomies, annotations, limitations, and potential for integration. Our analysis exposes a landscape that is modest in scale, fragmented across narrowly scoped tasks, and unevenly distributed across organs and modalities, which in turn limits the utility of existing medical image datasets for developing versatile and robust medical foundation models. To turn fragmentation into scale, we propose a metadata-driven fusion paradigm (MDFP) that integrates public datasets with shared modalities or tasks, thereby transforming multiple small data silos into larger, more coherent resources. Building on MDFP, we release an interactive discovery portal that enables end-to-end, automated medical image dataset integration, and compile all surveyed datasets into a unified, structured table that clearly summarizes their key characteristics and provides reference links, offering the community an accessible and comprehensive repository. By charting the current terrain and offering a principled path to dataset consolidation, our survey provides a practical roadmap for scaling medical imaging corpora, supporting faster data discovery, more principled dataset creation, and more capable medical foundation models.

Han Yue, Chunhui Zhang, Chuxu Zhang et al.

Recently, contrastiveness-based augmentation surges a new climax in the computer vision domain, where some operations, including rotation, crop, and flip, combined with dedicated algorithms, dramatically increase the model generalization and robustness. Following this trend, some pioneering attempts employ the similar idea to graph data. Nevertheless, unlike images, it is much more difficult to design reasonable augmentations without changing the nature of graphs. Although exciting, the current graph contrastive learning does not achieve as promising performance as visual contrastive learning. We conjecture the current performance of graph contrastive learning might be limited by the violation of the label-invariant augmentation assumption. In light of this, we propose a label-invariant augmentation for graph-structured data to address this challenge. Different from the node/edge modification and subgraph extraction, we conduct the augmentation in the representation space and generate the augmented samples in the most difficult direction while keeping the label of augmented data the same as the original samples. In the semi-supervised scenario, we demonstrate our proposed method outperforms the classical graph neural network based methods and recent graph contrastive learning on eight benchmark graph-structured data, followed by several in-depth experiments to further explore the label-invariant augmentation in several aspects.

Chunhui Zhang, Li Liu, Guanjie Huang et al.

Over the past decade, significant progress has been made in visual object tracking, largely due to the availability of large-scale datasets. However, these datasets have primarily focused on open-air scenarios and have largely overlooked underwater animal tracking-especially the complex challenges posed by camouflaged marine animals. To bridge this gap, we take a step forward by proposing the first large-scale multi-modal underwater camouflaged object tracking dataset, namely UW-COT220. Based on the proposed dataset, this work first comprehensively evaluates current advanced visual object tracking methods, including SAM- and SAM2-based trackers, in challenging underwater environments, \eg, coral reefs. Our findings highlight the improvements of SAM2 over SAM, demonstrating its enhanced ability to handle the complexities of underwater camouflaged objects. Furthermore, we propose a novel vision-language tracking framework called VL-SAM2, based on the video foundation model SAM2. Extensive experimental results demonstrate that the proposed VL-SAM2 achieves state-of-the-art performance across underwater and open-air object tracking datasets. The dataset and codes are available at~{\color{magenta}{https://github.com/983632847/Awesome-Multimodal-Object-Tracking}}.

Chunhui Zhang, Chao Huang, Yijun Tian et al.

Even pruned by the state-of-the-art network compression methods, Graph Neural Networks (GNNs) training upon non-Euclidean graph data often encounters relatively higher time costs, due to its irregular and nasty density properties, compared with data in the regular Euclidean space. Another natural property concomitantly with graph is class-imbalance which cannot be alleviated by the massive graph data while hindering GNNs' generalization. To fully tackle these unpleasant properties, (i) theoretically, we introduce a hypothesis about what extent a subset of the training data can approximate the full dataset's learning effectiveness. The effectiveness is further guaranteed and proved by the gradients' distance between the subset and the full set; (ii) empirically, we discover that during the learning process of a GNN, some samples in the training dataset are informative for providing gradients to update model parameters. Moreover, the informative subset is not fixed during training process. Samples that are informative in the current training epoch may not be so in the next one. We also notice that sparse subnets pruned from a well-trained GNN sometimes forget the information provided by the informative subset, reflected in their poor performances upon the subset. Based on these findings, we develop a unified data-model dynamic sparsity framework named Graph Decantation (GraphDec) to address challenges brought by training upon a massive class-imbalanced graph data. The key idea of GraphDec is to identify the informative subset dynamically during the training process by adopting sparse graph contrastive learning. Extensive experiments on benchmark datasets demonstrate that GraphDec outperforms baselines for graph and node tasks, with respect to classification accuracy and data usage efficiency.

Chunhui Zhang, Hongfu Liu, Jundong Li et al.

Prevailing deep graph learning models often suffer from label sparsity issue. Although many graph few-shot learning (GFL) methods have been developed to avoid performance degradation in face of limited annotated data, they excessively rely on labeled data, where the distribution shift in the test phase might result in impaired generalization ability. Additionally, they lack a general purpose as their designs are coupled with task or data-specific characteristics. To this end, we propose a general and effective Contrastive Graph Few-shot Learning framework (CGFL). CGFL leverages a self-distilled contrastive learning procedure to boost GFL. Specifically, our model firstly pre-trains a graph encoder with contrastive learning using unlabeled data. Later, the trained encoder is frozen as a teacher model to distill a student model with a contrastive loss. The distilled model is finally fed to GFL. CGFL learns data representation in a self-supervised manner, thus mitigating the distribution shift impact for better generalization and making model task and data-independent for a general graph mining purpose. Furthermore, we introduce an information-based method to quantitatively measure the capability of CGFL. Comprehensive experiments demonstrate that CGFL outperforms state-of-the-art baselines on several graph mining tasks in the few-shot scenario. We also provide quantitative measurement of CGFL's success.

Qianlong Wen, Zhongyu Ouyang, Chunhui Zhang et al.

Among different existing graph self-supervised learning strategies, graph contrastive learning (GCL) has been one of the most prevalent approaches to this problem. Despite the remarkable performance those GCL methods have achieved, existing GCL methods that heavily depend on various manually designed augmentation techniques still struggle to alleviate the feature suppression issue without risking losing task-relevant information. Consequently, the learned representation is either brittle or unilluminating. In light of this, we introduce the Graph Contrastive Learning with Cross-View Reconstruction (GraphCV), which follows the information bottleneck principle to learn minimal yet sufficient representation from graph data. Specifically, GraphCV aims to elicit the predictive (useful for downstream instance discrimination) and other non-predictive features separately. Except for the conventional contrastive loss which guarantees the consistency and sufficiency of the representation across different augmentation views, we introduce a cross-view reconstruction mechanism to pursue the disentanglement of the two learned representations. Besides, an adversarial view perturbed from the original view is added as the third view for the contrastive loss to guarantee the intactness of the global semantics and improve the representation robustness. We empirically demonstrate that our proposed model outperforms the state-of-the-art on graph classification task over multiple benchmark datasets.

Chunhui Zhang, Li Liu, Zhipeng Zhang et al.

Unmanned Aerial Vehicles (UAVs) offer wide-ranging applications but also pose significant safety and privacy violation risks in areas like airport and infrastructure inspection, spurring the rapid development of Anti-UAV technologies in recent years. However, current Anti-UAV research primarily focuses on RGB, infrared (IR), or RGB-IR videos captured by fixed ground cameras, with little attention to tracking target UAVs from another moving UAV platform. To fill this gap, we propose a new multi-modal visual tracking task termed UAV-Anti-UAV, which involves a pursuer UAV tracking a target adversarial UAV in the video stream. Compared to existing Anti-UAV tasks, UAV-Anti-UAV is more challenging due to severe dual-dynamic disturbances caused by the rapid motion of both the capturing platform and the target. To advance research in this domain, we construct a million-scale dataset consisting of 1,810 videos, each manually annotated with bounding boxes, a language prompt, and 15 tracking attributes. Furthermore, we propose MambaSTS, a Mamba-based baseline method for UAV-Anti-UAV tracking, which enables integrated spatial-temporal-semantic learning. Specifically, we employ Mamba and Transformer models to learn global semantic and spatial features, respectively, and leverage the state space model's strength in long-sequence modeling to establish video-level long-term context via a temporal token propagation mechanism. We conduct experiments on the UAV-Anti-UAV dataset to validate the effectiveness of our method. A thorough experimental evaluation of 50 modern deep tracking algorithms demonstrates that there is still significant room for improvement in the UAV-Anti-UAV domain. The dataset and codes will be available at {\color{magenta}https://github.com/983632847/Awesome-Multimodal-Object-Tracking}.

Xiangchi Yuan, Dachuan Shi, Chunhui Zhang et al.

Reinforcement learning (RL) is central to post-training, particularly for agentic models that require specialized reasoning behaviors. In this setting, model merging offers a practical mechanism for integrating multiple RL-trained agents from different tasks into a single generalist model. However, existing merging methods are designed for supervised fine-tuning (SFT), and they are suboptimal to preserve task-specific capabilities on RL-trained agentic models. The root is a task-vector mismatch between RL and SFT: on-policy RL induces task vectors that are highly sparse and heterogeneous, whereas SFT-style merging implicitly assumes dense and globally comparable task vectors. When standard global averaging is applied under this mismatch, RL's non-overlapping task vectors that encode critical task-specific behaviors are reduced and parameter updates are diluted. To address this issue, we propose Reinforced Agent Merging (RAM), a distribution-aware merging framework explicitly designed for RL-trained agentic models. RAM disentangles shared and task-specific unique parameter updates, averaging shared components while selectively preserving and rescaling unique ones to counteract parameter update dilution. Experiments across multiple agent domains and model architectures demonstrate that RAM not only surpasses merging baselines, but also unlocks synergistic potential among agents to achieve performance superior to that of specialized agents in their domains.

Xingjian Diao, Chunhui Zhang, Weiyi Wu et al.

Multimodal foundation models (MFMs) have demonstrated significant success in tasks such as visual captioning, question answering, and image-text retrieval. However, these models face inherent limitations due to their finite internal capacity, which restricts their ability to process extended temporal sequences, a crucial requirement for comprehensive video and audio analysis. To overcome these challenges, we introduce a specialized cognitive module, temporal working memory (TWM), which aims to enhance the temporal modeling capabilities of MFMs. It selectively retains task-relevant information across temporal dimensions, ensuring that critical details are preserved throughout the processing of video and audio content. The TWM uses a query-guided attention approach to focus on the most informative multimodal segments within temporal sequences. By retaining only the most relevant content, TWM optimizes the use of the model's limited capacity, enhancing its temporal modeling ability. This plug-and-play module can be easily integrated into existing MFMs. With our TWM, nine state-of-the-art models exhibit significant performance improvements across tasks such as video captioning, question answering, and video-text retrieval. By enhancing temporal modeling, TWM extends the capability of MFMs to handle complex, time-sensitive data effectively. Our code is available at https://github.com/xid32/NAACL_2025_TWM.

Xingjian Diao, Chunhui Zhang, Tingxuan Wu et al.

Music performances are representative scenarios for audio-visual modeling. Unlike common scenarios with sparse audio, music performances continuously involve dense audio signals throughout. While existing multimodal learning methods on the audio-video QA demonstrate impressive capabilities in general scenarios, they are incapable of dealing with fundamental problems within the music performances: they underexplore the interaction between the multimodal signals in performance and fail to consider the distinctive characteristics of instruments and music. Therefore, existing methods tend to answer questions regarding musical performances inaccurately. To bridge the above research gaps, (i) given the intricate multimodal interconnectivity inherent to music data, our primary backbone is designed to incorporate multimodal interactions within the context of music; (ii) to enable the model to learn music characteristics, we annotate and release rhythmic and music sources in the current music datasets; (iii) for time-aware audio-visual modeling, we align the model's music predictions with the temporal dimension. Our experiments show state-of-the-art effects on the Music AVQA datasets. Our code is available at https://github.com/xid32/Amuse.

Chunhui Zhang, Yawen Cui, Weilin Lin et al.

The recent wave of foundation models has witnessed tremendous success in computer vision (CV) and beyond, with the segment anything model (SAM) having sparked a passion for exploring task-agnostic visual foundation models. Empowered by its remarkable zero-shot generalization, SAM is currently challenging numerous traditional paradigms in CV, delivering extraordinary performance not only in various image segmentation and multi-modal segmentation (\eg, text-to-mask) tasks, but also in the video domain. Additionally, the latest released SAM 2 is once again sparking research enthusiasm in the realm of promptable visual segmentation for both images and videos. However, existing surveys mainly focus on SAM in various image processing tasks, a comprehensive and in-depth review in the video domain is notably absent. To address this gap, this work conducts a systematic review on SAM for videos in the era of foundation models. As the first to review the progress of SAM for videos, this work focuses on its applications to various tasks by discussing its recent advances, and innovation opportunities of developing foundation models on broad applications. We begin with a brief introduction to the background of SAM and video-related research domains. Subsequently, we present a systematic taxonomy that categorizes existing methods into three key areas: video understanding, video generation, and video editing, analyzing and summarizing their advantages and limitations. Furthermore, comparative results of SAM-based and current state-of-the-art methods on representative benchmarks, as well as insightful analysis are offered. Finally, we discuss the challenges faced by current research and envision several future research directions in the field of SAM for video and beyond.

Junqi Wang, Chunhui Zhang, Jiapeng Li et al.

Facing the current debate on whether Large Language Models (LLMs) attain near-human intelligence levels (Mitchell & Krakauer, 2023; Bubeck et al., 2023; Kosinski, 2023; Shiffrin & Mitchell, 2023; Ullman, 2023), the current study introduces a benchmark for evaluating social intelligence, one of the most distinctive aspects of human cognition. We developed a comprehensive theoretical framework for social dynamics and introduced two evaluation tasks: Inverse Reasoning (IR) and Inverse Inverse Planning (IIP). Our approach also encompassed a computational model based on recursive Bayesian inference, adept at elucidating diverse human behavioral patterns. Extensive experiments and detailed analyses revealed that humans surpassed the latest GPT models in overall performance, zero-shot learning, one-shot generalization, and adaptability to multi-modalities. Notably, GPT models demonstrated social intelligence only at the most basic order (order = 0), in stark contrast to human social intelligence (order >= 2). Further examination indicated a propensity of LLMs to rely on pattern recognition for shortcuts, casting doubt on their possession of authentic human-level social intelligence. Our codes, dataset, appendix and human data are released at https://github.com/bigai-ai/Evaluate-n-Model-Social-Intelligence.

Xingjian Diao, Chunhui Zhang, Keyi Kong et al.

While large language models have demonstrated impressive reasoning abilities, their extension to the audio modality, particularly within large audio-language models (LALMs), remains underexplored. Addressing this gap requires a systematic approach that involves a capable base model, high-quality reasoning-oriented audio data, and effective training algorithms. In this work, we present a comprehensive solution for audio logical reasoning (ALR) tasks: we introduce SoundMind, a dataset of 6,446 audio-text annotated samples specifically curated to support complex reasoning. Building on this resource, we propose SoundMind-RL, a rule-based reinforcement learning (RL) algorithm designed to equip audio-language models with robust audio-text reasoning capabilities. By fine-tuning Qwen2.5-Omni-7B on the proposed SoundMind dataset using SoundMind-RL, we achieve strong and consistent improvements over state-of-the-art baselines on the SoundMind benchmark. This work highlights the benefit of combining high-quality, reasoning-focused datasets with specialized RL techniques, and contributes to advancing auditory intelligence in language models. The code and dataset introduced in this work are publicly available at https://github.com/xid32/SoundMind.

Chunhui Zhang, Li Liu, Hao Wen et al.

Multi-modal object tracking (MMOT) is an emerging field that combines data from various modalities, \eg vision (RGB), depth, thermal infrared, event, language and audio, to estimate the state of an arbitrary object in a video sequence. It is of great significance for many applications such as autonomous driving and intelligent surveillance. In recent years, MMOT has received more and more attention. However, existing MMOT algorithms mainly focus on two modalities (\eg RGB+depth, RGB+thermal infrared, and RGB+language). To leverage more modalities, some recent efforts have been made to learn a unified visual object tracking model for any modality. Additionally, some large-scale multi-modal tracking benchmarks have been established by simultaneously providing more than two modalities, such as vision-language-audio (\eg WebUAV-3M) and vision-depth-language (\eg UniMod1K). To track the latest progress in MMOT, we conduct a comprehensive investigation in this report. Specifically, we first divide existing MMOT tasks into five main categories, \ie RGBL tracking, RGBE tracking, RGBD tracking, RGBT tracking, and miscellaneous (RGB+X), where X can be any modality, such as language, depth, and event. Then, we analyze and summarize each MMOT task, focusing on widely used datasets and mainstream tracking algorithms based on their technical paradigms (\eg self-supervised learning, prompt learning, knowledge distillation, generative models, and state space models). Finally, we maintain a continuously updated paper list for MMOT at https://github.com/983632847/Awesome-Multimodal-Object-Tracking.

Chunhui Zhang, Li Liu, Hao Wen et al.

Night unmanned aerial vehicle (UAV) tracking is impeded by the challenges of poor illumination, with previous daylight-optimized methods demonstrating suboptimal performance in low-light conditions, limiting the utility of UAV applications. To this end, we propose an efficient mamba-based tracker, leveraging dual enhancement techniques to boost night UAV tracking. The mamba-based low-light enhancer, equipped with an illumination estimator and a damage restorer, achieves global image enhancement while preserving the details and structure of low-light images. Additionally, we advance a cross-modal mamba network to achieve efficient interactive learning between vision and language modalities. Extensive experiments showcase that our method achieves advanced performance and exhibits significantly improved computation and memory efficiency. For instance, our method is 2.8$\times$ faster than CiteTracker and reduces 50.2$\%$ GPU memory. Our codes are available at \url{https://github.com/983632847/Awesome-Multimodal-Object-Tracking}.

Wenhao You, Xingjian Diao, Chunhui Zhang et al.

While recent Multimodal Large Language Models exhibit impressive capabilities for general multimodal tasks, specialized domains like music necessitate tailored approaches. Music Audio-Visual Question Answering (Music AVQA) particularly underscores this, presenting unique challenges with its continuous, densely layered audio-visual content, intricate temporal dynamics, and the critical need for domain-specific knowledge. Through a systematic analysis of Music AVQA datasets and methods, this position paper identifies that specialized input processing, architectures incorporating dedicated spatial-temporal designs, and music-specific modeling strategies are critical for success in this domain. Our study provides valuable insights for researchers by highlighting effective design patterns empirically linked to strong performance, proposing concrete future directions for incorporating musical priors, and aiming to establish a robust foundation for advancing multimodal musical understanding. This work is intended to inspire broader attention and further research, supported by a continuously updated anonymous GitHub repository of relevant papers: https://github.com/xid32/Survey4MusicAVQA.

Tianxin Xie, Wentao Lei, Guanjie Huang et al.

Text-to-audio-video (T2AV) generation underpins a wide range of applications demanding realistic audio-visual content, including virtual reality, world modeling, gaming, and filmmaking. However, existing T2AV models remain incapable of generating physically plausible sounds, primarily due to their limited understanding of physical principles. To situate current research progress, we present PhyAVBench, a challenging audio physics-sensitivity benchmark designed to systematically evaluate the audio physics grounding capabilities of existing T2AV models. PhyAVBench comprises 1,000 groups of paired text prompts with controlled physical variables that implicitly induce sound variations, enabling a fine-grained assessment of models' sensitivity to changes in underlying acoustic conditions. We term this evaluation paradigm the Audio-Physics Sensitivity Test (APST). Unlike prior benchmarks that primarily focus on audio-video synchronization, PhyAVBench explicitly evaluates models' understanding of the physical mechanisms underlying sound generation, covering 6 major audio physics dimensions, 4 daily scenarios (music, sound effects, speech, and their mix), and 50 fine-grained test points, ranging from fundamental aspects such as sound diffraction to more complex phenomena, e.g., Helmholtz resonance. Each test point consists of multiple groups of paired prompts, where each prompt is grounded by at least 20 newly recorded or collected real-world videos, thereby minimizing the risk of data leakage during model pre-training. Both prompts and videos are iteratively refined through rigorous human-involved error correction and quality control to ensure high quality. We argue that only models with a genuine grasp of audio-related physical principles can generate physically consistent audio-visual content. We hope PhyAVBench will stimulate future progress in this critical yet largely unexplored domain.

Yi-Long Lu, Chunhui Zhang, Wei Wang

Large Language Models (LLMs) are increasingly used in tasks such as psychological text analysis and decision-making in automated workflows. However, their reliability remains a concern due to potential biases inherited from their training process. In this study, we examine how different response format: binary versus continuous, may systematically influence LLMs' judgments. In a value statement judgments task and a text sentiment analysis task, we prompted LLMs to simulate human responses and tested both formats across several models, including both open-source and commercial models. Our findings revealed a consistent negative bias: LLMs were more likely to deliver "negative" judgments in binary formats compared to continuous ones. Control experiments further revealed that this pattern holds across both tasks. Our results highlight the importance of considering response format when applying LLMs to decision tasks, as small changes in task design can introduce systematic biases.

Chunhui Zhang, Li Liu, Yawen Cui et al.

Artificial intelligence (AI) is evolving towards artificial general intelligence, which refers to the ability of an AI system to perform a wide range of tasks and exhibit a level of intelligence similar to that of a human being. This is in contrast to narrow or specialized AI, which is designed to perform specific tasks with a high degree of efficiency. Therefore, it is urgent to design a general class of models, which we term foundation models, trained on broad data that can be adapted to various downstream tasks. The recently proposed segment anything model (SAM) has made significant progress in breaking the boundaries of segmentation, greatly promoting the development of foundation models for computer vision. To fully comprehend SAM, we conduct a survey study. As the first to comprehensively review the progress of segmenting anything task for vision and beyond based on the foundation model of SAM, this work focuses on its applications to various tasks and data types by discussing its historical development, recent progress, and profound impact on broad applications. We first introduce the background and terminology for foundation models including SAM, as well as state-of-the-art methods contemporaneous with SAM that are significant for segmenting anything task. Then, we analyze and summarize the advantages and limitations of SAM across various image processing applications, including software scenes, real-world scenes, and complex scenes. Importantly, many insights are drawn to guide future research to develop more versatile foundation models and improve the architecture of SAM. We also summarize massive other amazing applications of SAM in vision and beyond. Finally, we maintain a continuously updated paper list and an open-source project summary for foundation model SAM at \href{https://github.com/liliu-avril/Awesome-Segment-Anything}{\color{magenta}{here}}.

Chunhui Zhang, Guanjie Huang, Li Liu et al.

Unmanned aerial vehicle (UAV) tracking is of great significance for a wide range of applications, such as delivery and agriculture. Previous benchmarks in this area mainly focused on small-scale tracking problems while ignoring the amounts of data, types of data modalities, diversities of target categories and scenarios, and evaluation protocols involved, greatly hiding the massive power of deep UAV tracking. In this work, we propose WebUAV-3M, the largest public UAV tracking benchmark to date, to facilitate both the development and evaluation of deep UAV trackers. WebUAV-3M contains over 3.3 million frames across 4,500 videos and offers 223 highly diverse target categories. Each video is densely annotated with bounding boxes by an efficient and scalable semiautomatic target annotation (SATA) pipeline. Importantly, to take advantage of the complementary superiority of language and audio, we enrich WebUAV-3M by innovatively providing both natural language specifications and audio descriptions. We believe that such additions will greatly boost future research in terms of exploring language features and audio cues for multimodal UAV tracking. In addition, a fine-grained UAV tracking-under-scenario constraint (UTUSC) evaluation protocol and seven challenging scenario subtest sets are constructed to enable the community to develop, adapt and evaluate various types of advanced trackers. We provide extensive evaluations and detailed analyses of 43 representative trackers and envision future research directions in the field of deep UAV tracking and beyond. The dataset, toolkits and baseline results are available at \url{https://github.com/983632847/WebUAV-3M}.

Yixiong Chen, Chunhui Zhang, Li Liu et al.

Most deep neural networks (DNNs) based ultrasound (US) medical image analysis models use pretrained backbones (e.g., ImageNet) for better model generalization. However, the domain gap between natural and medical images causes an inevitable performance bottleneck. To alleviate this problem, an US dataset named US-4 is constructed for direct pretraining on the same domain. It contains over 23,000 images from four US video sub-datasets. To learn robust features from US-4, we propose an US semi-supervised contrastive learning method, named USCL, for pretraining. In order to avoid high similarities between negative pairs as well as mine abundant visual features from limited US videos, USCL adopts a sample pair generation method to enrich the feature involved in a single step of contrastive optimization. Extensive experiments on several downstream tasks show the superiority of USCL pretraining against ImageNet pretraining and other state-of-the-art (SOTA) pretraining approaches. In particular, USCL pretrained backbone achieves fine-tuning accuracy of over 94% on POCUS dataset, which is 10% higher than 84% of the ImageNet pretrained model. The source codes of this work are available at https://github.com/983632847/USCL.

Chunhui Zhang, Yuxuan Wang, Aoyang Qin et al.

Current embodied agents are often limited to passive instruction-following or reactive need-satisfaction, lacking a stable, high-order value framework essential for long-term, self-directed behavior and resolving motivational conflicts. We introduce \textit{ValuePlanner}, a hierarchical cognitive architecture that decouples high-level value scheduling from low-level action execution. \textit{ValuePlanner} employs an LLM-based cognitive module to generate symbolic subgoals by reasoning through abstract value trade-offs, which are then translated into executable action plans by a classical PDDL planner. This process is refined via a closed-loop feedback mechanism. Evaluating such autonomy requires methods beyond task-success rates, and we therefore propose a value-centric evaluation suite measuring cumulative value gain, preference alignment, and behavioral diversity. Experiments in the TongSim household environment demonstrate that \textit{ValuePlanner} arbitrates competing values to generate coherent, long-horizon, self-directed behavior absent from instruction-following and needs-driven baselines. Our work offers a structured approach to bridging intrinsic values and grounded behavior for autonomous agents.

Yaning Jia, Chunhui Zhang

The Lipschitz bound, a technique from robust statistics, can limit the maximum changes in the output concerning the input, taking into account associated irrelevant biased factors. It is an efficient and provable method for examining the output stability of machine learning models without incurring additional computation costs. Recently, Graph Neural Networks (GNNs), which operate on non-Euclidean data, have gained significant attention. However, no previous research has investigated the GNN Lipschitz bounds to shed light on stabilizing model outputs, especially when working on non-Euclidean data with inherent biases. Given the inherent biases in common graph data used for GNN training, it poses a serious challenge to constraining the GNN output perturbations induced by input biases, thereby safeguarding fairness during training. Recently, despite the Lipschitz constant's use in controlling the stability of Euclideanneural networks, the calculation of the precise Lipschitz constant remains elusive for non-Euclidean neural networks like GNNs, especially within fairness contexts. To narrow this gap, we begin with the general GNNs operating on an attributed graph, and formulate a Lipschitz bound to limit the changes in the output regarding biases associated with the input. Additionally, we theoretically analyze how the Lipschitz constant of a GNN model could constrain the output perturbations induced by biases learned from data for fairness training. We experimentally validate the Lipschitz bound's effectiveness in limiting biases of the model output. Finally, from a training dynamics perspective, we demonstrate why the theoretical Lipschitz bound can effectively guide the GNN training to better trade-off between accuracy and fairness.

Chunhui Zhang, Yiren Jian, Zhongyu Ouyang et al.

Developing video captioning models is computationally expensive. The dynamic nature of video also complicates the design of multimodal models that can effectively caption these sequences. However, we find that by using minimal computational resources and without complex modifications to address video dynamics, an image-based model can be repurposed to outperform several specialised video captioning systems. Our adapted model demonstrates top tier performance on major benchmarks, ranking 2nd on MSRVTT and MSVD, and 3rd on VATEX. We transform it into a competitive video captioner by post training a typical image captioning model BLIP2 with only 6,000 video text pairs and simply concatenating frames (significantly fewer data than other methods), which use 2.5 to 144 million pairs. From a resource optimization perspective, this video captioning study focuses on three fundamental factors: optimizing model scale, maximizing data efficiency, and incorporating reinforcement learning. This extensive study demonstrates that a lightweight, image based adaptation strategy can rival state-of-the-art video captioning systems, offering a practical solution for low-resource scenarios.

Xiangchi Yuan, Chunhui Zhang, Zheyuan Liu et al.

As scaled language models (LMs) approach human-level reasoning capabilities, self-improvement emerges as a solution to synthesizing high-quality data corpus. While previous research has identified model collapse as a risk in self-improvement, where model outputs become increasingly deterministic, we discover a more fundamental challenge: the superficial self-improved reasoners phenomenon. In particular, our analysis reveals that even when LMs show improved in-domain (ID) reasoning accuracy, they actually compromise their generalized reasoning capabilities on out-of-domain (OOD) tasks due to memorization rather than genuine. Through a systematic investigation of LM architecture, we discover that during self-improvement, LM weight updates are concentrated in less reasoning-critical layers, leading to superficial learning. To address this, we propose Iterative Model Merging (IMM), a method that strategically combines weights from original and self-improved models to preserve generalization while incorporating genuine reasoning improvements. Our approach effectively mitigates both LM collapse and superficial learning, moving towards more stable self-improving systems.

Xingjian Diao, Tianzhen Yang, Chunhui Zhang et al.

Music performances, characterized by dense and continuous audio as well as seamless audio-visual integration, present unique challenges for multimodal scene understanding and reasoning. Recent Music Performance Audio-Visual Question Answering (Music AVQA) datasets have been proposed to reflect these challenges, highlighting the continued need for more effective integration of audio-visual representations in complex question answering. However, existing Music AVQA methods often rely on dense and unoptimized representations, leading to inefficiencies in the isolation of key information, the reduction of redundancy, and the prioritization of critical samples. To address these challenges, we introduce Sparsify, a sparse learning framework specifically designed for Music AVQA. It integrates three sparsification strategies into an end-to-end pipeline and achieves state-of-the-art performance on the Music AVQA datasets. In addition, it reduces training time by 28.32% compared to its fully trained dense counterpart while maintaining accuracy, demonstrating clear efficiency gains. To further improve data efficiency, we propose a key-subset selection algorithm that selects and uses approximately 25% of MUSIC-AVQA v2.0 training data and retains 70-80% of full-data performance across models.

Chunhui Zhang, Sirui, Wang et al.

Language models (LMs) require robust episodic grounding-the capacity to learn from and apply past experiences-to excel at physical planning tasks. Current episodic grounding approaches struggle with scalability and integration, limiting their effectiveness, especially for medium-sized LMs (7B parameters). While larger LMs (70-405B parameters) possess superior hierarchical representations and extensive pre-trained knowledge, they encounter a fundamental scale paradox: despite their advanced abstraction capabilities, they lack efficient mechanisms to leverage experience streams. We propose a scalable weak-to-strong episodic learning framework that effectively transfers episodic behaviors from smaller to larger LMs. This framework integrates Monte Carlo tree search for structured experience collection with a novel distillation method, preserving the inherent LM capabilities while embedding episodic memory. Experiments demonstrate our method surpasses state-of-the-art proprietary LMs by 3.45% across diverse planning and question-answering tasks. Layer-wise probing further indicates significant improvements in task alignment, especially within deeper LM layers, highlighting stable generalization even for previously unseen scenarios with increased planning complexity-conditions where baseline methods degrade markedly.

Zheyuan Liu, Guangyao Dou, Xiangchi Yuan et al.

Generative models such as Large Language Models (LLMs) and Multimodal Large Language Models (MLLMs) trained on massive datasets can lead them to memorize and inadvertently reveal sensitive information, raising ethical and privacy concerns. While some prior works have explored this issue in the context of LLMs, it presents a unique challenge for MLLMs due to the entangled nature of knowledge across modalities, making comprehensive unlearning more difficult. To address this challenge, we propose Modality Aware Neuron Unlearning (MANU), a novel unlearning framework for MLLMs designed to selectively clip neurons based on their relative importance to the targeted forget data, curated for different modalities. Specifically, MANU consists of two stages: important neuron selection and selective pruning. The first stage identifies and collects the most influential neurons across modalities relative to the targeted forget knowledge, while the second stage is dedicated to pruning those selected neurons. MANU effectively isolates and removes the neurons that contribute most to the forget data within each modality, while preserving the integrity of retained knowledge. Our experiments conducted across various MLLM architectures illustrate that MANU can achieve a more balanced and comprehensive unlearning in each modality without largely affecting the overall model utility.

Ivory Yang, Weicheng Ma, Chunhui Zhang et al.

Endangered languages, such as Navajo - the most widely spoken Native American language - are significantly underrepresented in contemporary language technologies, exacerbating the challenges of their preservation and revitalization. This study evaluates Google's Language Identification (LangID) tool, which does not currently support any Native American languages. To address this, we introduce a random forest classifier trained on Navajo and twenty erroneously suggested languages by LangID. Despite its simplicity, the classifier achieves near-perfect accuracy (97-100%). Additionally, the model demonstrates robustness across other Athabaskan languages - a family of Native American languages spoken primarily in Alaska, the Pacific Northwest, and parts of the Southwestern United States - suggesting its potential for broader application. Our findings underscore the pressing need for NLP systems that prioritize linguistic diversity and adaptability over centralized, one-size-fits-all solutions, especially in supporting underrepresented languages in a multicultural world. This work directly contributes to ongoing efforts to address cultural biases in language models and advocates for the development of culturally localized NLP tools that serve diverse linguistic communities.

Chunhui Zhang, Zhongyu Ouyang, Kwonjoon Lee et al. · mit

Theory-of-Mind (ToM) enables humans to infer mental states-such as beliefs, desires, and intentions-forming the foundation of social cognition. However, existing computational ToM methods rely on structured workflows with ToM-specific priors or deep model fine-tuning, which struggle with scalability in multimodal environments and fail to generalize as task complexity increases. To address these limitations, we propose a scalable Bayesian ToM planner that decomposes ToM reasoning into stepwise Bayesian updates. Our framework introduces weak-to-strong control, allowing smaller language models (LMs) to specialize in ToM-specific likelihood estimation and transfer their reasoning behaviors to larger LMs (7B to 405B) for integration with social and world knowledge. This synergistic approach aligns large-model inference of human mental states with Bayesian principles. Extensive experiments show that our method achieves a 4.6% accuracy improvement over state-of-the-art techniques on multimodal ToM benchmarks, including challenging unseen scenarios, thereby establishing a new standard for modeling human mental states in complex environments.

Yi-Long Lu, Chunhui Zhang, Jiajun Song et al.

Theory of Mind (ToM), the ability to attribute mental states to others, is fundamental for human social intelligence and a critical capability for advanced Artificial Intelligence. Recent advancements in Large Language Models (LLMs) have shown promising performance on ToM benchmarks, raising the question: Do these benchmarks necessitate explicit human-like reasoning processes, or can models succeed through alternative strategies? We investigate this question empirically by applying Reinforcement Learning (RL) and Supervised Fine-Tuning (SFT) to LLMs of varying scales (0.5B to 7B parameters) and evaluating them across multiple ToM datasets. Our results reveal a scale-dependent impact of RL: while RL significantly improves accuracy and fosters high-quality, interpretable, and transferable belief-tracking reasoning in larger models (7B), it leads to "reasoning collapse" in smaller models ($\leq$3B), where high accuracy and generalization ability are achieved via drastically shortened, less meaningful responses. Surprisingly, further SFT achieves competitive and generalizable performance across these benchmarks, often matching or exceeding RL models in accuracy, despite not being explicitly trained to produce structured reasoning traces. These findings highlight a critical discrepancy between benchmark accuracy and the nature of learned reasoning. Our work suggests that current ToM benchmarks may be solvable without requiring the explicit, human-like simulation of mental states they were designed to probe. LLMs, particularly when scale is limited or training signals focus solely on output correctness, may leverage alternative rules effective for benchmark data structures.

Chunhui Zhang, Li Liu, Jialin Gao et al.

Transformer has recently demonstrated great potential in improving vision-language (VL) tracking algorithms. However, most of the existing VL trackers rely on carefully designed mechanisms to perform the multi-stage multi-modal fusion. Additionally, direct multi-modal fusion without alignment ignores distribution discrepancy between modalities in feature space, potentially leading to suboptimal representations. In this work, we propose COST, a contrastive one-stage transformer fusion framework for VL tracking, aiming to learn semantically consistent and unified VL representations. Specifically, we introduce a contrastive alignment strategy that maximizes mutual information (MI) between a video and its corresponding language description. This enables effective cross-modal alignment, yielding semantically consistent features in the representation space. By leveraging a visual-linguistic transformer, we establish an efficient multi-modal fusion and reasoning mechanism, empirically demonstrating that a simple stack of transformer encoders effectively enables unified VL representations. Moreover, we contribute a newly collected VL tracking benchmark dataset for small object tracking, named VL-SOT500, with bounding boxes and language descriptions. Our dataset comprises two challenging subsets, VL-SOT230 and VL-SOT270, dedicated to evaluating generic and high-speed small object tracking, respectively. Small object tracking is notoriously challenging due to weak appearance and limited features, and this dataset is, to the best of our knowledge, the first to explore the usage of language cues to enhance visual representation for small object tracking. Extensive experiments demonstrate that COST achieves state-of-the-art performance on five existing VL tracking datasets, as well as on our proposed VL-SOT500 dataset. Source codes and dataset will be made publicly available.

Zhongyu Ouyang, Chunhui Zhang, Yaning Jia et al.

In modern social media, recommender systems (RecSys) rely on the click-through rate (CTR) as the standard metric to evaluate user engagement. CTR prediction is traditionally framed as a binary classification task to predict whether a user will interact with a given item. However, this approach overlooks the complexity of real-world social modeling, where the user, item, and their interactive features change dynamically in fast-paced online environments. This dynamic nature often leads to model instability, reflected in overfitting short-term fluctuations rather than higher-level interactive patterns. While overfitting calls for more scaled and refined supervisions, current solutions often rely on binary labels that overly simplify fine-grained user preferences through the thresholding process, which significantly reduces the richness of the supervision. Therefore, we aim to alleviate the overfitting problem by increasing the supervision bandwidth in CTR training. Specifically, (i) theoretically, we formulate the impact of fine-grained preferences on model stability as a Lipschitz constrain; (ii) empirically, we discover that scaling the supervision bandwidth can act as an implicit Lipschitz regularizer, stably optimizing existing CTR models to achieve better generalizability. Extensive experiments show that this scaled supervision significantly and consistently improves the optimization process and the performance of existing CTR models, even without the need for additional hyperparameter tuning.

Xingjian Diao, Zheyuan Liu, Chunhui Zhang et al.

Large Vision-Language Models (LVLMs) have exhibited strong reasoning capabilities through chain-of-thought mechanisms that generate step-by-step rationales. However, such slow-thinking approaches often lead to overthinking, where models produce excessively verbose responses even for simple queries, resulting in test-time inefficiency and even degraded accuracy. Prior work has attempted to mitigate this issue via adaptive reasoning strategies, but these methods largely overlook a fundamental bottleneck: visual perception failures. We argue that stable reasoning critically depends on low-level visual grounding, and that reasoning errors often originate from imperfect perception rather than insufficient deliberation. To address this limitation, we propose Gated Perception-Reasoning Optimization (GPRO), a meta-reasoning controller that dynamically routes computation among three decision paths at each generation step: a lightweight fast path, a slow perception path for re-examining visual inputs, and a slow reasoning path for internal self-reflection. To learn this distinction, we derive large-scale failure attribution supervision from approximately 790k samples, using teacher models to distinguish perceptual hallucinations from reasoning errors. We then train the controller with multi-objective reinforcement learning to optimize the trade-off between task accuracy and computational cost under uncertainty. Experiments on five benchmarks demonstrate that GPRO substantially improves both accuracy and efficiency, outperforming recent slow-thinking methods while generating significantly shorter responses.

Yaning Jia, Chunhui Zhang, Xingjian Diao et al.

Curriculum learning (CL) - ordering training data from easy to hard - has become a popular strategy for improving reasoning in large language models (LLMs). Yet prior work employs disparate difficulty metrics and training setups, leaving open fundamental questions: When does curriculum help? Which direction - forward or reverse - is better? And does the answer depend on what we measure? We address these questions through a unified offline evaluation framework that decomposes curriculum difficulty into five complementary dimensions: Problem Difficulty, Model Surprisal, Confidence Margin, Predictive Uncertainty, and Decision Variability. Through controlled post-training experiments on mathematical reasoning benchmarks with Llama3.1-8B, Mistral-7B, and Gemma3-4B, we find that (i) no curriculum strategy dominates universally - the relative effectiveness of forward versus reverse CL depends jointly on model capability and task complexity; (ii) even within a single metric, samples at different difficulty levels produce distinct gains depending on task demands; and (iii) task-aligned curricula focus on shaping the model's final representations and generalization, whereas inner-state curricula modulate internal states such as confidence and uncertainty. Our findings challenge the notion of a universal curriculum strategy and offer actionable guidance across model and task regimes, with some metrics indicating that prioritizing decision-uncertain samples can further enhance learning outcomes.

Lianjun Liao, Chunhui Zhang, Tong Wu et al.

Image data captured outdoors often exhibit unbounded scenes and unconstrained, varying lighting conditions, making it challenging to decompose them into geometry, reflectance, and illumination. Recent works have focused on achieving this decomposition using Neural Radiance Fields (NeRF) or the 3D Gaussian Splatting (3DGS) representation but remain hindered by two key limitations: the high computational overhead associated with neural networks of NeRF and the use of low-frequency lighting representations, which often result in inefficient rendering and suboptimal relighting accuracy. We propose ROSGS, a two-stage pipeline designed to efficiently reconstruct relightable outdoor scenes using the Gaussian Splatting representation. By leveraging monocular normal priors, ROSGS first reconstructs the scene's geometry with the compact 2D Gaussian Splatting (2DGS) representation, providing an efficient and accurate geometric foundation. Building upon this reconstructed geometry, ROSGS then decomposes the scene's texture and lighting through a hybrid lighting model. This model effectively represents typical outdoor lighting by employing a spherical Gaussian function to capture the directional, high-frequency components of sunlight, while learning a radiance transfer function via Spherical Harmonic coefficients to model the remaining low-frequency skylight comprehensively. Both quantitative metrics and qualitative comparisons demonstrate that ROSGS achieves state-of-the-art performance in relighting outdoor scenes and highlight its ability to deliver superior relighting accuracy and rendering efficiency.

Yi-Long Lu, Jiajun Song, Chunhui Zhang et al.

Humans constantly generate a diverse range of tasks guided by internal motivations. While generative agents powered by large language models (LLMs) aim to simulate this complex behavior, it remains uncertain whether they operate on similar cognitive principles. To address this, we conducted a task-generation experiment comparing human responses with those of an LLM agent (GPT-4o). We find that human task generation is consistently influenced by psychological drivers, including personal values (e.g., Openness to Change) and cognitive style. Even when these psychological drivers are explicitly provided to the LLM, it fails to reflect the corresponding behavioral patterns. They produce tasks that are markedly less social, less physical, and thematically biased toward abstraction. Interestingly, while the LLM's tasks were perceived as more fun and novel, this highlights a disconnect between its linguistic proficiency and its capacity to generate human-like, embodied goals. We conclude that there is a core gap between the value-driven, embodied nature of human cognition and the statistical patterns of LLMs, highlighting the necessity of incorporating intrinsic motivation and physical grounding into the design of more human-aligned agents.

Zhongyu Ouyang, Qianlong Wen, Chunhui Zhang et al.

Sequential recommendation systems aspire to profile users by interpreting their interaction histories, echoing how humans make decisions by weighing experience, relative preference strength, and situational relevance. Yet, existing large language model (LLM)-based recommenders often fall short of mimicking the flexible, context-aware decision strategies humans exhibit, neglecting the structured, dynamic, and context-aware mechanisms fundamental to human behaviors. To bridge this gap, we propose RecPO, a preference optimization framework that models structured feedback and contextual delay to emulate human-like prioritization in sequential recommendation. RecPO exploits adaptive reward margins based on inferred preference hierarchies and temporal signals, enabling the model to favor immediately relevant items and to distinguish between varying degrees of preference and aversion. Extensive experiments across five real-world datasets demonstrate that RecPO not only yields performance gains over state-of-the-art baselines, but also mirrors key characteristics of human decision-making: favoring timely satisfaction, maintaining coherent preferences, and exercising discernment under shifting contexts.

Chenggian Ma, Xiangyu Zhao, Chunhui Zhang et al.

With the development of Large Language Models (LLM), numerous prompts have been proposed, each with a rich set of features and their own merits. This paper summarizes the prompt words for large language models (LLMs), categorizing them into stimulating and framework types, and proposes an Auto-Prompt Graphical Paradigm(APGP) that combines both stimulating and framework prompts to enhance the problem-solving capabilities of LLMs across multiple domains, then exemplifies it with a framework that adheres to this paradigm. The framework involves automated prompt generation and consideration of emotion-stimulus factors, guiding LLMs in problem abstraction, diversified solutions generation, comprehensive optimization, and self-verification after providing answers, ensuring solution accuracy. Compared to traditional stimuli and framework prompts, this framework integrates the advantages of both by adopting automated approaches inspired by APE work, overcoming the limitations of manually designed prompts. Test results on the ruozhiba and BBH datasets demonstrate that this framework can effectively improve the efficiency and accuracy of LLMs in problem-solving, paving the way for new applications of LLMs.

Chunhui Zhang, Xiaoming Yuan, Qianyun Zhang et al.

Neural networks often encounter various stringent resource constraints while deploying on edge devices. To tackle these problems with less human efforts, automated machine learning becomes popular in finding various neural architectures that fit diverse Artificial Intelligence of Things (AIoT) scenarios. Recently, to prevent the leakage of private information while enable automated machine intelligence, there is an emerging trend to integrate federated learning and neural architecture search (NAS). Although promising as it may seem, the coupling of difficulties from both tenets makes the algorithm development quite challenging. In particular, how to efficiently search the optimal neural architecture directly from massive non-independent and identically distributed (non-IID) data among AIoT devices in a federated manner is a hard nut to crack. In this paper, to tackle this challenge, by leveraging the advances in ProxylessNAS, we propose a Federated Direct Neural Architecture Search (FDNAS) framework that allows for hardware-friendly NAS from non- IID data across devices. To further adapt to both various data distributions and different types of devices with heterogeneous embedded hardware platforms, inspired by meta-learning, a Cluster Federated Direct Neural Architecture Search (CFDNAS) framework is proposed to achieve device-aware NAS, in the sense that each device can learn a tailored deep learning model for its particular data distribution and hardware constraint. Extensive experiments on non-IID datasets have shown the state-of-the-art accuracy-efficiency trade-offs achieved by the proposed solution in the presence of both data and device heterogeneity.

Kangkai Zhang, Chunhui Zhang, Shikun Li et al.

Knowledge distillation provides an effective way to transfer knowledge via teacher-student learning, where most existing distillation approaches apply a fixed pre-trained model as teacher to supervise the learning of student network. This manner usually brings in a big capability gap between teacher and student networks during learning. Recent researches have observed that a small teacher-student capability gap can facilitate knowledge transfer. Inspired by that, we propose an evolutionary knowledge distillation approach to improve the transfer effectiveness of teacher knowledge. Instead of a fixed pre-trained teacher, an evolutionary teacher is learned online and consistently transfers intermediate knowledge to supervise student network learning on-the-fly. To enhance intermediate knowledge representation and mimicking, several simple guided modules are introduced between corresponding teacher-student blocks. In this way, the student can simultaneously obtain rich internal knowledge and capture its growth process, leading to effective student network learning. Extensive experiments clearly demonstrate the effectiveness of our approach as well as good adaptability in the low-resolution and few-sample visual recognition scenarios.

Chunhui Zhang, Yongyuan Liang, Xiaoming Yuan et al.

To prevent the leakage of private information while enabling automated machine intelligence, there is an emerging trend to integrate federated learning and Neural Architecture Search (NAS). Although promising as it may seem, the coupling of difficulties from both two tenets makes the algorithm development quite challenging. In particular, how to efficiently search the optimal neural architecture directly from massive non-iid data of clients in a federated manner remains to be a hard nut to crack. To tackle this challenge, in this paper, by leveraging the advances in proxy-less NAS, we propose a Federated Direct Neural Architecture Search (FDNAS) framework that allows hardware-aware NAS from decentralized non-iid data of clients. To further adapt for various data distributions of clients, inspired by meta-learning, a cluster Federated Direct Neural Architecture Search (CFDNAS) framework is proposed to achieve client-aware NAS, in the sense that each client can learn a tailored deep learning model for its particular data distribution. Extensive experiments on real-world non-iid datasets show state-of-the-art accuracy-efficiency trade-offs for various hardware and data distributions of clients. Our codes will be released publicly upon paper acceptance.