Linchao Zhu

Juncheng Li, Junlin Xie, Linchao Zhu et al. · cmu

Understanding human emotions is a crucial ability for intelligent robots to provide better human-robot interactions. The existing works are limited to trimmed video-level emotion classification, failing to locate the temporal window corresponding to the emotion. In this paper, we introduce a new task, named Temporal Emotion Localization in videos~(TEL), which aims to detect human emotions and localize their corresponding temporal boundaries in untrimmed videos with aligned subtitles. TEL presents three unique challenges compared to temporal action localization: 1) The emotions have extremely varied temporal dynamics; 2) The emotion cues are embedded in both appearances and complex plots; 3) The fine-grained temporal annotations are complicated and labor-intensive. To address the first two challenges, we propose a novel dilated context integrated network with a coarse-fine two-stream architecture. The coarse stream captures varied temporal dynamics by modeling multi-granularity temporal contexts. The fine stream achieves complex plots understanding by reasoning the dependency between the multi-granularity temporal contexts from the coarse stream and adaptively integrates them into fine-grained video segment features. To address the third challenge, we introduce a cross-modal consensus learning paradigm, which leverages the inherent semantic consensus between the aligned video and subtitle to achieve weakly-supervised learning. We contribute a new testing set with 3,000 manually-annotated temporal boundaries so that future research on the TEL problem can be quantitatively evaluated. Extensive experiments show the effectiveness of our approach on temporal emotion localization. The repository of this work is at https://github.com/YYJMJC/Temporal-Emotion-Localization-in-Videos.

Juncheng Li, Xin He, Longhui Wei et al.

Large-scale vision-language pre-training has shown impressive advances in a wide range of downstream tasks. Existing methods mainly model the cross-modal alignment by the similarity of the global representations of images and texts, or advanced cross-modal attention upon image and text features. However, they fail to explicitly learn the fine-grained semantic alignment between visual regions and textual phrases, as only global image-text alignment information is available. In this paper, we introduce LOUPE, a fine-grained semantically aLigned visiOn-langUage PrE-training framework, which learns fine-grained semantic alignment from the novel perspective of game-theoretic interactions. To efficiently compute the game-theoretic interactions, we further propose an uncertainty-aware neural Shapley interaction learning module. Experiments show that LOUPE achieves state-of-the-art performance on a variety of vision-language tasks. Furthermore, without any object-level human annotations and fine-tuning, LOUPE achieves competitive performance on object detection and visual grounding. More importantly, LOUPE opens a new promising direction of learning fine-grained semantics from large-scale raw image-text pairs. The repository of this work is at https://github.com/YYJMJC/LOUPE.

Shuai Zhao, Linchao Zhu, Xiaohan Wang et al.

Recently, large-scale pre-training methods like CLIP have made great progress in multi-modal research such as text-video retrieval. In CLIP, transformers are vital for modeling complex multi-modal relations. However, in the vision transformer of CLIP, the essential visual tokenization process, which produces discrete visual token sequences, generates many homogeneous tokens due to the redundancy nature of consecutive and similar frames in videos. This significantly increases computation costs and hinders the deployment of video retrieval models in web applications. In this paper, to reduce the number of redundant video tokens, we design a multi-segment token clustering algorithm to find the most representative tokens and drop the non-essential ones. As the frame redundancy occurs mostly in consecutive frames, we divide videos into multiple segments and conduct segment-level clustering. Center tokens from each segment are later concatenated into a new sequence, while their original spatial-temporal relations are well maintained. We instantiate two clustering algorithms to efficiently find deterministic medoids and iteratively partition groups in high dimensional space. Through this token clustering and center selection procedure, we successfully reduce computation costs by removing redundant visual tokens. This method further enhances segment-level semantic alignment between video and text representations, enforcing the spatio-temporal interactions of tokens from within-segment frames. Our method, coined as CenterCLIP, surpasses existing state-of-the-art by a large margin on typical text-video benchmarks, while reducing the training memory cost by 35\% and accelerating the inference speed by 14\% at the best case. The code is available at \href{https://github.com/mzhaoshuai/CenterCLIP}{https://github.com/mzhaoshuai/CenterCLIP}.

Fan Ma, Mike Zheng Shou, Linchao Zhu et al.

As an important area in computer vision, object tracking has formed two separate communities that respectively study Single Object Tracking (SOT) and Multiple Object Tracking (MOT). However, current methods in one tracking scenario are not easily adapted to the other due to the divergent training datasets and tracking objects of both tasks. Although UniTrack \cite{wang2021different} demonstrates that a shared appearance model with multiple heads can be used to tackle individual tracking tasks, it fails to exploit the large-scale tracking datasets for training and performs poorly on single object tracking. In this work, we present the Unified Transformer Tracker (UTT) to address tracking problems in different scenarios with one paradigm. A track transformer is developed in our UTT to track the target in both SOT and MOT. The correlation between the target and tracking frame features is exploited to localize the target. We demonstrate that both SOT and MOT tasks can be solved within this framework. The model can be simultaneously end-to-end trained by alternatively optimizing the SOT and MOT objectives on the datasets of individual tasks. Extensive experiments are conducted on several benchmarks with a unified model trained on SOT and MOT datasets. Code will be available at https://github.com/Flowerfan/Trackron.

Shuai Zhao, Linchao Zhu, Xiaohan Wang et al.

Self-supervised learning makes significant progress in pre-training large models, but struggles with small models. Mainstream solutions to this problem rely mainly on knowledge distillation, which involves a two-stage procedure: first training a large teacher model and then distilling it to improve the generalization ability of smaller ones. In this work, we introduce another one-stage solution to obtain pre-trained small models without the need for extra teachers, namely, slimmable networks for contrastive self-supervised learning (SlimCLR). A slimmable network consists of a full network and several weight-sharing sub-networks, which can be pre-trained once to obtain various networks, including small ones with low computation costs. However, interference between weight-sharing networks leads to severe performance degradation in self-supervised cases, as evidenced by gradient magnitude imbalance and gradient direction divergence. The former indicates that a small proportion of parameters produce dominant gradients during backpropagation, while the main parameters may not be fully optimized. The latter shows that the gradient direction is disordered, and the optimization process is unstable. To address these issues, we introduce three techniques to make the main parameters produce dominant gradients and sub-networks have consistent outputs. These techniques include slow start training of sub-networks, online distillation, and loss re-weighting according to model sizes. Furthermore, theoretical results are presented to demonstrate that a single slimmable linear layer is sub-optimal during linear evaluation. Thus a switchable linear probe layer is applied during linear evaluation. We instantiate SlimCLR with typical contrastive learning frameworks and achieve better performance than previous arts with fewer parameters and FLOPs. The code is at https://github.com/mzhaoshuai/SlimCLR.

Difei Gao, Luowei Zhou, Lei Ji et al.

To build Video Question Answering (VideoQA) systems capable of assisting humans in daily activities, seeking answers from long-form videos with diverse and complex events is a must. Existing multi-modal VQA models achieve promising performance on images or short video clips, especially with the recent success of large-scale multi-modal pre-training. However, when extending these methods to long-form videos, new challenges arise. On the one hand, using a dense video sampling strategy is computationally prohibitive. On the other hand, methods relying on sparse sampling struggle in scenarios where multi-event and multi-granularity visual reasoning are required. In this work, we introduce a new model named Multi-modal Iterative Spatial-temporal Transformer (MIST) to better adapt pre-trained models for long-form VideoQA. Specifically, MIST decomposes traditional dense spatial-temporal self-attention into cascaded segment and region selection modules that adaptively select frames and image regions that are closely relevant to the question itself. Visual concepts at different granularities are then processed efficiently through an attention module. In addition, MIST iteratively conducts selection and attention over multiple layers to support reasoning over multiple events. The experimental results on four VideoQA datasets, including AGQA, NExT-QA, STAR, and Env-QA, show that MIST achieves state-of-the-art performance and is superior at computation efficiency and interpretability.

Yaowei Li, Ruijie Quan, Linchao Zhu et al.

Large-scale pre-training has brought unimodal fields such as computer vision and natural language processing to a new era. Following this trend, the size of multi-modal learning models constantly increases, leading to an urgent need to reduce the massive computational cost of finetuning these models for downstream tasks. In this paper, we propose an efficient and flexible multimodal fusion method, namely PMF, tailored for fusing unimodally pre-trained transformers. Specifically, we first present a modular multimodal fusion framework that exhibits high flexibility and facilitates mutual interactions among different modalities. In addition, we disentangle vanilla prompts into three types in order to learn different optimizing objectives for multimodal learning. It is also worth noting that we propose to add prompt vectors only on the deep layers of the unimodal transformers, thus significantly reducing the training memory usage. Experiment results show that our proposed method achieves comparable performance to several other multimodal finetuning methods with less than 3% trainable parameters and up to 66% saving of training memory usage.

Juncheng Li, Junlin Xie, Long Qian et al.

Temporal grounding in videos aims to localize one target video segment that semantically corresponds to a given query sentence. Thanks to the semantic diversity of natural language descriptions, temporal grounding allows activity grounding beyond pre-defined classes and has received increasing attention in recent years. The semantic diversity is rooted in the principle of compositionality in linguistics, where novel semantics can be systematically described by combining known words in novel ways (compositional generalization). However, current temporal grounding datasets do not specifically test for the compositional generalizability. To systematically measure the compositional generalizability of temporal grounding models, we introduce a new Compositional Temporal Grounding task and construct two new dataset splits, i.e., Charades-CG and ActivityNet-CG. Evaluating the state-of-the-art methods on our new dataset splits, we empirically find that they fail to generalize to queries with novel combinations of seen words. To tackle this challenge, we propose a variational cross-graph reasoning framework that explicitly decomposes video and language into multiple structured hierarchies and learns fine-grained semantic correspondence among them. Experiments illustrate the superior compositional generalizability of our approach. The repository of this work is at https://github.com/YYJMJC/ Compositional-Temporal-Grounding.

Wei Li, Linchao Zhu, Longyin Wen et al.

Large-scale pre-trained multi-modal models (e.g., CLIP) demonstrate strong zero-shot transfer capability in many discriminative tasks. Their adaptation to zero-shot image-conditioned text generation tasks has drawn increasing interest. Prior arts approach to zero-shot captioning by either utilizing the existing large language models (e.g., GPT-2) or pre-training the encoder-decoder network in an end-to-end manner. In this work, we propose a simple framework, named DeCap, for zero-shot captioning. We introduce a lightweight visual-aware language decoder. This decoder is both data-efficient and computation-efficient: 1) it only requires the text data for training, easing the burden on the collection of paired data. 2) it does not require end-to-end training. When trained with text-only data, the decoder takes the text embedding extracted from the off-the-shelf CLIP encoder as a prefix embedding. The challenge is that the decoder is trained on the text corpus but at the inference stage, it needs to generate captions based on visual inputs. The modality gap issue is widely observed in multi-modal contrastive models that prevents us from directly taking the visual embedding as the prefix embedding. We propose a training-free mechanism to reduce the modality gap. We project the visual embedding into the CLIP text embedding space, while the projected embedding retains the information of the visual input. Taking the projected embedding as the prefix embedding, the decoder generates high-quality descriptions that match the visual input. The experiments show that DeCap outperforms other zero-shot captioning methods and unpaired captioning methods on the typical image captioning benchmarks, i.e., MSCOCO and NoCaps.

Fan Ma, Xiaojie Jin, Heng Wang et al.

Video-Language Pre-training models have recently significantly improved various multi-modal downstream tasks. Previous dominant works mainly adopt contrastive learning to achieve global feature alignment across modalities. However, the local associations between videos and texts are not modeled, restricting the pre-training models' generality, especially for tasks requiring the temporal video boundary for certain query texts. This work introduces a novel text-video localization pre-text task to enable fine-grained temporal and semantic alignment such that the trained model can accurately perceive temporal boundaries in videos given the text description. Specifically, text-video localization consists of moment retrieval, which predicts start and end boundaries in videos given the text description, and text localization which matches the subset of texts with the video features. To produce temporal boundaries, frame features in several videos are manually merged into a long video sequence that interacts with a text sequence. With the localization task, our method connects the fine-grained frame representations with the word representations and implicitly distinguishes representations of different instances in the single modality. Notably, comprehensive experimental results show that our method significantly improves the state-of-the-art performance on various benchmarks, covering text-to-video retrieval, video question answering, video captioning, temporal action localization and temporal moment retrieval. The code will be released soon.

Yu Lu, Yuanzhi Liang, Linchao Zhu et al.

Video diffusion models have made substantial progress in various video generation applications. However, training models for long video generation tasks require significant computational and data resources, posing a challenge to developing long video diffusion models. This paper investigates a straightforward and training-free approach to extend an existing short video diffusion model (e.g. pre-trained on 16-frame videos) for consistent long video generation (e.g. 128 frames). Our preliminary observation has found that directly applying the short video diffusion model to generate long videos can lead to severe video quality degradation. Further investigation reveals that this degradation is primarily due to the distortion of high-frequency components in long videos, characterized by a decrease in spatial high-frequency components and an increase in temporal high-frequency components. Motivated by this, we propose a novel solution named FreeLong to balance the frequency distribution of long video features during the denoising process. FreeLong blends the low-frequency components of global video features, which encapsulate the entire video sequence, with the high-frequency components of local video features that focus on shorter subsequences of frames. This approach maintains global consistency while incorporating diverse and high-quality spatiotemporal details from local videos, enhancing both the consistency and fidelity of long video generation. We evaluated FreeLong on multiple base video diffusion models and observed significant improvements. Additionally, our method supports coherent multi-prompt generation, ensuring both visual coherence and seamless transitions between scenes.

Juncheng Li, Siliang Tang, Linchao Zhu et al.

Temporal grounding is the task of locating a specific segment from an untrimmed video according to a query sentence. This task has achieved significant momentum in the computer vision community as it enables activity grounding beyond pre-defined activity classes by utilizing the semantic diversity of natural language descriptions. The semantic diversity is rooted in the principle of compositionality in linguistics, where novel semantics can be systematically described by combining known words in novel ways (compositional generalization). However, existing temporal grounding datasets are not carefully designed to evaluate the compositional generalizability. To systematically benchmark the compositional generalizability of temporal grounding models, we introduce a new Compositional Temporal Grounding task and construct two new dataset splits, i.e., Charades-CG and ActivityNet-CG. When evaluating the state-of-the-art methods on our new dataset splits, we empirically find that they fail to generalize to queries with novel combinations of seen words. We argue that the inherent structured semantics inside the videos and language is the crucial factor to achieve compositional generalization. Based on this insight, we propose a variational cross-graph reasoning framework that explicitly decomposes video and language into hierarchical semantic graphs, respectively, and learns fine-grained semantic correspondence between the two graphs. Furthermore, we introduce a novel adaptive structured semantics learning approach to derive the structure-informed and domain-generalizable graph representations, which facilitate the fine-grained semantic correspondence reasoning between the two graphs. Extensive experiments validate the superior compositional generalizability of our approach.

Chao Liang, Zongxin Yang, Linchao Zhu et al.

In real-world scenarios, collected and annotated data often exhibit the characteristics of multiple classes and long-tailed distribution. Additionally, label noise is inevitable in large-scale annotations and hinders the applications of learning-based models. Although many deep learning based methods have been proposed for handling long-tailed multi-label recognition or label noise respectively, learning with noisy labels in long-tailed multi-label visual data has not been well-studied because of the complexity of long-tailed distribution entangled with multi-label correlation. To tackle such a critical yet thorny problem, this paper focuses on reducing noise based on some inherent properties of multi-label classification and long-tailed learning under noisy cases. In detail, we propose a Stitch-Up augmentation to synthesize a cleaner sample, which directly reduces multi-label noise by stitching up multiple noisy training samples. Equipped with Stitch-Up, a Heterogeneous Co-Learning framework is further designed to leverage the inconsistency between long-tailed and balanced distributions, yielding cleaner labels for more robust representation learning with noisy long-tailed data. To validate our method, we build two challenging benchmarks, named VOC-MLT-Noise and COCO-MLT-Noise, respectively. Extensive experiments are conducted to demonstrate the effectiveness of our proposed method. Compared to a variety of baselines, our method achieves superior results.

Shannan Guan, Haiyan Lu, Linchao Zhu et al.

Existing 3D skeleton-based action recognition approaches reach impressive performance by encoding handcrafted action features to image format and decoding by CNNs. However, such methods are limited in two ways: a) the handcrafted action features are difficult to handle challenging actions, and b) they generally require complex CNN models to improve action recognition accuracy, which usually occur heavy computational burden. To overcome these limitations, we introduce a novel AFE-CNN, which devotes to enhance the features of 3D skeleton-based actions to adapt to challenging actions. We propose feature enhance modules from key joint, bone vector, key frame and temporal perspectives, thus the AFE-CNN is more robust to camera views and body sizes variation, and significantly improve the recognition accuracy on challenging actions. Moreover, our AFE-CNN adopts a light-weight CNN model to decode images with action feature enhanced, which ensures a much lower computational burden than the state-of-the-art methods. We evaluate the AFE-CNN on three benchmark skeleton-based action datasets: NTU RGB+D, NTU RGB+D 120, and UTKinect-Action3D, with extensive experimental results demonstrate our outstanding performance of AFE-CNN.

Yucheng Suo, Linchao Zhu, Yi Yang

In this paper, we study a challenging task of zero-shot referring image segmentation. This task aims to identify the instance mask that is most related to a referring expression without training on pixel-level annotations. Previous research takes advantage of pre-trained cross-modal models, e.g., CLIP, to align instance-level masks with referring expressions. %Yet, CLIP only considers image-text pair level alignment, which neglects fine-grained image region and complex sentence matching. Yet, CLIP only considers the global-level alignment of image-text pairs, neglecting fine-grained matching between the referring sentence and local image regions. To address this challenge, we introduce a Text Augmented Spatial-aware (TAS) zero-shot referring image segmentation framework that is training-free and robust to various visual encoders. TAS incorporates a mask proposal network for instance-level mask extraction, a text-augmented visual-text matching score for mining the image-text correlation, and a spatial rectifier for mask post-processing. Notably, the text-augmented visual-text matching score leverages a $P$ score and an $N$-score in addition to the typical visual-text matching score. The $P$-score is utilized to close the visual-text domain gap through a surrogate captioning model, where the score is computed between the surrogate model-generated texts and the referring expression. The $N$-score considers the fine-grained alignment of region-text pairs via negative phrase mining, encouraging the masked image to be repelled from the mined distracting phrases. Extensive experiments are conducted on various datasets, including RefCOCO, RefCOCO+, and RefCOCOg. The proposed method clearly outperforms state-of-the-art zero-shot referring image segmentation methods.

Yu Lu, Linchao Zhu, Hehe Fan et al.

Text-to-video (T2V) generation is a rapidly growing research area that aims to translate the scenes, objects, and actions within complex video text into a sequence of coherent visual frames. We present FlowZero, a novel framework that combines Large Language Models (LLMs) with image diffusion models to generate temporally-coherent videos. FlowZero uses LLMs to understand complex spatio-temporal dynamics from text, where LLMs can generate a comprehensive dynamic scene syntax (DSS) containing scene descriptions, object layouts, and background motion patterns. These elements in DSS are then used to guide the image diffusion model for video generation with smooth object motions and frame-to-frame coherence. Moreover, FlowZero incorporates an iterative self-refinement process, enhancing the alignment between the spatio-temporal layouts and the textual prompts for the videos. To enhance global coherence, we propose enriching the initial noise of each frame with motion dynamics to control the background movement and camera motion adaptively. By using spatio-temporal syntaxes to guide the diffusion process, FlowZero achieves improvement in zero-shot video synthesis, generating coherent videos with vivid motion.

Zenan Huang, Jun Wen, Siheng Chen et al.

Domain adaptation methods reduce domain shift typically by learning domain-invariant features. Most existing methods are built on distribution matching, e.g., adversarial domain adaptation, which tends to corrupt feature discriminability. In this paper, we propose Discriminative Radial Domain Adaptation (DRDA) which bridges source and target domains via a shared radial structure. It's motivated by the observation that as the model is trained to be progressively discriminative, features of different categories expand outwards in different directions, forming a radial structure. We show that transferring such an inherently discriminative structure would enable to enhance feature transferability and discriminability simultaneously. Specifically, we represent each domain with a global anchor and each category a local anchor to form a radial structure and reduce domain shift via structure matching. It consists of two parts, namely isometric transformation to align the structure globally and local refinement to match each category. To enhance the discriminability of the structure, we further encourage samples to cluster close to the corresponding local anchors based on optimal-transport assignment. Extensively experimenting on multiple benchmarks, our method is shown to consistently outperforms state-of-the-art approaches on varied tasks, including the typical unsupervised domain adaptation, multi-source domain adaptation, domain-agnostic learning, and domain generalization.

Yuanzhi Liang, Linchao Zhu, Yi Yang

Recent advancements in natural language and Large Language Models (LLMs) have enabled AI agents to simulate human-like interactions within virtual worlds. However, these interactions still face limitations in complexity and flexibility, particularly in scenarios involving multiple characters and novel objects. Pre-defining all interactable objects in the agent's world model presents challenges, and conveying implicit intentions to multiple characters through complex interactions remains difficult. To address these issues, we propose integrating virtual Game Masters (GMs) into the agent's world model, drawing inspiration from Tabletop Role-Playing Games (TRPGs). GMs play a crucial role in overseeing information, estimating players' intentions, providing environment descriptions, and offering feedback, compensating for current world model deficiencies. To facilitate future explorations for complex interactions, we introduce a benchmark named Tachikuma, comprising a Multiple character and novel Object based interaction Estimation (MOE) task and a supporting dataset. MOE challenges models to understand characters' intentions and accurately determine their actions within intricate contexts involving multi-character and novel object interactions. Besides, the dataset captures log data from real-time communications during gameplay, providing diverse, grounded, and complex interactions for further explorations. Finally, we present a simple prompting baseline and evaluate its performance, demonstrating its effectiveness in enhancing interaction understanding. We hope that our dataset and task will inspire further research in complex interactions with natural language, fostering the development of more advanced AI agents.

Xu Zhang, Zhedong Zheng, Linchao Zhu et al.

Composed image retrieval extends content-based image retrieval systems by enabling users to search using reference images and captions that describe their intention. Despite great progress in developing image-text compositors to extract discriminative visual-linguistic features, we identify a hitherto overlooked issue, triplet ambiguity, which impedes robust feature extraction. Triplet ambiguity refers to a type of semantic ambiguity that arises between the reference image, the relative caption, and the target image. It is mainly due to the limited representation of the annotated text, resulting in many noisy triplets where multiple visually dissimilar candidate images can be matched to an identical reference pair (i.e., a reference image + a relative caption). To address this challenge, we propose the Consensus Network (Css-Net), inspired by the psychological concept that groups outperform individuals. Css-Net comprises two core components: (1) a consensus module with four diverse compositors, each generating distinct image-text embeddings, fostering complementary feature extraction and mitigating dependence on any single, potentially biased compositor; (2) a Kullback-Leibler divergence loss that encourages learning of inter-compositor interactions to promote consensual outputs. During evaluation, the decisions of the four compositors are combined through a weighting scheme, enhancing overall agreement. On benchmark datasets, particularly FashionIQ, Css-Net demonstrates marked improvements. Notably, it achieves significant recall gains, with a 2.77% increase in R@10 and 6.67% boost in R@50, underscoring its competitiveness in addressing the fundamental limitations of existing methods.

Shannan Guan, Haiyan Lu, Linchao Zhu et al.

3D pose estimation has recently gained substantial interests in computer vision domain. Existing 3D pose estimation methods have a strong reliance on large size well-annotated 3D pose datasets, and they suffer poor model generalization on unseen poses due to limited diversity of 3D poses in training sets. In this work, we propose PoseGU, a novel human pose generator that generates diverse poses with access only to a small size of seed samples, while equipping the Counterfactual Risk Minimization to pursue an unbiased evaluation objective. Extensive experiments demonstrate PoseGU outforms almost all the state-of-the-art 3D human pose methods under consideration over three popular benchmark datasets. Empirical analysis also proves PoseGU generates 3D poses with improved data diversity and better generalization ability.

Chao Liang, Linchao Zhu, Humphrey Shi et al.

Modern deep learning systems are data-hungry. Learning with web data is one of the feasible solutions, but will introduce label noise inevitably, which can hinder the performance of deep neural networks. Sample selection is an effective way to deal with label noise. The key is to separate clean samples based on some criterion. Previous methods pay more attention to the small loss criterion where small-loss samples are regarded as clean ones. Nevertheless, such a strategy relies on the learning dynamics of each data instance. Some noisy samples are still memorized due to frequently occurring corrupted learning patterns. To tackle this problem, a training-free surrogate model is preferred, freeing from the effect of memorization. In this work, we propose to leverage the vision-language surrogate model CLIP to filter noisy samples automatically. CLIP brings external knowledge to facilitate the selection of clean samples with its ability of text-image alignment. Furthermore, a margin adaptive loss is designed to regularize the selection bias introduced by CLIP, providing robustness to label noise. We validate the effectiveness of our proposed method on both real-world and synthetic noisy datasets. Our method achieves significant improvement without CLIP involved during the inference stage.

Xuexiang Wen, Hang Yu, Linchao Zhu et al.

While Reinforcement Learning with Verifiable Rewards (RLVR) has recently emerged as a promising post-training paradigm for Large Language Models (LLMs), its dependency on the gold label or domain-specific verifiers limits its scalability to new tasks and domains. In this work, we propose Verifier-free Intrinsic Gradient-Norm Reward (VIGOR), a simple reward that uses only the policy model itself. Given a prompt, VIGOR samples a group of completions and assigns higher within-group rewards to outputs that induce smaller $\ell_2$ norms of the teacher-forced negative log-likelihood gradients under the current parameters. Intuitively, lower gradient norms suggest the completion aligns better with the current policy, serving as an intrinsic preference signal for policy optimization. To make this intrinsic signal practical for RL, we correct the systematic length bias of averaged token-level gradients with a $\sqrt{T}$ scaling, and apply group-wise rank shaping to stabilize reward scales across prompts. Across mathematical reasoning benchmarks, VIGOR outperforms the state-of-the-art Reinforcement Learning from Internal Feedback (RLIF) baseline, and it also exhibits cross-domain transfer to code benchmarks when trained only on math data. For instance, on Qwen2.5-7B-Base post-trained on MATH, VIGOR improves the average math accuracy by +3.31% and the average code accuracy by +1.91% over this baseline, while exhibiting more stable training dynamics. The code is available at https://github.com/ZJUSCL/VIGOR.

Yunzhu Zhang, Zeyu Pan, Zhengwen Zeng et al.

GUI grounding, which translates natural language instructions into precise pixel coordinates, is essential for developing practical GUI agents. However, we observe that existing grounding models exhibit significant coordinate prediction instability, minor visual perturbations (e.g. cropping a few pixels) can drastically alter predictions, flipping results between correct and incorrect. This instability severely undermines model performance, especially for samples with high-resolution and small UI elements. To address this issue, we propose Multi-View Prediction (MVP), a training-free framework that enhances grounding performance through multi-view inference. Our key insight is that while single-view predictions may be unstable, aggregating predictions from multiple carefully cropped views can effectively distinguish correct coordinates from outliers. MVP comprises two components: (1) Attention-Guided View Proposal, which derives diverse views guided by instruction-to-image attention scores, and (2) Multi-Coordinates Clustering, which ensembles predictions by selecting the centroid of the densest spatial cluster. Extensive experiments demonstrate MVP's effectiveness across various models and benchmarks. Notably, on ScreenSpot-Pro, MVP boosts UI-TARS-1.5-7B to 56.1%, GTA1-7B to 61.7%, Qwen3VL-8B-Instruct to 65.3%, and Qwen3VL-32B-Instruct to 74.0%. The code is available at https://github.com/ZJUSCL/MVP.

Chao Liang, Fan Ma, Linchao Zhu et al.

We concentrate on a novel human-centric image synthesis task, that is, given only one reference facial photograph, it is expected to generate specific individual images with diverse head positions, poses, facial expressions, and illuminations in different contexts. To accomplish this goal, we argue that our generative model should be capable of the following favorable characteristics: (1) a strong visual and semantic understanding of our world and human society for basic object and human image generation. (2) generalizable identity preservation ability. (3) flexible and fine-grained head control. Recently, large pre-trained text-to-image diffusion models have shown remarkable results, serving as a powerful generative foundation. As a basis, we aim to unleash the above two capabilities of the pre-trained model. In this work, we present a new framework named CapHuman. We embrace the "encode then learn to align" paradigm, which enables generalizable identity preservation for new individuals without cumbersome tuning at inference. CapHuman encodes identity features and then learns to align them into the latent space. Moreover, we introduce the 3D facial prior to equip our model with control over the human head in a flexible and 3D-consistent manner. Extensive qualitative and quantitative analyses demonstrate our CapHuman can produce well-identity-preserved, photo-realistic, and high-fidelity portraits with content-rich representations and various head renditions, superior to established baselines. Code and checkpoint will be released at https://github.com/VamosC/CapHuman.

Xiangpeng Yang, Linchao Zhu, Hehe Fan et al.

Recent advancements in diffusion models have significantly improved video generation and editing capabilities. However, multi-grained video editing, which encompasses class-level, instance-level, and part-level modifications, remains a formidable challenge. The major difficulties in multi-grained editing include semantic misalignment of text-to-region control and feature coupling within the diffusion model. To address these difficulties, we present VideoGrain, a zero-shot approach that modulates space-time (cross- and self-) attention mechanisms to achieve fine-grained control over video content. We enhance text-to-region control by amplifying each local prompt's attention to its corresponding spatial-disentangled region while minimizing interactions with irrelevant areas in cross-attention. Additionally, we improve feature separation by increasing intra-region awareness and reducing inter-region interference in self-attention. Extensive experiments demonstrate our method achieves state-of-the-art performance in real-world scenarios. Our code, data, and demos are available at https://knightyxp.github.io/VideoGrain_project_page/

Hongyuan Tao, Ying Zhang, Zhenhao Tang et al.

Recent advances in Large Language Models (LLMs) have shown promise in function-level code generation, yet repository-level software engineering tasks remain challenging. Current solutions predominantly rely on proprietary LLM agents, which introduce unpredictability and limit accessibility, raising concerns about data privacy and model customization. This paper investigates whether open-source LLMs can effectively address repository-level tasks without requiring agent-based approaches. We demonstrate this is possible by enabling LLMs to comprehend functions and files within codebases through their semantic information and structural dependencies. To this end, we introduce Code Graph Models (CGMs), which integrate repository code graph structures into the LLM's attention mechanism and map node attributes to the LLM's input space using a specialized adapter. When combined with an agentless graph RAG framework, our approach achieves a 43.00% resolution rate on the SWE-bench Lite benchmark using the open-source Qwen2.5-72B model. This performance ranks first among open weight models, second among methods with open-source systems, and eighth overall, surpassing the previous best open-source model-based method by 12.33%.

Veuns-Team, Changlong Gao, Zhangxuan Gu et al.

GUI agents have emerged as a powerful paradigm for automating interactions in digital environments, yet achieving both broad generality and consistently strong task performance remains challenging.In this report, we present UI-Venus-1.5, a unified, end-to-end GUI Agent designed for robust real-world applications.The proposed model family comprises two dense variants (2B and 8B) and one mixture-of-experts variant (30B-A3B) to meet various downstream application scenarios.Compared to our previous version, UI-Venus-1.5 introduces three key technical advances: (1) a comprehensive Mid-Training stage leveraging 10 billion tokens across 30+ datasets to establish foundational GUI semantics; (2) Online Reinforcement Learning with full-trajectory rollouts, aligning training objectives with long-horizon, dynamic navigation in large-scale environments; and (3) a single unified GUI Agent constructed via Model Merging, which synthesizes domain-specific models (grounding, web, and mobile) into one cohesive checkpoint. Extensive evaluations demonstrate that UI-Venus-1.5 establishes new state-of-the-art performance on benchmarks such as ScreenSpot-Pro (69.6%), VenusBench-GD (75.0%), and AndroidWorld (77.6%), significantly outperforming previous strong baselines. In addition, UI-Venus-1.5 demonstrates robust navigation capabilities across a variety of Chinese mobile apps, effectively executing user instructions in real-world scenarios. Code: https://github.com/inclusionAI/UI-Venus; Model: https://huggingface.co/collections/inclusionAI/ui-venus

Shuai Zhao, Yunqiu Xu, Linchao Zhu et al.

Large language models~(LLMs) are expected to be helpful, harmless, and honest. In different alignment scenarios, such as safety, confidence, and general preference alignment, binary preference data collection and reward modeling are resource-intensive but play a central role in transferring human preferences. In this work, we explore using the similarity between sampled generations and reference answers as a supplementary reward function for alignment. When unary reference answers are available, such similarity-based rewards can circumvent the need for binary preference data and explicit reward modeling. We introduce \textit{RefAlign}, a versatile REINFORCE-style alignment algorithm that does not rely on reward or reference models. RefAlign utilizes language generation evaluation metrics, such as BERTScore, between sampled generations and reference answers as surrogate rewards. Beyond general preference optimization, RefAlign can be naturally extended to diverse scenarios, including safety and confidence alignment, by combining similarity-based rewards with task-specific objectives. Across multiple scenarios, RefAlign achieves performance comparable to prior alignment methods while operating without binary preference data or reward models. The code is available at https://github.com/mzhaoshuai/RefAlign.

Heng Jia, Linchao Zhu, Na Zhao

Despite recent advances in feed-forward 3D Gaussian Splatting, generalizable 3D reconstruction remains challenging, particularly in multi-view correspondence modeling. Existing approaches face a fundamental trade-off: explicit methods achieve geometric precision but struggle with ambiguous regions, while implicit methods provide robustness but suffer from slow convergence. We present H3R, a hybrid framework that addresses this limitation by integrating volumetric latent fusion with attention-based feature aggregation. Our framework consists of two complementary components: an efficient latent volume that enforces geometric consistency through epipolar constraints, and a camera-aware Transformer that leverages Plücker coordinates for adaptive correspondence refinement. By integrating both paradigms, our approach enhances generalization while converging 2$\times$ faster than existing methods. Furthermore, we show that spatial-aligned foundation models (e.g., SD-VAE) substantially outperform semantic-aligned models (e.g., DINOv2), resolving the mismatch between semantic representations and spatial reconstruction requirements. Our method supports variable-number and high-resolution input views while demonstrating robust cross-dataset generalization. Extensive experiments show that our method achieves state-of-the-art performance across multiple benchmarks, with significant PSNR improvements of 0.59 dB, 1.06 dB, and 0.22 dB on the RealEstate10K, ACID, and DTU datasets, respectively. Code is available at https://github.com/JiaHeng-DLUT/H3R.

Yubo Dong, Xukun Zhu, Zhengzhe Pan et al.

In this paper, we aim to evaluate multi-agent systems against complex dependencies, including spatial, causal, and temporal constraints. First, we construct a new benchmark, named VillagerBench, within the Minecraft environment.VillagerBench comprises diverse tasks crafted to test various aspects of multi-agent collaboration, from workload distribution to dynamic adaptation and synchronized task execution. Second, we introduce a Directed Acyclic Graph Multi-Agent Framework VillagerAgent to resolve complex inter-agent dependencies and enhance collaborative efficiency. This solution incorporates a task decomposer that creates a directed acyclic graph (DAG) for structured task management, an agent controller for task distribution, and a state manager for tracking environmental and agent data. Our empirical evaluation on VillagerBench demonstrates that VillagerAgent outperforms the existing AgentVerse model, reducing hallucinations and improving task decomposition efficacy. The results underscore VillagerAgent's potential in advancing multi-agent collaboration, offering a scalable and generalizable solution in dynamic environments. The source code is open-source on GitHub (https://github.com/cnsdqd-dyb/VillagerAgent).

Xiangpeng Yang, Linchao Zhu, Xiaohan Wang et al.

Text-video retrieval is a critical multi-modal task to find the most relevant video for a text query. Although pretrained models like CLIP have demonstrated impressive potential in this area, the rising cost of fully finetuning these models due to increasing model size continues to pose a problem. To address this challenge, prompt tuning has emerged as an alternative. However, existing works still face two problems when adapting pretrained image-text models to downstream video-text tasks: (1) The visual encoder could only encode frame-level features and failed to extract global-level general video information. (2) Equipping the visual and text encoder with separated prompts failed to mitigate the visual-text modality gap. To this end, we propose DGL, a cross-modal Dynamic prompt tuning method with Global-Local video attention. In contrast to previous prompt tuning methods, we employ the shared latent space to generate local-level text and frame prompts that encourage inter-modal interaction. Furthermore, we propose modeling video in a global-local attention mechanism to capture global video information from the perspective of prompt tuning. Extensive experiments reveal that when only 0.67% parameters are tuned, our cross-modal prompt tuning strategy DGL outperforms or is comparable to fully finetuning methods on MSR-VTT, VATEX, LSMDC, and ActivityNet datasets. Code will be available at https://github.com/knightyxp/DGL

Shuai Zhao, Xiaohan Wang, Linchao Zhu et al.

One fascinating aspect of pre-trained vision-language models~(VLMs) learning under language supervision is their impressive zero-shot generalization capability. However, this ability is hindered by distribution shifts between the training and testing data. Previous test time adaptation~(TTA) methods for VLMs in zero-shot classification rely on minimizing the entropy of model outputs, tending to be stuck in incorrect model predictions. In this work, we propose TTA with feedback to rectify the model output and prevent the model from becoming blindly confident. Specifically, a CLIP model is adopted as the reward model during TTA and provides feedback for the VLM. Given a single test sample, the VLM is forced to maximize the CLIP reward between the input and sampled results from the VLM output distribution. The proposed \textit{reinforcement learning with CLIP feedback~(RLCF)} framework is highly flexible and universal. Beyond the classification task, with task-specific sampling strategies and a proper reward baseline choice, RLCF can be easily extended to not only discrimination tasks like retrieval but also generalization tasks like image captioning, improving the zero-shot generalization capacity of VLMs. According to the characteristics of these VL tasks, we build different fully TTA pipelines with RLCF to improve the zero-shot generalization ability of various VLMs. Extensive experiments along with promising empirical results demonstrate the effectiveness of RLCF. The code is available at https://github.com/mzhaoshuai/RLCF.

Kezhou Lin, Xiaohan Wang, Linchao Zhu et al.

In this paper, we tackle the problem of sign language translation (SLT) without gloss annotations. Although intermediate representation like gloss has been proven effective, gloss annotations are hard to acquire, especially in large quantities. This limits the domain coverage of translation datasets, thus handicapping real-world applications. To mitigate this problem, we design the Gloss-Free End-to-end sign language translation framework (GloFE). Our method improves the performance of SLT in the gloss-free setting by exploiting the shared underlying semantics of signs and the corresponding spoken translation. Common concepts are extracted from the text and used as a weak form of intermediate representation. The global embedding of these concepts is used as a query for cross-attention to find the corresponding information within the learned visual features. In a contrastive manner, we encourage the similarity of query results between samples containing such concepts and decrease those that do not. We obtained state-of-the-art results on large-scale datasets, including OpenASL and How2Sign. The code and model will be available at https://github.com/HenryLittle/GloFE.

Fan Ma, Linchao Zhu, Yi Yang et al.

In this paper, we study an intermediate form of supervision, i.e., single-frame supervision, for temporal action localization (TAL). To obtain the single-frame supervision, the annotators are asked to identify only a single frame within the temporal window of an action. This can significantly reduce the labor cost of obtaining full supervision which requires annotating the action boundary. Compared to the weak supervision that only annotates the video-level label, the single-frame supervision introduces extra temporal action signals while maintaining low annotation overhead. To make full use of such single-frame supervision, we propose a unified system called SF-Net. First, we propose to predict an actionness score for each video frame. Along with a typical category score, the actionness score can provide comprehensive information about the occurrence of a potential action and aid the temporal boundary refinement during inference. Second, we mine pseudo action and background frames based on the single-frame annotations. We identify pseudo action frames by adaptively expanding each annotated single frame to its nearby, contextual frames and we mine pseudo background frames from all the unannotated frames across multiple videos. Together with the ground-truth labeled frames, these pseudo-labeled frames are further used for training the classifier. In extensive experiments on THUMOS14, GTEA, and BEOID, SF-Net significantly improves upon state-of-the-art weakly-supervised methods in terms of both segment localization and single-frame localization. Notably, SF-Net achieves comparable results to its fully-supervised counterpart which requires much more resource intensive annotations. The code is available at https://github.com/Flowerfan/SF-Net.

Linchao Zhu, Yanbin Liu, Yi Yang

In this paper, we present our solution to Google YouTube-8M Video Classification Challenge 2017. We leveraged both video-level and frame-level features in the submission. For video-level classification, we simply used a 200-mixture Mixture of Experts (MoE) layer, which achieves GAP 0.802 on the validation set with a single model. For frame-level classification, we utilized several variants of recurrent neural networks, sequence aggregation with attention mechanism and 1D convolutional models. We achieved GAP 0.8408 on the private testing set with the ensemble model. The source code of our models can be found in \url{https://github.com/ffmpbgrnn/yt8m}.

Wenyi Xiao, Ziwei Huang, Leilei Gan et al.

The rapidly developing Large Vision Language Models (LVLMs) have shown notable capabilities on a range of multi-modal tasks, but still face the hallucination phenomena where the generated texts do not align with the given contexts, significantly restricting the usages of LVLMs. Most previous work detects and mitigates hallucination at the coarse-grained level or requires expensive annotation (e.g., labeling by proprietary models or human experts). To address these issues, we propose detecting and mitigating hallucinations in LVLMs via fine-grained AI feedback. The basic idea is that we generate a small-size sentence-level hallucination annotation dataset by proprietary models, whereby we train a hallucination detection model which can perform sentence-level hallucination detection, covering primary hallucination types (i.e., object, attribute, and relationship). Then, we propose a detect-then-rewrite pipeline to automatically construct preference dataset for training hallucination mitigating model. Furthermore, we propose differentiating the severity of hallucinations, and introducing a Hallucination Severity-Aware Direct Preference Optimization (HSA-DPO) for mitigating hallucination in LVLMs by incorporating the severity of hallucinations into preference learning. Extensive experiments demonstrate the effectiveness of our method.

Yucheng Suo, Fan Ma, Linchao Zhu et al.

We study the zero-shot Composed Image Retrieval (ZS-CIR) task, which is to retrieve the target image given a reference image and a description without training on the triplet datasets. Previous works generate pseudo-word tokens by projecting the reference image features to the text embedding space. However, they focus on the global visual representation, ignoring the representation of detailed attributes, e.g., color, object number and layout. To address this challenge, we propose a Knowledge-Enhanced Dual-stream zero-shot composed image retrieval framework (KEDs). KEDs implicitly models the attributes of the reference images by incorporating a database. The database enriches the pseudo-word tokens by providing relevant images and captions, emphasizing shared attribute information in various aspects. In this way, KEDs recognizes the reference image from diverse perspectives. Moreover, KEDs adopts an extra stream that aligns pseudo-word tokens with textual concepts, leveraging pseudo-triplets mined from image-text pairs. The pseudo-word tokens generated in this stream are explicitly aligned with fine-grained semantics in the text embedding space. Extensive experiments on widely used benchmarks, i.e. ImageNet-R, COCO object, Fashion-IQ and CIRR, show that KEDs outperforms previous zero-shot composed image retrieval methods.

Yang Wan, Zheng Cao, Zhenhao Zhang et al.

Large language models excel as few-shot learners when provided with appropriate demonstrations, yet this strength becomes problematic in multiturn agent scenarios, where LLMs erroneously mimic their own previous responses as few-shot examples. Through attention analysis, we identify conversational inertia, a phenomenon where models exhibit strong diagonal attention to previous responses, which is associated with imitation bias that constrains exploration. This reveals a tension when transforming few-shot LLMs into agents: longer context enriches environmental feedback for exploitation, yet also amplifies conversational inertia that undermines exploration. Our key insight is that for identical states, actions generated with longer contexts exhibit stronger inertia than those with shorter contexts, enabling construction of preference pairs without environment rewards. Based on this, we propose Context Preference Learning to calibrate model preferences to favor low-inertia responses over highinertia ones. We further provide context management strategies at inference time to balance exploration and exploitation. Experimental results across eight agentic environments and one deep research scenario validate that our framework reduces conversational inertia and achieves performance improvements.

Xiangpeng Yang, Linchao Zhu, Hehe Fan et al.

Current diffusion-based video editing primarily focuses on local editing (\textit{e.g.,} object/background editing) or global style editing by utilizing various dense correspondences. However, these methods often fail to accurately edit the foreground and background simultaneously while preserving the original layout. We find that the crux of the issue stems from the imprecise distribution of attention weights across designated regions, including inaccurate text-to-attribute control and attention leakage. To tackle this issue, we introduce EVA, a \textbf{zero-shot} and \textbf{multi-attribute} video editing framework tailored for human-centric videos with complex motions. We incorporate a Spatial-Temporal Layout-Guided Attention mechanism that leverages the intrinsic positive and negative correspondences of cross-frame diffusion features. To avoid attention leakage, we utilize these correspondences to boost the attention scores of tokens within the same attribute across all video frames while limiting interactions between tokens of different attributes in the self-attention layer. For precise text-to-attribute manipulation, we use discrete text embeddings focused on specific layout areas within the cross-attention layer. Benefiting from the precise attention weight distribution, EVA can be easily generalized to multi-object editing scenarios and achieves accurate identity mapping. Extensive experiments demonstrate EVA achieves state-of-the-art results in real-world scenarios. Full results are provided at https://knightyxp.github.io/EVA/

Yunqiu Xu, Linchao Zhu, Yi Yang

While multimodal large language models (MLLMs) have demonstrated extraordinary vision-language understanding capabilities, their abilities to solve instance-level visual-language problems beyond a single image warrant further exploration. To assess these unproven abilities of MLLMs, this paper proposes a new visual grounding task called multi-context visual grounding, which aims to localize instances of interest across multiple images based on open-ended text prompts. In order to facilitate this research, we construct a new dataset MC-Bench that features 2K high-quality and manually annotated samples. Each sample consists of an instance-level labeled image pair and a corresponding text prompt that indicates the target instances in the images. These text prompts are highly open-ended and follow three distinct styles, covering 20 practical skills. We benchmark over 20 state-of-the-art MLLMs and foundation models with potential multi-context visual grounding capabilities, along with our developed simple yet effective agentic baseline and a finetuned baseline by multi-context instruction tuning. Our evaluation reveals a non-trivial performance gap between existing MLLMs and humans, along with some insightful observations that suggest potential future directions. We hope that MC-Bench and our empirical findings encourage the research community to further advance the untapped potentials of MLLMs in instance-level tasks, particularly in multi-image contexts. Project page: https://xuyunqiu.github.io/MC-Bench.

Xiao Liang, Yunzhu Zhang, Linchao Zhu

Diffusion models have achieved remarkable success in video generation; however, the high computational cost of the denoising process remains a major bottleneck. Existing approaches have shown promise in reducing the number of diffusion steps, but they often suffer from significant quality degradation when applied to video generation. We propose Guided Progressive Distillation (GPD), a framework that accelerates the diffusion process for fast and high-quality video generation. GPD introduces a novel training strategy in which a teacher model progressively guides a student model to operate with larger step sizes. The framework consists of two key components: (1) an online-generated training target that reduces optimization difficulty while improving computational efficiency, and (2) frequency-domain constraints in the latent space that promote the preservation of fine-grained details and temporal dynamics. Applied to the Wan2.1 model, GPD reduces the number of sampling steps from 48 to 6 while maintaining competitive visual quality on VBench. Compared with existing distillation methods, GPD demonstrates clear advantages in both pipeline simplicity and quality preservation.

Yucheng Suo, Fan Ma, Linchao Zhu et al.

Multi-modal Large language models (MLLMs) show remarkable ability in video understanding. Nevertheless, understanding long videos remains challenging as the models can only process a finite number of frames in a single inference, potentially omitting crucial visual information. To address the challenge, we propose generating multiple predictions through visual context sampling, followed by a scoring mechanism to select the final prediction. Specifically, we devise a bin-wise sampling strategy that enables MLLMs to generate diverse answers based on various combinations of keyframes, thereby enriching the visual context. To determine the final prediction from the sampled answers, we employ a self-reward by linearly combining three scores: (1) a frequency score indicating the prevalence of each option, (2) a marginal confidence score reflecting the inter-intra sample certainty of MLLM predictions, and (3) a reasoning score for different question types, including clue-guided answering for global questions and temporal self-refocusing for local questions. The frequency score ensures robustness through majority correctness, the confidence-aligned score reflects prediction certainty, and the typed-reasoning score addresses cases with sparse key visual information using tailored strategies. Experiments show that this approach covers the correct answer for a high percentage of long video questions, on seven datasets show that our method improves the performance of three MLLMs.

Yunzhu Zhang, Yu Lu, Tianyi Wang et al.

Long-form video understanding poses a significant challenge for video large language models (VideoLLMs) due to prohibitively high computational and memory demands. In this paper, we propose FlexSelect, a flexible and efficient token selection strategy for processing long videos. FlexSelect identifies and retains the most semantically relevant content by leveraging cross-modal attention patterns from a reference transformer layer. It comprises two key components: (1) a training-free token ranking pipeline that leverages faithful cross-modal attention weights to estimate each video token's importance, and (2) a rank-supervised lightweight selector that is trained to replicate these rankings and filter redundant tokens. This generic approach can be seamlessly integrated into various VideoLLM architectures, such as LLaVA-Video, InternVL and Qwen-VL, serving as a plug-and-play module to extend their temporal context length. Empirically, FlexSelect delivers strong gains across multiple long-video benchmarks including VideoMME, MLVU, LongVB, and LVBench. Moreover, it achieves significant speed-ups (for example, up to 9 times on a LLaVA-Video-7B model), highlighting FlexSelect's promise for efficient long-form video understanding. Project page available at: https://yunzhuzhang0918.github.io/flex_select

Kaixin Shen, Ruijie Quan, Linchao Zhu et al.

Maintaining temporal stability is crucial in multi-agent trajectory prediction. Insufficient regularization to uphold this stability often results in fluctuations in kinematic states, leading to inconsistent predictions and the amplification of errors. In this study, we introduce a framework called Multi-Agent Trajectory prediction via neural interaction Energy (MATE). This framework assesses the interactive motion of agents by employing neural interaction energy, which captures the dynamics of interactions and illustrates their influence on the future trajectories of agents. To bolster temporal stability, we introduce two constraints: inter-agent interaction constraint and intra-agent motion constraint. These constraints work together to ensure temporal stability at both the system and agent levels, effectively mitigating prediction fluctuations inherent in multi-agent systems. Comparative evaluations against previous methods on four diverse datasets highlight the superior prediction accuracy and generalization capabilities of our model.

Shuai Zhao, Linchao Zhu, Ruijie Quan et al.

A primary concern regarding training large language models (LLMs) is whether they abuse copyrighted online text. With the increasing training data scale and the prevalence of LLMs in daily lives, two problems arise: \textbf{1)} false positive membership inference results misled by similar examples; \textbf{2)} membership inference methods are usually too complex for end users to understand and use. To address these issues, we propose an alternative \textit{insert-and-detect} methodology, advocating that web users and content platforms employ \textbf{\textit{unique identifiers}} for reliable and independent membership inference. Users and platforms can create their identifiers, embed them in copyrighted text, and independently detect them in future LLMs. As an initial demonstration, we introduce \textit{\textbf{ghost sentences}} and a user-friendly last-$k$ words test, allowing end users to chat with LLMs for membership inference. Ghost sentences consist primarily of unique passphrases of random natural words, which can come with customized elements to bypass possible filter rules. The last-$k$ words test requires a significant repetition time of ghost sentences~($\ge10$). For cases with fewer repetitions, we designed an extra perplexity test, as LLMs exhibit high perplexity when encountering unnatural passphrases. We also conduct a comprehensive study on the memorization and membership inference of ghost sentences, examining factors such as training data scales, model sizes, repetition times, insertion positions, wordlist of passphrases, alignment, \textit{etc}. Our study shows the possibility of applying ghost sentences in real scenarios and provides instructions for the potential application.

Yucheng Suo, Fan Ma, Kaixin Shen et al.

Visual instructions for long-horizon tasks are crucial as they intuitively clarify complex concepts and enhance retention across extended steps. Directly generating a series of images using text-to-image models without considering the context of previous steps results in inconsistent images, increasing cognitive load. Additionally, the generated images often miss objects or the attributes such as color, shape, and state of the objects are inaccurate. To address these challenges, we propose LIGER, the first training-free framework for Long-horizon Instruction GEneration with logic and attribute self-Reflection. LIGER first generates a draft image for each step with the historical prompt and visual memory of previous steps. This step-by-step generation approach maintains consistency between images in long-horizon tasks. Moreover, LIGER utilizes various image editing tools to rectify errors including wrong attributes, logic errors, object redundancy, and identity inconsistency in the draft images. Through this self-reflection mechanism, LIGER improves the logic and object attribute correctness of the images. To verify whether the generated images assist human understanding, we manually curated a new benchmark consisting of various long-horizon tasks. Human-annotated ground truth expressions reflect the human-defined criteria for how an image should appear to be illustrative. Experiments demonstrate the visual instructions generated by LIGER are more comprehensive compared with baseline methods.

Xihang Yue, Yi Yang, Linchao Zhu

Recent advances in operator learning have produced two distinct approaches for solving partial differential equations (PDEs): attention-based methods offering point-level adaptability but lacking spectral constraints, and spectral-based methods providing domain-level continuity priors but limited in local flexibility. This dichotomy has hindered the development of PDE solvers with both strong flexibility and generalization capability. This work introduces Holistic Physics Mixer (HPM), a simple framework that bridges this gap by integrating spectral and physical information in a unified space. HPM unifies both approaches as special cases while enabling more powerful spectral-physical interactions beyond either method alone. This enables HPM to inherit both the strong generalization of spectral methods and the flexibility of attention mechanisms while avoiding their respective limitations. Through extensive experiments across diverse PDE problems, we demonstrate that HPM consistently outperforms state-of-the-art methods in both accuracy and computational efficiency, while maintaining strong generalization capabilities with limited training data and excellent zero-shot performance on unseen resolutions.

Kaixin Shen, Ruijie Quan, Linchao Zhu et al.

Audio-driven visual scene editing endeavors to manipulate the visual background while leaving the foreground content unchanged, according to the given audio signals. Unlike current efforts focusing primarily on image editing, audio-driven video scene editing has not been extensively addressed. In this paper, we introduce AudioScenic, an audio-driven framework designed for video scene editing. AudioScenic integrates audio semantics into the visual scene through a temporal-aware audio semantic injection process. As our focus is on background editing, we further introduce a SceneMasker module, which maintains the integrity of the foreground content during the editing process. AudioScenic exploits the inherent properties of audio, namely, audio magnitude and frequency, to guide the editing process, aiming to control the temporal dynamics and enhance the temporal consistency. First, we present an audio Magnitude Modulator module that adjusts the temporal dynamics of the scene in response to changes in audio magnitude, enhancing the visual dynamics. Second, the audio Frequency Fuser module is designed to ensure temporal consistency by aligning the frequency of the audio with the dynamics of the video scenes, thus improving the overall temporal coherence of the edited videos. These integrated features enable AudioScenic to not only enhance visual diversity but also maintain temporal consistency throughout the video. We present a new metric named temporal score for more comprehensive validation of temporal consistency. We demonstrate substantial advancements of AudioScenic over competing methods on DAVIS and Audioset datasets.

Chao Liang, Linchao Zhu, Zongxin Yang et al.

We focus on the challenging problem of learning an unbiased classifier from a large number of potentially relevant but noisily labeled web images given only a few clean labeled images. This problem is particularly practical because it reduces the expensive annotation costs by utilizing freely accessible web images with noisy labels. Typically, prototypes are representative images or features used to classify or identify other images. However, in the few clean and many noisy scenarios, the class prototype can be severely biased due to the presence of irrelevant noisy images. The resulting prototypes are less compact and discriminative, as previous methods do not take into account the diverse range of images in the noisy web image collections. On the other hand, the relation modeling between noisy and clean images is not learned for the class prototype generation in an end-to-end manner, which results in a suboptimal class prototype. In this article, we introduce a similarity maximization loss named SimNoiPro. Our SimNoiPro first generates noise-tolerant hybrid prototypes composed of clean and noise-tolerant prototypes and then pulls them closer to each other. Our approach considers the diversity of noisy images by explicit division and overcomes the optimization discrepancy issue. This enables better relation modeling between clean and noisy images and helps extract judicious information from the noisy image set. The evaluation results on two extended few-shot classification benchmarks confirm that our SimNoiPro outperforms prior methods in measuring image relations and cleaning noisy data.

Yubo Dong, Linchao Zhu

Latent Video Diffusion Models (LVDMs) rely on Variational Autoencoders (VAEs) to compress videos into compact latent representations. For continuous Variational Autoencoders (VAEs), achieving higher compression rates is desirable; yet, the efficiency notably declines when extra sampling layers are added without expanding the dimensions of hidden channels. In this paper, we present a technique to convert fixed compression rate VAEs into models that support multi-level temporal compression, providing a straightforward and minimal fine-tuning approach to counteract performance decline at elevated compression rates.Moreover, we examine how varying compression levels impact model performance over video segments with diverse characteristics, offering empirical evidence on the effectiveness of our proposed approach. We also investigate the integration of our multi-level temporal compression VAE with diffusion-based generative models, DiT, highlighting successful concurrent training and compatibility within these frameworks. This investigation illustrates the potential uses of multi-level temporal compression.