Meng Cao

Zhengfeng Lai, Haotian Zhang, Bowen Zhang et al.

Large-scale web-crawled datasets are fundamental for the success of pre-training vision-language models, such as CLIP. However, the inherent noise and potential irrelevance of web-crawled AltTexts pose challenges in achieving precise image-text alignment. Existing methods utilizing large language models (LLMs) for caption rewriting have shown promise on small, curated datasets like CC3M and CC12M. This study introduces a scalable pipeline for noisy caption rewriting. Unlike recent LLM rewriting techniques, we emphasize the incorporation of visual concepts into captions, termed as Visual-enriched Captions (VeCap). To ensure data diversity, we propose a novel mixed training scheme that optimizes the utilization of AltTexts alongside newly generated VeCap. We showcase the adaptation of this method for training CLIP on large-scale web-crawled datasets, termed VeCLIP. Employing this cost-effective pipeline, we effortlessly scale our dataset up to 300 million samples named VeCap dataset. Our results show significant advantages in image-text alignment and overall model performance. For example, VeCLIP achieves up to +25.2% gain in COCO and Flickr30k retrieval tasks under the 12M setting. For data efficiency, VeCLIP achieves +3% gain while only using 14% of the data employed in the vanilla CLIP and 11% in ALIGN. We also note the VeCap data is complementary with other well curated datasets good for zero-shot classification tasks. When combining VeCap and DFN, our model can achieve strong performance on both of image-text retrieval and zero-shot classification tasks, e.g. 83.1% accuracy@1 on ImageNet zero-shot for a H/14 model. We release the pre-trained models at https://github.com/apple/ml-veclip.

Liang Zhang, Jionghao Lin, John Sabatini et al. · cmu

Learning performance data describe correct and incorrect answers or problem-solving attempts in adaptive learning, such as in intelligent tutoring systems (ITSs). Learning performance data tend to be highly sparse (80\%\(\sim\)90\% missing observations) in most real-world applications due to adaptive item selection. This data sparsity presents challenges to using learner models to effectively predict future performance explore new hypotheses about learning. This article proposes a systematic framework for augmenting learner data to address data sparsity in learning performance data. First, learning performance is represented as a three-dimensional tensor of learners' questions, answers, and attempts, capturing longitudinal knowledge states during learning. Second, a tensor factorization method is used to impute missing values in sparse tensors of collected learner data, thereby grounding the imputation on knowledge tracing tasks that predict missing performance values based on real observations. Third, a module for generating patterns of learning is used. This study contrasts two forms of generative Artificial Intelligence (AI), including Generative Adversarial Networks (GANs) and Generate Pre-Trained Transformers (GPT) to generate data associated with different clusters of learner data. We tested this approach on an adult literacy dataset from AutoTutor lessons developed for Adult Reading Comprehension (ARC). We found that: (1) tensor factorization improved the performance in tracing and predicting knowledge mastery compared with other knowledge tracing techniques without data augmentation, showing higher relative fidelity for this imputation method, and (2) the GAN-based simulation showed greater overall stability and less statistical bias based on a divergence evaluation with varying simulation sample sizes compared to GPT.

Pavan Kumar Anasosalu Vasu, Cem Koc, Fartash Faghri et al. · utoronto

Streaming vision-language models (VLMs) continuously generate responses given an instruction prompt and an online stream of input frames. This is a core mechanism for real-time visual assistants. Existing VLM frameworks predominantly assess models in offline settings. In contrast, the performance of a streaming VLM depends on additional metrics beyond pure video understanding, including proactiveness, which reflects the timeliness of the model's responses, and consistency, which captures the robustness of its responses over time. To address this limitation, we propose VSAS-Bench, a new framework and benchmark for Visual Streaming Assistants. In contrast to prior benchmarks that primarily employ single-turn question answering on video inputs, VSAS-Bench features temporally dense annotations with over 18,000 annotations across diverse input domains and task types. We introduce standardized synchronous and asynchronous evaluation protocols, along with metrics that isolate and measure distinct capabilities of streaming VLMs. Using this framework, we conduct large-scale evaluations of recent video and streaming VLMs, analyzing the accuracy-latency trade-off under key design factors such as memory buffer length, memory access policy, and input resolution, yielding several practical insights. Finally, we show empirically that conventional VLMs can be adapted to streaming settings without additional training, and demonstrate that these adapted models outperform recent streaming VLMs. For example, Qwen3-VL-4B surpasses Dispider, the best streaming VLM on our benchmark, by 3% under the asynchronous protocol. The benchmark and code will be available at https://github.com/apple/ml-vsas-bench.

Haoping Bai, Shancong Mou, Tatiana Likhomanenko et al. · apple-ml

Despite progress in vision-based inspection algorithms, real-world industrial challenges -- specifically in data availability, quality, and complex production requirements -- often remain under-addressed. We introduce the VISION Datasets, a diverse collection of 14 industrial inspection datasets, uniquely poised to meet these challenges. Unlike previous datasets, VISION brings versatility to defect detection, offering annotation masks across all splits and catering to various detection methodologies. Our datasets also feature instance-segmentation annotation, enabling precise defect identification. With a total of 18k images encompassing 44 defect types, VISION strives to mirror a wide range of real-world production scenarios. By supporting two ongoing challenge competitions on the VISION Datasets, we hope to foster further advancements in vision-based industrial inspection.

Can Zhang, Tianyu Yang, Junwu Weng et al.

Unsupervised video representation learning has made remarkable achievements in recent years. However, most existing methods are designed and optimized for video classification. These pre-trained models can be sub-optimal for temporal localization tasks due to the inherent discrepancy between video-level classification and clip-level localization. To bridge this gap, we make the first attempt to propose a self-supervised pretext task, coined as Pseudo Action Localization (PAL) to Unsupervisedly Pre-train feature encoders for Temporal Action Localization tasks (UP-TAL). Specifically, we first randomly select temporal regions, each of which contains multiple clips, from one video as pseudo actions and then paste them onto different temporal positions of the other two videos. The pretext task is to align the features of pasted pseudo action regions from two synthetic videos and maximize the agreement between them. Compared to the existing unsupervised video representation learning approaches, our PAL adapts better to downstream TAL tasks by introducing a temporal equivariant contrastive learning paradigm in a temporally dense and scale-aware manner. Extensive experiments show that PAL can utilize large-scale unlabeled video data to significantly boost the performance of existing TAL methods. Our codes and models will be made publicly available at https://github.com/zhang-can/UP-TAL.

Meng Cao, Mehdi Fatemi, Jackie Chi Kit Cheung et al. · microsoft-research

With adversarial or otherwise normal prompts, existing large language models (LLM) can be pushed to generate toxic discourses. One way to reduce the risk of LLMs generating undesired discourses is to alter the training of the LLM. This can be very restrictive due to demanding computation requirements. Other methods rely on rule-based or prompt-based token elimination, which are limited as they dismiss future tokens and the overall meaning of the complete discourse. Here, we center detoxification on the probability that the finished discourse is ultimately considered toxic. That is, at each point, we advise against token selections proportional to how likely a finished text from this point will be toxic. To this end, we formally extend the dead-end theory from the recent reinforcement learning (RL) literature to also cover uncertain outcomes. Our approach, called rectification, utilizes a separate but significantly smaller model for detoxification, which can be applied to diverse LLMs as long as they share the same vocabulary. Importantly, our method does not require access to the internal representations of the LLM, but only the token probability distribution at each decoding step. This is crucial as many LLMs today are hosted in servers and only accessible through APIs. When applied to various LLMs, including GPT-3, our approach significantly improves the generated discourse compared to the base LLMs and other techniques in terms of both the overall language and detoxification performance.

Xiaozhong Lyu, Stefan Grafberger, Samantha Biegel et al. · eth-zurich

Retrieval augmentation enables large language models to take advantage of external knowledge, for example on tasks like question answering and data imputation. However, the performance of such retrieval-augmented models is limited by the data quality of their underlying retrieval corpus. In this paper, we propose an algorithm based on multilinear extension for evaluating the data importance of retrieved data points. There are exponentially many terms in the multilinear extension, and one key contribution of this paper is a polynomial time algorithm that computes exactly, given a retrieval-augmented model with an additive utility function and a validation set, the data importance of data points in the retrieval corpus using the multilinear extension of the model's utility function. We further proposed an even more efficient (ε, δ)-approximation algorithm. Our experimental results illustrate that we can enhance the performance of large language models by only pruning or reweighting the retrieval corpus, without requiring further training. For some tasks, this even allows a small model (e.g., GPT-JT), augmented with a search engine API, to outperform GPT-3.5 (without retrieval augmentation). Moreover, we show that weights based on multilinear extension can be computed efficiently in practice (e.g., in less than ten minutes for a corpus with 100 million elements).

Zidi Xiu, David Q. Sun, Kevin Cheng et al. · apple-ml

Next-generation AI must manage vast personal data, diverse tools, and multi-step reasoning, yet most benchmarks remain context-free and single-turn. We present ASTRA-bench (Assistant Skills in Tool-use, Reasoning \& Action-planning), a benchmark that uniquely unifies time-evolving personal context with an interactive toolbox and complex user intents. Our event-driven pipeline generates 2,413 scenarios across four protagonists, grounded in longitudinal life events and annotated by referential, functional, and informational complexity. Evaluation of state-of-the-art models (e.g., Claude-4.5-Opus, DeepSeek-V3.2) reveals significant performance degradation under high-complexity conditions, with argument generation emerging as the primary bottleneck. These findings expose critical limitations in current agents' ability to ground reasoning within messy personal context and orchestrate reliable multi-step plans. We release ASTRA-bench with a full execution environment and evaluation scripts to provide a diagnostic testbed for developing truly context-aware AI assistants.

Yu Lu Liu, Meng Cao, Su Lin Blodgett et al. · microsoft-research

AI and NLP publication venues have increasingly encouraged researchers to reflect on possible ethical considerations, adverse impacts, and other responsible AI issues their work might engender. However, for specific NLP tasks our understanding of how prevalent such issues are, or when and why these issues are likely to arise, remains limited. Focusing on text summarization -- a common NLP task largely overlooked by the responsible AI community -- we examine research and reporting practices in the current literature. We conduct a multi-round qualitative analysis of 333 summarization papers from the ACL Anthology published between 2020-2022. We focus on how, which, and when responsible AI issues are covered, which relevant stakeholders are considered, and mismatches between stated and realized research goals. We also discuss current evaluation practices and consider how authors discuss the limitations of both prior work and their own work. Overall, we find that relatively few papers engage with possible stakeholders or contexts of use, which limits their consideration of potential downstream adverse impacts or other responsible AI issues. Based on our findings, we make recommendations on concrete practices and research directions.

Guoli Yin, Haoping Bai, Shuang Ma et al.

Recent advances in large language models (LLMs) have increased the demand for comprehensive benchmarks to evaluate their capabilities as human-like agents. Existing benchmarks, while useful, often focus on specific application scenarios, emphasizing task completion but failing to dissect the underlying skills that drive these outcomes. This lack of granularity makes it difficult to deeply discern where failures stem from. Additionally, setting up these environments requires considerable effort, and issues of unreliability and reproducibility sometimes arise, especially in interactive tasks. To address these limitations, we introduce the Massive Multitask Agent Understanding (MMAU) benchmark, featuring comprehensive offline tasks that eliminate the need for complex environment setups. It evaluates models across five domains, including Tool-use, Directed Acyclic Graph (DAG) QA, Data Science and Machine Learning coding, Contest-level programming and Mathematics, and covers five essential capabilities: Understanding, Reasoning, Planning, Problem-solving, and Self-correction. With a total of 20 meticulously designed tasks encompassing over 3K distinct prompts, MMAU provides a comprehensive framework for evaluating the strengths and limitations of LLM agents. By testing 18 representative models on MMAU, we provide deep and insightful analyses. Ultimately, MMAU not only sheds light on the capabilities and limitations of LLM agents but also enhances the interpretability of their performance. Datasets and evaluation scripts of MMAU are released at https://github.com/apple/axlearn/tree/main/docs/research/mmau.

Peilin Zhou, Meng Cao, You-Liang Huang et al.

Large Multimodal Models (LMMs) have demonstrated impressive performance across various vision and language tasks, yet their potential applications in recommendation tasks with visual assistance remain unexplored. To bridge this gap, we present a preliminary case study investigating the recommendation capabilities of GPT-4V(ison), a recently released LMM by OpenAI. We construct a series of qualitative test samples spanning multiple domains and employ these samples to assess the quality of GPT-4V's responses within recommendation scenarios. Evaluation results on these test samples prove that GPT-4V has remarkable zero-shot recommendation abilities across diverse domains, thanks to its robust visual-text comprehension capabilities and extensive general knowledge. However, we have also identified some limitations in using GPT-4V for recommendations, including a tendency to provide similar responses when given similar inputs. This report concludes with an in-depth discussion of the challenges and research opportunities associated with utilizing GPT-4V in recommendation scenarios. Our objective is to explore the potential of extending LMMs from vision and language tasks to recommendation tasks. We hope to inspire further research into next-generation multimodal generative recommendation models, which can enhance user experiences by offering greater diversity and interactivity. All images and prompts used in this report will be accessible at https://github.com/PALIN2018/Evaluate_GPT-4V_Rec.

Meng Cao, Yue Dong, Jingyi He et al. · mila

State-of-the-art abstractive summarization systems frequently hallucinate content that is not supported by the source document, mainly due to noise in the training dataset. Existing methods opt to drop the noisy samples or tokens from the training set entirely, reducing the effective training set size and creating an artificial propensity to copy words from the source. In this work, we propose a training objective for abstractive summarization based on rejection learning, in which the model learns whether or not to reject potentially noisy tokens. We further propose a regularized decoding objective that penalizes non-factual candidate summaries during inference by using the rejection probability learned during training. We show that our method considerably improves the factuality of generated summaries in automatic and human evaluations when compared to five baseline models and that it does so while increasing the abstractiveness of the generated summaries.

Pengfei Zhu, Xinjie Yao, Yu Wang et al.

Multi-view learning has progressed rapidly in recent years. Although many previous studies assume that each instance appears in all views, it is common in real-world applications for instances to be missing from some views, resulting in incomplete multi-view data. To tackle this problem, we propose a novel Latent Heterogeneous Graph Network (LHGN) for incomplete multi-view learning, which aims to use multiple incomplete views as fully as possible in a flexible manner. By learning a unified latent representation, a trade-off between consistency and complementarity among different views is implicitly realized. To explore the complex relationship between samples and latent representations, a neighborhood constraint and a view-existence constraint are proposed, for the first time, to construct a heterogeneous graph. Finally, to avoid any inconsistencies between training and test phase, a transductive learning technique is applied based on graph learning for classification tasks. Extensive experimental results on real-world datasets demonstrate the effectiveness of our model over existing state-of-the-art approaches.

Meng Cao, Tianyu Yang, Junwu Weng et al.

Video-Text Pre-training (VTP) aims to learn transferable representations for various downstream tasks from large-scale web videos. To date, almost all existing VTP methods are limited to retrieval-based downstream tasks, e.g., video retrieval, whereas their transfer potentials on localization-based tasks, e.g., temporal grounding, are under-explored. In this paper, we experimentally analyze and demonstrate the incompatibility of current VTP methods with localization tasks, and propose a novel Localization-oriented Video-Text Pre-training framework, dubbed as LocVTP. Specifically, we perform the fine-grained contrastive alignment as a complement to the coarse-grained one by a clip-word correspondence discovery scheme. To further enhance the temporal reasoning ability of the learned feature, we propose a context projection head and a temporal aware contrastive loss to perceive the contextual relationships. Extensive experiments on four downstream tasks across six datasets demonstrate that our LocVTP achieves state-of-the-art performance on both retrieval-based and localization-based tasks. Furthermore, we conduct comprehensive ablation studies and thorough analyses to explore the optimum model designs and training strategies.

Weikai Li, Meng Cao, Songcan Chen

Unsupervised Source (data) Free domain adaptation (USFDA) aims to transfer knowledge from a well-trained source model to a related but unlabeled target domain. In such a scenario, all conventional adaptation methods that require source data fail. To combat this challenge, existing USFDAs turn to transfer knowledge by aligning the target feature to the latent distribution hidden in the source model. However, such information is naturally limited. Thus, the alignment in such a scenario is not only difficult but also insufficient, which degrades the target generalization performance. To relieve this dilemma in current USFDAs, we are motivated to explore a new perspective to boost their performance. For this purpose and gaining necessary insight, we look back upon the origin of the domain adaptation and first theoretically derive a new-brand target generalization error bound based on the model smoothness. Then, following the theoretical insight, a general and model-smoothness-guided Jacobian norm (JN) regularizer is designed and imposed on the target domain to mitigate this dilemma. Extensive experiments are conducted to validate its effectiveness. In its implementation, just with a few lines of codes added to the existing USFDAs, we achieve superior results on various benchmark datasets.

Hussain Hussain, Meng Cao, Sandipan Sikdar et al.

We present evidence for the existence and effectiveness of adversarial attacks on graph neural networks (GNNs) that aim to degrade fairness. These attacks can disadvantage a particular subgroup of nodes in GNN-based node classification, where nodes of the underlying network have sensitive attributes, such as race or gender. We conduct qualitative and experimental analyses explaining how adversarial link injection impairs the fairness of GNN predictions. For example, an attacker can compromise the fairness of GNN-based node classification by injecting adversarial links between nodes belonging to opposite subgroups and opposite class labels. Our experiments on empirical datasets demonstrate that adversarial fairness attacks can significantly degrade the fairness of GNN predictions (attacks are effective) with a low perturbation rate (attacks are efficient) and without a significant drop in accuracy (attacks are deceptive). This work demonstrates the vulnerability of GNN models to adversarial fairness attacks. We hope our findings raise awareness about this issue in our community and lay a foundation for the future development of GNN models that are more robust to such attacks.

Meng Cao, Ji Jiang, Long Chen et al.

We investigate the problem of video Referring Expression Comprehension (REC), which aims to localize the referent objects described in the sentence to visual regions in the video frames. Despite the recent progress, existing methods suffer from two problems: 1) inconsistent localization results across video frames; 2) confusion between the referent and contextual objects. To this end, we propose a novel Dual Correspondence Network (dubbed as DCNet) which explicitly enhances the dense associations in both the inter-frame and cross-modal manners. Firstly, we aim to build the inter-frame correlations for all existing instances within the frames. Specifically, we compute the inter-frame patch-wise cosine similarity to estimate the dense alignment and then perform the inter-frame contrastive learning to map them close in feature space. Secondly, we propose to build the fine-grained patch-word alignment to associate each patch with certain words. Due to the lack of this kind of detailed annotations, we also predict the patch-word correspondence through the cosine similarity. Extensive experiments demonstrate that our DCNet achieves state-of-the-art performance on both video and image REC benchmarks. Furthermore, we conduct comprehensive ablation studies and thorough analyses to explore the optimal model designs. Notably, our inter-frame and cross-modal contrastive losses are plug-and-play functions and are applicable to any video REC architectures. For example, by building on top of Co-grounding, we boost the performance by 1.48% absolute improvement on Accu.@0.5 for VID-Sentence dataset.

Hongxiang Li, Meng Cao, Xuxin Cheng et al.

The recent video grounding works attempt to introduce vanilla contrastive learning into video grounding. However, we claim that this naive solution is suboptimal. Contrastive learning requires two key properties: (1) \emph{alignment} of features of similar samples, and (2) \emph{uniformity} of the induced distribution of the normalized features on the hypersphere. Due to two annoying issues in video grounding: (1) the co-existence of some visual entities in both ground truth and other moments, \ie semantic overlapping; (2) only a few moments in the video are annotated, \ie sparse annotation dilemma, vanilla contrastive learning is unable to model the correlations between temporally distant moments and learned inconsistent video representations. Both characteristics lead to vanilla contrastive learning being unsuitable for video grounding. In this paper, we introduce Geodesic and Game Localization (G2L), a semantically aligned and uniform video grounding framework via geodesic and game theory. We quantify the correlations among moments leveraging the geodesic distance that guides the model to learn the correct cross-modal representations. Furthermore, from the novel perspective of game theory, we propose semantic Shapley interaction based on geodesic distance sampling to learn fine-grained semantic alignment in similar moments. Experiments on three benchmarks demonstrate the effectiveness of our method.

Yangjun Mao, Jun Xiao, Dong Zhang et al.

Distinctive Image Captioning (DIC) -- generating distinctive captions that describe the unique details of a target image -- has received considerable attention over the last few years. A recent DIC method proposes to generate distinctive captions by comparing the target image with a set of semantic-similar reference images, i.e., reference-based DIC (Ref-DIC). It aims to force the generated captions to distinguish between the target image and the reference image. To ensure Ref-DIC models really perceive the unique objects (or attributes) in target images, we propose two new Ref-DIC benchmarks and develop a Transformer-based Ref-DIC baseline TransDIC. The model only extracts visual features from the target image, but also encodes the differences between objects in the target and reference images. Taking one step further, we propose a stronger TransDIC++, which consists of an extra contrastive learning module to make full use of the reference images. This new module is model-agnostic, which can be easily incorporated into various Ref-DIC architectures. Finally, for more trustworthy benchmarking, we propose a new evaluation metric named DisCIDEr for Ref-DIC, which evaluates both the accuracy and distinctiveness of the generated captions. Experimental results demonstrate that our TransDIC++ can generate distinctive captions. Besides, it outperforms several state-of-the-art models on the two new benchmarks over different metrics.

Wentao Zhang, Yexin Wang, Zhenbang You et al.

Graph Neural Networks (GNNs) have achieved great success in various tasks, but their performance highly relies on a large number of labeled nodes, which typically requires considerable human effort. GNN-based Active Learning (AL) methods are proposed to improve the labeling efficiency by selecting the most valuable nodes to label. Existing methods assume an oracle can correctly categorize all the selected nodes and thus just focus on the node selection. However, such an exact labeling task is costly, especially when the categorization is out of the domain of individual expert (oracle). The paper goes further, presenting a soft-label approach to AL on GNNs. Our key innovations are: i) relaxed queries where a domain expert (oracle) only judges the correctness of the predicted labels (a binary question) rather than identifying the exact class (a multi-class question), and ii) new criteria of maximizing information gain propagation for active learner with relaxed queries and soft labels. Empirical studies on public datasets demonstrate that our method significantly outperforms the state-of-the-art GNN-based AL methods in terms of both accuracy and labeling cost.

Hongxiang Li, Meng Cao, Xuxin Cheng et al.

Video grounding aims to locate a moment of interest matching the given query sentence from an untrimmed video. Previous works ignore the {sparsity dilemma} in video annotations, which fails to provide the context information between potential events and query sentences in the dataset. In this paper, we contend that exploiting easily available captions which describe general actions, i.e., auxiliary captions defined in our paper, will significantly boost the performance. To this end, we propose an Auxiliary Caption Network (ACNet) for video grounding. Specifically, we first introduce dense video captioning to generate dense captions and then obtain auxiliary captions by Non-Auxiliary Caption Suppression (NACS). To capture the potential information in auxiliary captions, we propose Caption Guided Attention (CGA) project the semantic relations between auxiliary captions and query sentences into temporal space and fuse them into visual representations. Considering the gap between auxiliary captions and ground truth, we propose Asymmetric Cross-modal Contrastive Learning (ACCL) for constructing more negative pairs to maximize cross-modal mutual information. Extensive experiments on three public datasets (i.e., ActivityNet Captions, TACoS and ActivityNet-CG) demonstrate that our method significantly outperforms state-of-the-art methods.

Qichen Ye, Junling Liu, Dading Chong et al.

Integrating large language models (LLMs) into healthcare holds great potential but faces challenges. Pre-training LLMs from scratch for domains like medicine is resource-heavy and often unfeasible. On the other hand, sole reliance on Supervised Fine-tuning (SFT) can result in overconfident predictions and may not tap into domain-specific insights. In response, we present a multi-stage training method combining Domain-specific Continued Pre-training (DCPT), SFT, and Direct Preference Optimization (DPO). In addition, we publish a 3Gb Chinese Medicine (ChiMed) dataset, encompassing medical question answering, plain texts, knowledge graphs, and dialogues, segmented into three training stages. The medical LLM trained with our pipeline, Qilin-Med, shows substantial performance improvement. In the CPT and SFT phases, Qilin-Med achieved 38.4% and 40.0% accuracy on the CMExam test set, respectively. It outperformed the basemodel Baichuan-7B (accuracy: 33.5%), by 7.5%. In the DPO phase, it scored 16.66 in BLEU-1 and 27.44 in ROUGE-1 on the Huatuo-26M test set, bringing further improvement to the SFT phase (12.69 in BLEU-1 and 24.21 in ROUGE-1). Additionally, we have further enhanced the model's performance through the Retrieval Augmented Generation (RAG) approach. Experiments demonstrate that Qilin-Med-RAG achieves an accuracy rate of 42.8% on CMExam. These results highlight the contribution of our novel training approach in building LLMs for medical applications.

Ji Jiang, Meng Cao, Tengtao Song et al.

Video Referring Expression Comprehension (REC) aims to localize a target object in videos based on the queried natural language. Recent improvements in video REC have been made using Transformer-based methods with learnable queries. However, we contend that this naive query design is not ideal given the open-world nature of video REC brought by text supervision. With numerous potential semantic categories, relying on only a few slow-updated queries is insufficient to characterize them. Our solution to this problem is to create dynamic queries that are conditioned on both the input video and language to model the diverse objects referred to. Specifically, we place a fixed number of learnable bounding boxes throughout the frame and use corresponding region features to provide prior information. Also, we noticed that current query features overlook the importance of cross-modal alignment. To address this, we align specific phrases in the sentence with semantically relevant visual areas, annotating them in existing video datasets (VID-Sentence and VidSTG). By incorporating these two designs, our proposed model (called ConFormer) outperforms other models on widely benchmarked datasets. For example, in the testing split of VID-Sentence dataset, ConFormer achieves 8.75% absolute improvement on Accu.@0.6 compared to the previous state-of-the-art model.

Ji Jiang, Meng Cao, Tengtao Song et al.

Video Referring Expression Comprehension (REC) aims to localize a target object in video frames referred by the natural language expression. Recently, the Transformerbased methods have greatly boosted the performance limit. However, we argue that the current query design is suboptima and suffers from two drawbacks: 1) the slow training convergence process; 2) the lack of fine-grained alignment. To alleviate this, we aim to couple the pure learnable queries with the content information. Specifically, we set up a fixed number of learnable bounding boxes across the frame and the aligned region features are employed to provide fruitful clues. Besides, we explicitly link certain phrases in the sentence to the semantically relevant visual areas. To this end, we introduce two new datasets (i.e., VID-Entity and VidSTG-Entity) by augmenting the VIDSentence and VidSTG datasets with the explicitly referred words in the whole sentence, respectively. Benefiting from this, we conduct the fine-grained cross-modal alignment at the region-phrase level, which ensures more detailed feature representations. Incorporating these two designs, our proposed model (dubbed as ContFormer) achieves the state-of-the-art performance on widely benchmarked datasets. For example on VID-Entity dataset, compared to the previous SOTA, ContFormer achieves 8.75% absolute improvement on Accu.@0.6.

Shancong Mou, Xiaoyi Gu, Meng Cao et al.

Generative adversarial networks (GANs), trained on a large-scale image dataset, can be a good approximator of the natural image manifold. GAN-inversion, using a pre-trained generator as a deep generative prior, is a promising tool for image restoration under corruptions. However, the performance of GAN-inversion can be limited by a lack of robustness to unknown gross corruptions, i.e., the restored image might easily deviate from the ground truth. In this paper, we propose a Robust GAN-inversion (RGI) method with a provable robustness guarantee to achieve image restoration under unknown \textit{gross} corruptions, where a small fraction of pixels are completely corrupted. Under mild assumptions, we show that the restored image and the identified corrupted region mask converge asymptotically to the ground truth. Moreover, we extend RGI to Relaxed-RGI (R-RGI) for generator fine-tuning to mitigate the gap between the GAN learned manifold and the true image manifold while avoiding trivial overfitting to the corrupted input image, which further improves the image restoration and corrupted region mask identification performance. The proposed RGI/R-RGI method unifies two important applications with state-of-the-art (SOTA) performance: (i) mask-free semantic inpainting, where the corruptions are unknown missing regions, the restored background can be used to restore the missing content; (ii) unsupervised pixel-wise anomaly detection, where the corruptions are unknown anomalous regions, the retrieved mask can be used as the anomalous region's segmentation mask.

Shancong Mou, Meng Cao, Haoping Bai et al.

Unsupervised pixel-level defective region segmentation is an important task in image-based anomaly detection for various industrial applications. The state-of-the-art methods have their own advantages and limitations: matrix-decomposition-based methods are robust to noise but lack complex background image modeling capability; representation-based methods are good at defective region localization but lack accuracy in defective region shape contour extraction; reconstruction-based methods detected defective region match well with the ground truth defective region shape contour but are noisy. To combine the best of both worlds, we present an unsupervised patch autoencoder based deep image decomposition (PAEDID) method for defective region segmentation. In the training stage, we learn the common background as a deep image prior by a patch autoencoder (PAE) network. In the inference stage, we formulate anomaly detection as an image decomposition problem with the deep image prior and domain-specific regularizations. By adopting the proposed approach, the defective regions in the image can be accurately extracted in an unsupervised fashion. We demonstrate the effectiveness of the PAEDID method in simulation studies and an industrial dataset in the case study.

Haoran Tang, Meng Cao, Jinfa Huang et al.

Text-Video Retrieval (TVR) aims to align and associate relevant video content with corresponding natural language queries. Most existing TVR methods are based on large-scale pre-trained vision-language models (e.g., CLIP). However, due to the inherent plain structure of CLIP, few TVR methods explore the multi-scale representations which offer richer contextual information for a more thorough understanding. To this end, we propose MUSE, a multi-scale mamba with linear computational complexity for efficient cross-resolution modeling. Specifically, the multi-scale representations are generated by applying a feature pyramid on the last single-scale feature map. Then, we employ the Mamba structure as an efficient multi-scale learner to jointly learn scale-wise representations. Furthermore, we conduct comprehensive studies to investigate different model structures and designs. Extensive results on three popular benchmarks have validated the superiority of MUSE.

Hang Ding, Peidong Liu, Junqiao Wang et al.

The development of autonomous web agents, powered by Large Language Models (LLMs) and reinforcement learning (RL), represents a significant step towards general-purpose AI assistants. However, training these agents is severely hampered by the challenges of interacting with the live internet, which is inefficient, costly, and fraught with risks. Model-based reinforcement learning (MBRL) offers a promising solution by learning a world model of the environment to enable simulated interaction. This paper introduces DynaWeb, a novel MBRL framework that trains web agents through interacting with a web world model trained to predict naturalistic web page representations given agent actions. This model serves as a synthetic web environment where an agent policy can dream by generating vast quantities of rollout action trajectories for efficient online reinforcement learning. Beyond free policy rollouts, DynaWeb incorporates real expert trajectories from training data, which are randomly interleaved with on-policy rollouts during training to improve stability and sample efficiency. Experiments conducted on the challenging WebArena and WebVoyager benchmarks demonstrate that DynaWeb consistently and significantly improves the performance of state-of-the-art open-source web agent models. Our findings establish the viability of training web agents through imagination, offering a scalable and efficient way to scale up online agentic RL.

Tom Gunter, Zirui Wang, Chong Wang et al.

We present foundation language models developed to power Apple Intelligence features, including a ~3 billion parameter model designed to run efficiently on devices and a large server-based language model designed for Private Cloud Compute. These models are designed to perform a wide range of tasks efficiently, accurately, and responsibly. This report describes the model architecture, the data used to train the model, the training process, how the models are optimized for inference, and the evaluation results. We highlight our focus on Responsible AI and how the principles are applied throughout the model development.

Shancong Mou, Meng Cao, Zhendong Hong et al.

Display front-of-screen (FOS) quality inspection is essential for the mass production of displays in the manufacturing process. However, the severe imbalanced data, especially the limited number of defect samples, has been a long-standing problem that hinders the successful application of deep learning algorithms. Synthetic defect data generation can help address this issue. This paper reviews the state-of-the-art synthetic data generation methods and the evaluation metrics that can potentially be applied to display FOS quality inspection tasks.

Peilin Zhou, Bruce Leon, Xiang Ying et al.

As large language models (LLMs) evolve into tool-using agents, the ability to browse the web in real-time has become a critical yardstick for measuring their reasoning and retrieval competence. Existing benchmarks such as BrowseComp concentrate on English and overlook the linguistic, infrastructural, and censorship-related complexities of other major information ecosystems -- most notably Chinese. To address this gap, we introduce BrowseComp-ZH, a high-difficulty benchmark purpose-built to comprehensively evaluate LLM agents on the Chinese web. BrowseComp-ZH consists of 289 multi-hop questions spanning 11 diverse domains. Each question is reverse-engineered from a short, objective, and easily verifiable answer (e.g., a date, number, or proper noun). A two-stage quality control protocol is applied to strive for high question difficulty and answer uniqueness. We benchmark over 20 state-of-the-art language models and agentic search systems on our proposed BrowseComp-ZH. Despite their strong conversational and retrieval capabilities, most models struggle severely: a large number achieve accuracy rates below 10%, and only a handful exceed 20%. Even the best-performing system, OpenAI's DeepResearch, reaches just 42.9%. These results demonstrate the considerable difficulty of BrowseComp-ZH, where success demands not only effective retrieval strategies, but also sophisticated reasoning and information reconciliation -- capabilities that current models still struggle to master. Our dataset, construction guidelines, and benchmark results have been publicly released at https://github.com/PALIN2018/BrowseComp-ZH.

Meng Cao, Jiexi Liu, Songcan Chen

Imbalanced Domain Generalization (IDG) focuses on mitigating both domain and label shifts, both of which fundamentally shape the model's decision boundaries, particularly under heterogeneous long-tailed distributions across domains. Despite its practical significance, it remains underexplored, primarily due to the technical complexity of handling their entanglement and the paucity of theoretical foundations. In this paper, we begin by theoretically establishing the generalization bound for IDG, highlighting the role of posterior discrepancy and decision margin. This bound motivates us to focus on directly steering decision boundaries, marking a clear departure from existing methods. Subsequently, we technically propose a novel Negative-Dominant Contrastive Learning (NDCL) for IDG to enhance discriminability while enforce posterior consistency across domains. Specifically, inter-class decision-boundary separation is enhanced by placing greater emphasis on negatives as the primary signal in our contrastive learning, naturally amplifying gradient signals for minority classes to avoid the decision boundary being biased toward majority classes. Meanwhile, intra-class compactness is encouraged through a re-weighted cross-entropy strategy, and posterior consistency across domains is enforced through a prediction-central alignment strategy. Finally, rigorous yet challenging experiments on benchmarks validate the effectiveness of our NDCL. The code is available at https://github.com/Alrash/NDCL.

Jiahua Dong, Wenqi Liang, Hongliu Li et al.

Custom diffusion models (CDMs) have attracted widespread attention due to their astonishing generative ability for personalized concepts. However, most existing CDMs unreasonably assume that personalized concepts are fixed and cannot change over time. Moreover, they heavily suffer from catastrophic forgetting and concept neglect on old personalized concepts when continually learning a series of new concepts. To address these challenges, we propose a novel Concept-Incremental text-to-image Diffusion Model (CIDM), which can resolve catastrophic forgetting and concept neglect to learn new customization tasks in a concept-incremental manner. Specifically, to surmount the catastrophic forgetting of old concepts, we develop a concept consolidation loss and an elastic weight aggregation module. They can explore task-specific and task-shared knowledge during training, and aggregate all low-rank weights of old concepts based on their contributions during inference. Moreover, in order to address concept neglect, we devise a context-controllable synthesis strategy that leverages expressive region features and noise estimation to control the contexts of generated images according to user conditions. Experiments validate that our CIDM surpasses existing custom diffusion models. The source codes are available at https://github.com/JiahuaDong/CIFC.

Xiaoshuang Huang, Hongxiang Li, Meng Cao et al.

Recent developments underscore the potential of textual information in enhancing learning models for a deeper understanding of medical visual semantics. However, language-guided medical image segmentation still faces a challenging issue. Previous works employ implicit and ambiguous architectures to embed textual information. This leads to segmentation results that are inconsistent with the semantics represented by the language, sometimes even diverging significantly. To this end, we propose a novel cross-modal conditioned Reconstruction for Language-guided Medical Image Segmentation (RecLMIS) to explicitly capture cross-modal interactions, which assumes that well-aligned medical visual features and medical notes can effectively reconstruct each other. We introduce conditioned interaction to adaptively predict patches and words of interest. Subsequently, they are utilized as conditioning factors for mutual reconstruction to align with regions described in the medical notes. Extensive experiments demonstrate the superiority of our RecLMIS, surpassing LViT by 3.74% mIoU on the publicly available MosMedData+ dataset and achieving an average increase of 1.89% mIoU for cross-domain tests on our QATA-CoV19 dataset. Simultaneously, we achieve a relative reduction of 20.2% in parameter count and a 55.5% decrease in computational load. The code will be available at https://github.com/ShashankHuang/RecLMIS.

Meng Cao, Haoran Tang, Haoze Zhao et al.

Recent advancements in video-based large language models (Video LLMs) have witnessed the emergence of diverse capabilities to reason and interpret dynamic visual content. Among them, gameplay videos stand out as a distinctive data source, often containing glitches that defy physics commonsense. This characteristic renders them an effective benchmark for assessing the under-explored capability of physical commonsense understanding in video LLMs. In this paper, we propose PhysGame as a pioneering benchmark to evaluate physical commonsense violations in gameplay videos. PhysGame comprises 880 videos associated with glitches spanning four fundamental domains (i.e., mechanics, kinematics, optics, and material properties) and across 12 distinct physical commonsense. Through extensively evaluating various state-ofthe-art video LLMs, our findings reveal that the performance of current open-source video LLMs significantly lags behind that of proprietary counterparts. To bridge this gap, we curate an instruction tuning dataset PhysInstruct with 140,057 question-answering pairs to facilitate physical commonsense learning. In addition, we also propose a preference optimization dataset PhysDPO with 34,358 training pairs, where the dis-preferred responses are generated conditioned on misleading titles (i.e., meta information hacking), fewer frames (i.e., temporal hacking) and lower spatial resolutions (i.e., spatial hacking). Based on the suite of datasets, we propose PhysVLM as a physical knowledge-enhanced video LLM. Extensive experiments on both physical-oriented benchmark PhysGame and general video understanding benchmarks demonstrate the state-ofthe-art performance of PhysVLM.

Zhiyang Xu, Tian Qin, Bowen Jin et al.

Multimodal large language models (MLLMs) have recently shown strong performance in visual understanding, yet they often lack temporal awareness, particularly in egocentric settings where reasoning depends on the correct ordering and evolution of events. This deficiency stems in part from training objectives that fail to explicitly reward temporal reasoning and instead rely on frame-level spatial shortcuts. To address this limitation, we propose Temporal Global Policy Optimization (TGPO), a reinforcement learning with verifiable rewards (RLVR) algorithm designed to incentivize temporal awareness in MLLMs. TGPO contrasts model outputs generated from temporally ordered versus shuffled video frames to derive calibrated, globally normalized reward signals that explicitly favor temporally coherent reasoning. Integrated with GRPO and GSPO, TGPO supports cold-start RL training and effectively suppresses spatial shortcut behaviors learned by existing MLLMs. Experiments across five egocentric video benchmarks demonstrate that TGPO consistently improves temporal grounding and causal coherence, outperforming prior RL-based video reasoning approaches. Our results suggest that TGPO offers a simple and scalable pathway toward temporally robust MLLMs for egocentric video understanding.

Yongxin Wang, Meng Cao, Haokun Lin et al.

Multimodal large language models (MLLMs) have achieved remarkable progress on various visual question answering and reasoning tasks leveraging instruction fine-tuning specific datasets. They can also learn from preference data annotated by human to enhance their reasoning ability and mitigate hallucinations. Most of preference data is generated from the model itself. However, existing methods require high-quality critical labels, which are costly and rely on human or proprietary models like GPT-4V. In this work, we propose Enhancing Alignment in MLLMs via Critical Observation (EACO), which aligns MLLMs by self-generated preference data using only 5k images economically. Our approach begins with collecting and refining a Scoring Evaluation Instruction-tuning dataset to train a critical evaluation model, termed the Critic. This Critic observes model responses across multiple dimensions, selecting preferred and non-preferred outputs for refined Direct Preference Optimization (DPO) tuning. To further enhance model performance, we employ an additional supervised fine-tuning stage after preference tuning. EACO reduces the overall hallucinations by 65.6% on HallusionBench and improves the reasoning ability by 21.8% on MME-Cognition. EACO achieves an 8.5% improvement over LLaVA-v1.6-Mistral-7B across multiple benchmarks. Remarkably, EACO also shows the potential critical ability in open-source MLLMs, demonstrating that EACO is a viable path to boost the competence of MLLMs.

Meng Cao, Mehdi Fatemi, Jackie Chi Kit Cheung et al.

While large language models (LLMs) have achieved impressive performance in generating fluent and realistic text, controlling the generated text so that it exhibits properties such as safety, factuality, and non-toxicity remains challenging. % such as DExperts, GeDi, and rectification Existing decoding-based methods are static in terms of the dimension of control; if the target subject is changed, they require new training. Moreover, it can quickly become prohibitive to concurrently control multiple subjects. In this work, we introduce SF-GEN, which is grounded in two primary concepts: successor features (SFs) to decouple the LLM's dynamics from task-specific rewards, and language model rectification to proportionally adjust the probability of selecting a token based on the likelihood that the finished text becomes undesired. SF-GEN seamlessly integrates the two to enable dynamic steering of text generation with no need to alter the LLM's parameters. Thanks to the decoupling effect induced by successor features, our method proves to be memory-wise and computationally efficient for training as well as decoding, especially when dealing with multiple target subjects. To the best of our knowledge, our research represents the first application of successor features in text generation. In addition to its computational efficiency, the resultant language produced by our method is comparable to the SOTA (and outperforms baselines) in both control measures as well as language quality, which we demonstrate through a series of experiments in various controllable text generation tasks.

Xiuwei Chen, Wentao Hu, Hanhui Li et al.

Recent advances in multimodal large language models (MLLMs) have shown impressive reasoning capabilities. However, further enhancing existing MLLMs necessitates high-quality vision-language datasets with carefully curated task complexities, which are both costly and challenging to scale. Although recent self-improving models that iteratively refine themselves offer a feasible solution, they still suffer from two core challenges: (i) most existing methods augment visual or textual data separately, resulting in discrepancies in data complexity (e.g., over-simplified diagrams paired with redundant textual descriptions); and (ii) the evolution of data and models is also separated, leading to scenarios where models are exposed to tasks with mismatched difficulty levels. To address these issues, we propose C2-Evo, an automatic, closed-loop self-improving framework that jointly evolves both training data and model capabilities. Specifically, given a base dataset and a base model, C2-Evo enhances them by a cross-modal data evolution loop and a data-model evolution loop. The former loop expands the base dataset by generating complex multimodal problems that combine structured textual sub-problems with iteratively specified geometric diagrams, while the latter loop adaptively selects the generated problems based on the performance of the base model, to conduct supervised fine-tuning and reinforcement learning alternately. Consequently, our method continuously refines its model and training data, and consistently obtains considerable performance gains across multiple mathematical reasoning benchmarks. Our code, models, and datasets will be released.

David Ma, Yuanxing Zhang, Jincheng Ren et al.

Existing evaluation frameworks for Multimodal Large Language Models (MLLMs) primarily focus on image reasoning or general video understanding tasks, largely overlooking the significant role of image context in video comprehension. To bridge this gap, we propose IV-Bench, the first comprehensive benchmark for evaluating Image-Grounded Video Perception and Reasoning. IV-Bench consists of 967 videos paired with 2,585 meticulously annotated image-text queries across 13 tasks (7 perception and 6 reasoning tasks) and 5 representative categories. Extensive evaluations of state-of-the-art open-source (e.g., InternVL2.5, Qwen2.5-VL) and closed-source (e.g., GPT-4o, Gemini2-Flash and Gemini2-Pro) MLLMs demonstrate that current models substantially underperform in image-grounded video Perception and Reasoning, merely achieving at most 28.9% accuracy. Further analysis reveals key factors influencing model performance on IV-Bench, including inference pattern, frame number, and resolution. Additionally, through a simple data synthesis approach, we demonstratethe challenges of IV- Bench extend beyond merely aligning the data format in the training proecss. These findings collectively provide valuable insights for future research. Our codes and data are released in https://github.com/multimodal-art-projection/IV-Bench.

Kaidong Zhang, Jian Zhang, Rongtao Xu et al.

Vision--Language--Action (VLA) models have emerged as a powerful paradigm for open-world robot manipulation, but their practical deployment is often constrained by \emph{cost}: billion-scale VLM backbones and iterative diffusion/flow-based action heads incur high latency and compute, making real-time control expensive on commodity hardware. We present A1, a fully open-source and transparent VLA framework designed for low-cost, high-throughput inference without sacrificing manipulation success; Our approach leverages pretrained VLMs that provide implicit affordance priors for action generation. We release the full training stack (training code, data/data-processing pipeline, intermediate checkpoints, and evaluation scripts) to enable end-to-end reproducibility. Beyond optimizing the VLM alone, A1 targets the full inference pipeline by introducing a budget-aware adaptive inference scheme that jointly accelerates the backbone and the \emph{action head}. Specifically, we monitor action consistency across intermediate VLM layers to trigger early termination, and propose Inter-Layer Truncated Flow Matching that warm-starts denoising across layers, enabling accurate actions with substantially fewer effective denoising iterations. Across simulation benchmarks (LIBERO, VLABench) and real robots (Franka, AgiBot), A1 achieves state-of-the-art success rates while significantly reducing inference cost (e.g., up to 72% lower per-episode latency for flow-matching inference and up to 76.6% backbone computation reduction with minor performance degradation). On RoboChallenge, A1 achieves an average success rate of 29.00%, outperforming baselines including pi0(28.33%), X-VLA (21.33%), and RDT-1B (15.00%).

Meng Cao, Haoran Tang, Haoze Zhao et al.

Understanding the physical world, including object dynamics, material properties, and causal interactions, remains a core challenge in artificial intelligence. Although recent multi-modal large language models (MLLMs) have demonstrated impressive general reasoning capabilities, they still fall short of achieving human-level understanding of physical principles. Existing datasets for physical reasoning either rely on real-world videos, which incur high annotation costs, or on synthetic simulations, which suffer from limited realism and diversity. In this paper, we propose a novel paradigm that leverages glitches in gameplay videos, referring to visual anomalies that violate predefined physical laws, as a rich and scalable supervision source for physical world understanding. We introduce PhysGame, an meta information guided instruction-tuning dataset containing 140,057 glitch-centric question-answer pairs across five physical domains and sixteen fine-grained categories. To ensure data accuracy, we design a prompting strategy that utilizes gameplay metadata such as titles and descriptions to guide high-quality QA generation. Complementing PhysGame, we construct GameBench, an expert-annotated benchmark with 880 glitch-identified gameplay videos designed to evaluate physical reasoning capabilities. Extensive experiments show that PhysGame significantly enhances both Game2Real transferability, improving the real world physical reasoning performance of Qwen2.5VL by 2.5% on PhysBench, and Game2General transferability, yielding a 1.9% gain on the MVBench benchmark. Moreover, PhysGame-tuned models achieve a 3.7% absolute improvement on GameBench, demonstrating enhanced robustness in detecting physical implausibilities. These results indicate that learning from gameplay anomalies offers a scalable and effective pathway toward advancing physical world understanding in multimodal intelligence.

Yu Sun, Meng Cao, Ping Yang et al.

Vision-Language-Action (VLA) models and world models have recently emerged as promising paradigms for general-purpose robotic intelligence, yet their progress is hindered by the lack of reliable evaluation protocols that reflect real-world deployment. Existing benchmarks are largely simulator-centric, which provide controllability but fail to capture the reality gap caused by perception noise, complex contact dynamics, hardware constraints, and system latency. Moreover, fragmented real-world evaluations across different robot platforms prevent fair and reproducible comparison. To address these challenges, we introduce ManipArena, a standardized evaluation framework designed to bridge simulation and real-world execution. ManipArena comprises 20 diverse tasks across 10,812 expert trajectories emphasizing reasoning-oriented manipulation tasks requiring semantic and spatial reasoning, supports multi-level generalization through controlled out-of-distribution settings, and incorporates long-horizon mobile manipulation beyond tabletop scenarios. The framework further provides rich sensory diagnostics, including low-level motor signals, and synchronized real-to-sim environments constructed via high-quality 3D scanning. Together, these features enable fair, realistic, and reproducible evaluation for both VLA and world model approaches, providing a scalable foundation for diagnosing and advancing embodied intelligence systems.

Meng Cao

Automatic summarization is the process of shortening a set of textual data computationally, to create a subset (a summary) that represents the most important pieces of information in the original text. Existing summarization methods can be roughly divided into two types: extractive and abstractive. An extractive summarizer explicitly selects text snippets (words, phrases, sentences, etc.) from the source document, while an abstractive summarizer generates novel text snippets to convey the most salient concepts prevalent in the source.

David Ma, Huaqing Yuan, Xingjian Wang et al.

Although long-video understanding demands that models capture hierarchical temporal information -- from clip (seconds) and shot (tens of seconds) to event (minutes) and story (hours) -- existing benchmarks either neglect this multi-scale design or scatter scale-specific questions across different videos, preventing direct comparison of model performance across timescales on the same content. To address this, we introduce ScaleLong, the first benchmark to disentangle these factors by embedding questions targeting four hierarchical timescales -- clip (seconds), shot (tens of seconds), event (minutes), and story (hours) -- all within the same video content. This within-content multi-timescale questioning design enables direct comparison of model performance across timescales on identical videos. ScaleLong features 269 long videos (avg.\ 86\,min) from 5 main categories and 36 sub-categories, with 4--8 carefully designed questions, including at least one question for each timescale. Evaluating 23 MLLMs reveals a U-shaped performance curve, with higher accuracy at the shortest and longest timescales and a dip at intermediate levels. Furthermore, ablation studies show that increased visual token capacity consistently enhances reasoning across all timescales. ScaleLong offers a fine-grained, multi-timescale benchmark for advancing MLLM capabilities in long-video understanding. The code and dataset are available https://github.com/multimodal-art-projection/ScaleLong.

Jihao Gu, Yingyao Wang, Pi Bu et al.

The evaluation of factual accuracy in large vision language models (LVLMs) has lagged behind their rapid development, making it challenging to fully reflect these models' knowledge capacity and reliability. In this paper, we introduce the first factuality-based visual question-answering benchmark in Chinese, named ChineseSimpleVQA, aimed at assessing the visual factuality of LVLMs across 8 major topics and 56 subtopics. The key features of this benchmark include a focus on the Chinese language, diverse knowledge types, a multi-hop question construction, high-quality data, static consistency, and easy-to-evaluate through short answers. Moreover, we contribute a rigorous data construction pipeline and decouple the visual factuality into two parts: seeing the world (i.e., object recognition) and discovering knowledge. This decoupling allows us to analyze the capability boundaries and execution mechanisms of LVLMs. Subsequently, we evaluate 34 advanced open-source and closed-source models, revealing critical performance gaps within this field. Our evaluation-friendly code and data have already been open-sourced.

Junbin Zhang, Meng Cao, Feng Tan et al.

Achieving fine-grained and structurally sound controllability is a cornerstone of advanced visual generation. Existing part-based frameworks treat user-provided parts as an unordered set and therefore ignore their intrinsic spatial and semantic relationships, which often results in compositions that lack structural integrity. To bridge this gap, we propose Graph-PiT, a framework that explicitly models the structural dependencies of visual components using a graph prior. Specifically, we represent visual parts as nodes and their spatial-semantic relationships as edges. At the heart of our method is a Hierarchical Graph Neural Network (HGNN) module that performs bidirectional message passing between coarse-grained part-level super-nodes and fine-grained IP+ token sub-nodes, refining part embeddings before they enter the generative pipeline. We also introduce a graph Laplacian smoothness loss and an edge-reconstruction loss so that adjacent parts acquire compatible, relation-aware embeddings. Quantitative experiments on controlled synthetic domains (character, product, indoor layout, and jigsaw), together with qualitative transfer to real web images, show that Graph-PiT improves structural coherence over vanilla PiT while remaining compatible with the original IP-Prior pipeline. Ablation experiments confirm that explicit relational reasoning is crucial for enforcing user-specified adjacency constraints. Our approach not only enhances the plausibility of generated concepts but also offers a scalable and interpretable mechanism for complex, multi-part image synthesis. The code is available at https://github.com/wolf-bailang/Graph-PiT.

Zhangci Xiong, Meng Cao

In recent years, Multi-View Clustering (MVC) has attracted increasing attention for its potential to reduce the annotation burden associated with large datasets. The aim of MVC is to exploit the inherent consistency and complementarity among different views, thereby integrating information from multiple perspectives to improve clustering outcomes. Despite extensive research in MVC, most existing methods focus predominantly on harnessing complementary information across views to enhance clustering effectiveness, often neglecting the structural information among samples, which is crucial for exploring sample correlations. To address this gap, we introduce a novel framework, termed Structured Latent Representation Learning based Multi-View Clustering method (SLRL). SLRL leverages both the complementary and structural information. Initially, it learns a common latent representation for all views. Subsequently, to exploit the structural information among samples, a k-nearest neighbor graph is constructed from this common latent representation. This graph facilitates enhanced sample interaction through graph learning techniques, leading to a structured latent representation optimized for clustering. Extensive experiments demonstrate that SLRL not only competes well with existing methods but also sets new benchmarks in various multi-view datasets.

Pengfei Hu, Meng Cao, Yingyao Wang et al.

Long video understanding is essential for human-like intelligence, enabling coherent perception and reasoning over extended temporal contexts. While the emerging thinking-with-frames paradigm, which alternates between global temporal reasoning and local frame examination, has advanced the reasoning capabilities of video multi-modal large language models (MLLMs), it suffers from a significant efficiency bottleneck due to the progressively growing and redundant multi-modal context. To address this, we propose SpecTemp, a reinforcement learning-based Speculative Temporal reasoning framework that decouples temporal perception from reasoning via a cooperative dual-model design. In SpecTemp, a lightweight draft MLLM rapidly explores and proposes salient frames from densely sampled temporal regions, while a powerful target MLLM focuses on temporal reasoning and verifies the draft's proposals, iteratively refining its attention until convergence. This design mirrors the collaborative pathways of the human brain, balancing efficiency with accuracy. To support training, we construct the SpecTemp-80K dataset, featuring synchronized dual-level annotations for coarse evidence spans and fine-grained frame-level evidence. Experiments across multiple video understanding benchmarks demonstrate that SpecTemp not only maintains competitive accuracy but also significantly accelerates inference compared with existing thinking-with-frames methods.

Meng Cao, Haokun Lin, Haoyuan Li et al.

Spatial reasoning, the ability to understand and interpret the 3D structure of the world, is a critical yet underdeveloped capability in Multimodal Large Language Models (MLLMs). Current methods predominantly rely on verbal descriptive tuning, which suffers from visual illiteracy, i.e., they learn spatial concepts through textual symbols alone, devoid of connection to their visual manifestations. To bridge this gap, this paper introduces MILO, an Implicit spatIaL wOrld modeling paradigm that simulates human-like spatial imagination. MILO integrates a visual generator to provide geometry-aware feedback, thereby implicitly grounding the MLLM's symbolic reasoning in perceptual experience. Complementing this paradigm, we propose RePE (Relative Positional Encoding), a novel encoding scheme that captures relative camera-pose transformations, offering superior performance over absolute coordinate systems. To support the training, we construct GeoGen, a large-scale Geometry-aware Generative dataset with approximately 2,241 videos and 67,827 observation-action-outcome triplets. Experiments demonstrate that our approach significantly enhances spatial reasoning capabilities across multiple baselines and benchmarks, offering a more holistic understanding of 3D space.