Cross-Video Reasoning (CVR) presents a significant challenge in video understanding, which requires simultaneous understanding of multiple videos to aggregate and compare information across groups of videos. Most existing video understanding benchmarks focus on single-video analysis, failing to assess the ability of multimodal large language models (MLLMs) to simultaneously reason over various videos. Recent benchmarks evaluate MLLMs' capabilities on multi-view videos that capture different perspectives of the same scene. However, their limited tasks hinder a thorough assessment of MLLMs in diverse real-world CVR scenarios. To this end, we introduce CrossVid, the first benchmark designed to comprehensively evaluate MLLMs' spatial-temporal reasoning ability in cross-video contexts. Firstly, CrossVid encompasses a wide spectrum of hierarchical tasks, comprising four high-level dimensions and ten specific tasks, thereby closely reflecting the complex and varied nature of real-world video understanding. Secondly, CrossVid provides 5,331 videos, along with 9,015 challenging question-answering pairs, spanning single-choice, multiple-choice, and open-ended question formats. Through extensive experiments on various open-source and closed-source MLLMs, we observe that Gemini-2.5-Pro performs best on CrossVid, achieving an average accuracy of 50.4%. Notably, our in-depth case study demonstrates that most current MLLMs struggle with CVR tasks, primarily due to their inability to integrate or compare evidence distributed across multiple videos for reasoning. These insights highlight the potential of CrossVid to guide future advancements in enhancing MLLMs’ CVR capabilities.

Wei Wu, Liyi Chen, Congxi Xiao, Tianfu Wang, Qimeng Wang, Chengqiang Lu, Yan Gao, Yiwu, Yao Hu, Hui Xiong. Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2026.

Peiwen Yuan, Shaoxiong Feng, Yiwei Li, Xinglin Wang, Yueqi Zhang, Jiayi Shi, Chuyi Tan, Boyuan Pan, Yao Hu, Kan Li. Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2026.

Zhiwei Zhang, Fei Zhao, Rui Wang, Zezhong Wang, Bin Liang, Jiakang Wang, Yao Hu, Shaosheng Cao, Kam-Fai Wong. Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2026.

Xuyang Zhi, Peilun Zhou, Chengqiang Lu, Hang Lv, Yiwei Liang, Rongyang Zhang, Yan Gao, Yiwu, Yao Hu, Hongchao Gu, Defu Lian, Hao Wang, Enhong Chen. Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2026.

Vector approximate nearest neighbor search (ANNS) underpins search engines, recommendation systems, and advertising services. Recent advances in ANNS indexes make CPU a cost-effective choice for serving million-scale, in-memory vector search, yet per-core throughput remains constrained by memory access latency of vector reading and the compute intensity of distance evaluations in production deployments. With the growing scale of the business and advances in hardware, modern CCD-based multi-core CPUs have been widely deployed for high throughput in our services. However, we find that simply increasing core counts does not yield optimal performance scaling. To improve the efficiency of more cores from the CCD-based architecture, we analyze the distributions of real-world requests in our production environments. We observe high access locality in vector search in our online services and low cache utilization, resulting from overlooking the multi-chiplet nature of CCD based CPUs. Hence, we propose a workload- and hardware-aware thread orchestration framework at CCD-level that (i) provides a uniform interface for both inter-query parallel HNSW search and intra-query parallel IVF search, (ii) achieves cache-friendly and workload-adaptive mapping of task dispatching, and (iii) employs CCD-aware task stealing to address load imbalance. Applied to real production workloads from search, recommendation, and advertising services of Xiaohongshu (RedNote), our approach delivers up to 3.7x higher throughput and 30-90% reductions in P50 and P999 latency. In detail, compared with the original framework, the cache-miss ratio decreases by 6-30%, and the total CPU stall is reduced by 20-80%.

Ranking relevance is a fundamental task in search engines, aiming to identify the items most relevant to a given user query. Traditional relevance models typically produce scalar scores or directly predict relevance labels, limiting both interpretability and the modeling of complex relevance signals. Inspired by recent advances in Chain-of-Thought (CoT) reasoning for complex tasks, we investigate whether explicit reasoning can enhance both interpretability and performance in relevance modeling. However, existing reasoning-based Generative Relevance Models (GRMs) primarily rely on supervised fine-tuning on large amounts of human-annotated or synthetic CoT data, which often leads to limited generalization. Moreover, domain-agnostic, free-form reasoning tends to be overly generic and insufficiently grounded, limiting its potential to handle the diverse and ambiguous cases prevalent in open-domain search. In this work, we formulate relevance modeling in Xiaohongshu search as a reasoning task and introduce a Reinforcement Learning (RL)-based training framework to enhance the grounded reasoning capabilities of GRMs. Specifically, we incorporate practical business-specific relevance criteria into the multi-step reasoning prompt design and propose Stepwise Advantage Masking (SAM), a lightweight process-supervision strategy which facilitates effective learning of these criteria through improved credit assignment. To enable industrial deployment, we further distill the large-scale RL-tuned model to a lightweight version suitable for real-world search systems. Extensive offline evaluations and online A/B tests demonstrate that our approach consistently delivers significant improvements across key relevance and business metrics, validating its effectiveness, robustness, and practicality for large-scale industrial search systems.

Recently, the generative retrieval paradigm has emerged as a transformative framework that significantly enhances the efficiency of large-scale industrial recommendation systems. This innovative approach systematically maps items to meaningful semantic identifiers (SIDs) and employs advanced sequence generation techniques to construct high-quality candidate sets, thereby enabling more accurate modeling of users' evolving interests and behavioral patterns. Nevertheless, two critical challenges remain inadequately addressed in current research: (1) Existing methodologies predominantly focus on modeling a single task such as predicting users' click behavior, while overlooking other tasks including predicting users' dwell-time and engagement behaviors, which are very important for video/content recommendation at the same time. The independent modeling of each task inevitably results in substantial computational overhead, thereby raising the pivotal question of whether the sophisticated multi-task learning capabilities inherent in LLMs can be effectively leveraged to achieve unified and efficient multi-task learning for generative retrieval. (2) The mapping mechanism from SIDs to concrete items requires substantial refinement to ensure precise and reliable retrieval performance. To tackle these issues, we propose RedGR, a generative retrieval model that unifies the modeling of multiple complex retrieval tasks. RedGR first applies the RQ-Kmeans algorithm to map items into SIDs, and then conducts pre-training on large-scale user behavior datasets to learn general knowledge. Then the RedGR model is finetuned on multi-task retrieval data with a unique instruction prompt for each task. This enables RedGR to generate the corresponding set of SIDs for each task. And the union of all sets of SIDs is the multi-task retrieval result. Finally, the Swing algorithm incorporates explicit, high-quality collaborative signals to strengthen the mapping from SIDs to specific items, thereby facilitating efficient retrieval of high-quality items. RedGR has been fully depolyed in the homefeed recommendation scenario of RedNote,serving hundreds of millions of users every day. Online A/B test results show a 0.178% increase in pagetime, a 0.734% increase in average user engagement, and a 0.076% growth in homefeed active users (FAU). These metrics collectively validate the superior performance of RedGR's unified retrieval modeling approach in complex multi-task scenarios.

Multimodal recommendation improves user modeling by integrating collaborative signals with heterogeneous item content. In real applications, user interests evolve over time and exhibit nonstationary dynamics, where different preference factors change at different rates. This challenge is amplified in multimodal settings because visual and textual cues can dominate decisions under different temporal regimes. Despite strong progress, most multimodal recommenders still rely on static interaction graphs or coarse temporal heuristics, which limits their ability to model continuous preference evolution with fine-grained temporal adaptation. To address these limitations, we propose TimeMM, a time-conditioned spectral filtering framework for dynamic multimodal recommendation. TimeMM instantiates Time-as-Operator by mapping interaction recency to a family of parametric temporal kernels that reweight edges on the user--item graph, producing component-specific representations without explicit eigendecomposition. To capture non-stationary interests, we introduce Adaptive Spectral Filtering that mixes the operator bank according to temporal context, yielding prediction-specific effective spectral responses. To account for modality-specific temporal sensitivity, we further propose Spectral-Aware Modality Routing that calibrates visual and textual contributions conditioned on the same temporal context. Finally, a ranking-space Spectral Diversity Regularization encourages complementary expert behaviors and prevents filter-bank collapse. Extensive experiments on real-world benchmarks demonstrate that TimeMM consistently outperforms state-of-the-art multimodal recommenders while maintaining linear-time scalability.

Large language models (LLMs) have recently demonstrated strong potential for sequential recommendation. However, current LLM-based approaches face critical limitations in modeling users' long-term and diverse interests. First, due to inference latency and feature fetching bandwidth constraints, existing methods typically truncate user behavior sequences to include only the most recent interactions, resulting in the loss of valuable long-range preference signals. Second, most current methods rely on next-item prediction with a single predicted embedding, overlooking the multifaceted nature of user interests and limiting recommendation diversity. To address these challenges, we propose HyMiRec, a hybrid multi-interest sequential recommendation framework, which leverages a lightweight recommender to extracts coarse interest embeddings from long user sequences and an LLM-based recommender to captures refined interest embeddings. To alleviate the overhead of fetching features, we introduce a residual codebook based on cosine similarity, enabling efficient compression and reuse of user history embeddings. To model the diverse preferences of users, we design a disentangled multi-interest learning module, which leverages multiple interest queries to learn disentangles multiple interest signals adaptively, allowing the model to capture different facets of user intent. Extensive experiments are conducted on both benchmark datasets and a collected industrial dataset, demonstrating our effectiveness over existing state-of-the-art methods. Furthermore, online A/B testing shows that HyMiRec brings consistent improvements in real-world recommendation systems.

People's social relationships are often manifested through their surroundings, with certain objects or interactions acting as symbols for specific relationships, e.g., wedding rings, roses, hugs, or holding hands. This brings unique challenges to recognizing social relationships, requiring understanding and capturing the essence of these contexts from visual appearances. However, current methods of social relationship understanding rely on the basic classification paradigm of detected persons and objects, which fails to understand the comprehensive context and often overlooks decisive social factors, especially subtle visual cues. To highlight the social-aware context and intricate details, we propose a novel approach that recognizes contextual social relationships (ConSoRs) from a social cognitive perspective. Specifically, to incorporate social-aware semantics, we build a lightweight adapter upon the frozen contrastive language-image pretraining (CLIP) to learn social concepts via our novel multimodal side adapter tuning mechanism. Furthermore, we construct social-aware descriptive language prompts (e.g., scene, activity, objects, and emotions) with social relationships for each image, and then compel ConSoR to concentrate more intensively on the decisive visual social factors via visual-linguistic contrasting. Impressively, ConSoR outperforms previous methods with a 7.6% gain on the people-in-social-context (PISC) dataset and a 9.8% increase on the people-in-photo-album (PIPA) benchmark. Furthermore, we observe that ConSoR excels at finding critical visual evidence to reveal social relationships. Code is available at https://github.com/starsholic/ConSoR.

Retrieval-Augmented Generation (RAG) utilizes external knowledge to augment Large Language Models’ (LLMs) reliability. For flexibility, agentic RAG employs autonomous, multi-round retrieval and reasoning to resolve queries. Although recent agentic RAG has improved via reinforcement learning, they often incur substantial token overhead from search and reasoning. This tradeoff prioritizes accuracy over efficiency. To address this issue, this work proposes TeaRAG, a T oken- e fficient a gentic RAG framework capable of compressing both retrieval content and reasoning steps. (1) First, the retrieved content is compressed by augmenting chunk-based semantic retrieval with a graph retrieval using concise triplets. A knowledge association graph is then built from semantic similarity and co-occurrence. Finally, Personalized PageRank is leveraged to highlight key knowledge within this graph, reducing the number of tokens per retrieval. (2) Besides, to reduce reasoning steps, Iterative Process-aware Direct Preference Optimization (IP-DPO) is proposed. Specifically, our reward function evaluates the knowledge sufficiency by a knowledge matching mechanism, while penalizing excessive reasoning steps. This design can produce high-quality preference-pair datasets, supporting iterative DPO to improve reasoning conciseness. Across six datasets, TeaRAG improves the average Exact Match by \(4\%\) and \(2\%\) while reducing output tokens by \(61\%\) and \(59\%\) on Llama3-8B-Instruct and Qwen2.5-14B-Instruct, respectively. Code is available at https://github.com/Applied-Machine-Learning-Lab/TeaRAG .

Peiwen Yuan, Chuyi Tan, Shaoxiong Feng, Yiwei Li, Xinglin Wang, Yueqi Zhang, Jiayi Shi, Boyuan Pan, Yao Hu, Kan Li. Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2025.

Evaluating models on large benchmarks can be very resource-intensive, especially during a period of rapid model evolution. Existing efficient evaluation methods estimate the performance of target models by testing them on a small, static coreset derived from the publicly available evaluation results of source models, which are separate from the target models. However, these approaches rely on the assumption that target models have high prediction consistency with source models, which doesn’t generalize well in practice. To fill this gap, we propose TailoredBench, a method that conducts customized evaluation tailored to each target model. Specifically, a Global-coreset is first constructed as a probe to identify the most consistent source models for each target model with an adaptive source model selection strategy. Afterwards, a scalable K-Medoids clustering algorithm is proposed to extend the Global-coreset to a tailored Native-coreset for each target model. According to the predictions on respective Native-coreset, we estimate the overall performance of target models with a calibrated estimation strategy. Comprehensive experiments on five benchmarks across over 300 models demonstrate that compared to best performing baselines, TailoredBench achieves an average reduction of 31.4% in MAE of accuracy estimates under the same inference budgets, showcasing strong effectiveness and generalizability.

Language model (LM) distillation aims at distilling the knowledge in a large teacher LM to a small student one. As a critical issue facing LM distillation, a superior student often arises from a teacher of a relatively small scale instead of a larger one, especially in the presence of substantial capacity gap between the teacher and student. This issue, often referred to as the curse of capacity gap, suggests that there is likely an optimal teacher yielding the best-performing student along the scaling course of the teacher. Consequently, distillation trials on teachers of a wide range of scales are called for to determine the optimal teacher, which becomes computationally intensive in the context of large LMs (LLMs). This paper addresses this critical bottleneck by providing the law of capacity gap inducted from a preliminary study on distilling a broad range of small-scale (<3B) LMs, where the optimal teacher consistently scales linearly with the student scale across different model and data scales. By extending the law to LLM distillation on a larger scale (7B), we succeed in obtaining versatile LLMs that outperform a wide array of competitors.

Large Language models (LLMs) have become a research hotspot. To accelerate the inference of LLMs, storing computed caches in memory has become the standard technique. However, as the inference length increases, growing KV caches might lead to out-of-memory issues. Many existing methods address this issue through KV cache compression, primarily by preserving key tokens throughout all layers to reduce information loss. Most of them allocate a uniform budget size for each layer to retain. However, we observe that the minimum budget sizes needed to retain essential information vary across layers and models based on the perspectives of attention and hidden state output. Building on this observation, this paper proposes a simple yet effective KV cache compression method that leverages layer uncertainty to allocate budget size for each layer. Experimental results show that the proposed method can reduce memory usage of the KV caches to only ~20% when compared to full KV inference while achieving nearly lossless performance.

Enabling LLMs to handle lengthy context is currently a research hotspot. Most LLMs are built upon rotary position embedding (RoPE), a popular position encoding method. Therefore, a prominent path is to extrapolate the RoPE trained on comparably short texts to far longer texts. A heavy bunch of efforts have been dedicated to boosting the extrapolation via extending the formulations of the RoPE, however, few of them have attempted to showcase their inner workings comprehensively. In this paper, we are driven to offer a straightforward yet in-depth understanding of RoPE extensions from an attention perspective and on two benchmarking tasks. A broad array of experiments reveals several valuable findings: 1) Maintaining attention patterns to those at the pretrained length improves extrapolation; 2) Large attention uncertainty leads to retrieval errors; 3) Using longer continual pretraining lengths for RoPE extensions could reduce attention uncertainty and significantly enhance extrapolation.

With the rapid advancement of multimodal information retrieval, increasingly complex retrieval tasks have emerged. Existing methods predominately rely on task-specific fine-tuning of vision-language models, often those trained with image-text contrastive learning. In this paper, we explore the possibility of re-purposing generative Large Multimodal Models (LMMs) for retrieval. This approach enables unifying all retrieval tasks under the same formulation and, more importantly, allows for extrapolation towards unseen retrieval tasks without additional training. Our contributions can be summarised in the following aspects: (i) We introduce LamRA, a versatile framework designed to empower LMMs with sophisticated retrieval and reranking capabilities. (ii) For retrieval, we adopt a two-stage training strategy comprising language-only pre-training and multimodal instruction tuning to progressively enhance LMM’s retrieval performance. (iii) For reranking, we employ joint training of both pointwise and listwise reranking, offering two distinct ways to further boost the retrieval performance. (iv) Extensive experiments underscore the efficacy of our method in handling more than ten retrieval tasks, demonstrating robust performance in both supervised and zero-shot settings, including scenarios involving previously unseen retrieval tasks. Project page: https://code-kunkun.github.io/LamRA/.

Existing retrieval-augmented code generation (RACG) methods typically use an external retrieval module to fetch semantically similar code snippets used for generating subsequent fragments.However, even for consecutive code fragments, the content often diverges due to logical progression, resulting in a content gap.This gap undermines the performance of current RACG methods, as external retrieval modules based on content matching fail to infer the specific information need of LLMs to generate the next code fragment.Therefore, we propose SelfRACG, a novel paradigm that enables large language models (LLMs) to Self-express their information needs to enhance RACG.Specifically, SelfRACG includes an information need expression module and a two-stage information need-guided training strategy, which encourages LLMs to express their information need.Extensive experiments demonstrate that SelfRACG can retrieve external knowledge that better aligns with the LLM's own information needs, resulting in superior generation performance compared to vanilla RACG.Moreover, both the training and deployment costs for retrieval in our framework are much lower than those of the strongest retrieval model. 1

Video Large Language Models (VideoLLMs) have recently demonstrated remarkable progress in general video understanding. However, existing models primarily focus on high-level comprehension and are limited to text-only responses, restricting the flexibility for object-centric, multiround interactions. In this paper, we make three contributions: (i) we address these limitations by introducing a VideoLLM model, capable of performing both object referring for input and grounding for output in video reasoning tasks, i.e., allowing users to interact with videos using both textual and visual prompts; (ii) we propose STOM (Spatial-Temporal Overlay Module), a novel approach that propagates arbitrary visual prompts input at any single timestamp to the remaining frames within a video; (iii) we present VideoInfer, a manually curated object-centric video instruction dataset featuring questionanswering pairs that require reasoning. We conduct comprehensive experiments on VideoInfer and other existing benchmarks across video question answering and referring object segmentation. The results on 12 benchmarks of 6 tasks show that our proposed model consistently outperforms baselines in both video question answering and segmentation, underscoring its robustness in multimodal, object-centric video and image understanding. Project page: https://qirui-chen.github.io/RGA3-release/.

Test-time prompt tuning enhances zero-shot generalization of vision-language models but tends to ignore the relatedness among test samples during inference. Online test-time prompt tuning provides a simple way to leverage the information in previous test samples, albeit with the risk of prompt collapse due to error accumulation. To enhance test-time prompt tuning, we propose DynaPrompt, short for dynamic test-time prompt tuning, exploiting relevant data distribution information while reducing error accumulation. Built on an online prompt buffer, DynaPrompt adaptively selects and optimizes the relevant prompts for each test sample during tuning. Specifically, we introduce a dynamic prompt selection strategy based on two metrics: prediction entropy and probability difference. For unseen test data information, we develop dynamic prompt appending, which allows the buffer to append new prompts and delete the inactive ones. By doing so, the prompts are optimized to exploit beneficial information on specific test data, while alleviating error accumulation. Experiments on fourteen datasets demonstrate the effectiveness of dynamic test-time prompt tuning.

With the rapid development of generative models, discerning AI-generated content has evoked increasing attention from both industry and academia. In this paper, we conduct a sanity check on whether the task of AI-generated image detection has been solved. To start with, we present Chameleon dataset, consisting of AI-generated images that are genuinely challenging for human perception. To quantify the generalization of existing methods, we evaluate 9 off-the-shelf AI-generated image detectors on Chameleon dataset. Upon analysis, almost all models misclassify AI-generated images as real ones. Later, we propose AIDE AI-generated Image DEtector with Hybrid Features, which leverages multiple experts to simultaneously extract visual artifacts and noise patterns. Specifically, to capture the high-level semantics, we utilize CLIP to compute the visual embedding. This effectively enables the model to discern AI-generated images based on semantics and contextual information. Secondly, we select the highest and lowest frequency patches in the image, and compute the low-level patchwise features, aiming to detect AI-generated images by low-level artifacts, for example, noise patterns, anti-aliasing effects. While evaluating on existing benchmarks, for example, AIGCDetectBenchmark and GenImage, AIDE achieves +3.5% and +4.6% improvements to state-of-the-art methods, and on our proposed challenging Chameleon benchmarks, it also achieves promising results, despite the problem of detecting AI-generated images remains far from being solved.

Human preference plays a significant role in measuring large language models and guiding them to align with human values. Unfortunately, current comparing-based evaluation (CBE) methods typically focus on a single optimization objective, failing to effectively utilize scarce yet valuable preference signals. To address this, we delve into key factors that can enhance the accuracy, convergence, and scalability of CBE: suppressing sampling bias, balancing descending process of uncertainty, and mitigating updating uncertainty. Following the derived guidelines, we propose UniCBE, a unified uniformity-driven CBE framework which simultaneously optimize these core objectives by constructing and integrating three decoupled sampling probability matrices, each designed to ensure uniformity in specific aspects. We further ablate the optimal tuple sampling and preference aggregation strategies to achieve efficient CBE. On the AlpacaEval benchmark, UniCBE saves over 17% of evaluation budgets while achieving a Pearson correlation with ground truth exceeding 0.995, demonstrating excellent accuracy and convergence. In scenarios where new models are continuously introduced, UniCBE can even save over 50% of evaluation costs, highlighting its improved scalability.

With the rapid advancement of Social Networking Services (SNS), the need for intelligent and efficient interaction within diverse platforms has become more crucial. Large Language Models (LLMs) play an important role in SNS as they possess the potential to revolutionize user experience, content generation, and communication dynamics. However, recent studies focus on isolated SNS tasks rather than a comprehensive evaluation. In this paper, we introduce SNS-Bench, specially constructed for assessing the abilities of large language models from different Social Networking Services, with a wide range of SNS-related information. SNS-Bench encompasses 8 different tasks such as note classification, query content relevance, and highlight words generation in comments. Finally, 6,658 questions of social media text, including subjective questions, single-choice, and multiple-choice questions, are concluded in SNS-Bench. Further, we evaluate the performance of over 25+ current diverse LLMs on our SNS-Bench. Models with different sizes exhibit performance variations, yet adhere to the scaling law. Moreover, we hope provide more insights to revolutionize the techniques of social network services with LLMs.

With recent generative models facilitating photo-realistic image synthesis, the proliferation of synthetic images has also engendered certain negative impacts on social platforms, thereby raising an urgent imperative to develop effective detectors. Current synthetic image detection (SID) pipelines are primarily dedicated to crafting universal artifact features, accompanied by an oversight about SID training paradigm. In this paper, we re-examine the SID problem and identify two prevalent biases in current training paradigms, i.e., weakened artifact features and overfitted artifact features. Meanwhile, we discover that the imaging mechanism of synthetic images contributes to heightened local correlations among pixels, suggesting that detectors should be equipped with local awareness. In this light, we propose SAFE, a lightweight and effective detector with three simple image transformations. Firstly, for weakened artifact features, we substitute the down-sampling operator with the crop operator in image pre-processing to help circumvent artifact distortion. Secondly, for overfitted artifact features, we include ColorJitter and RandomRotation as additional data augmentations, to help alleviate irrelevant biases from color discrepancies and semantic differences in limited training samples. Thirdly, for local awareness, we propose a patch-based random masking strategy tailored for SID, forcing the detector to focus on local regions at training. Comparative experiments are conducted on an open-world dataset, comprising synthetic images generated by 26 distinct generative models. Our pipeline achieves a new state-of-the-art performance, with remarkable improvements of 4.5% in accuracy and 2.9% in average precision against existing methods. Our code is available at: https://github.com/Ouxiang-Li/SAFE.

Trajectory distillation based on consistency models (CMs) provides an effective framework for accelerating diffusion models by reducing inference steps. However, we find that existing CMs degrade style similarity and compromise aesthetic quality in stylization tasks-especially when handling image-to-image or video-to-video transformations that start denoising from partially noised inputs. The core limitation stems from existing methods enforcing initial-step alignment between the probability flow ODE (PF-ODE) trajectories of student models and their imperfect teacher models. This partial alignment strategy inevitably fails to guarantee full trajectory consistency, thereby compromising the overall generation quality. To address this issue, we propose Single Trajectory Distillation (STD), a training framework initiated from partial noise states. To counteract the additional time overhead introduced by STD, we design a trajectory bank that pre-stores intermediate states of the teacher model's PF-ODE trajectories, effectively offsetting the computational cost during student model training. This mechanism ensures STD maintains equivalent training efficiency compared to conventional consistency models. Furthermore, we incorporate an asymmetric adversarial loss to explicitly enhance style consistency and perceptual quality in generated outputs. Extensive experiments on image and video stylization demonstrate that our method surpasses existing acceleration models in terms of style similarity and aesthetic evaluations. Our code and results are available on the project page: https://single-trajectory-distillation.github.io/.

Xinglin Wang, Peiwen Yuan, Shaoxiong Feng, Yiwei Li, Boyuan Pan, Heda Wang, Yao Hu, Kan Li. Proceedings of the 2025 Conference of the Nations of the Americas Chapter of the Association for Computational Linguistics: Human Language Technologies (Volume 1: Long Papers). 2025.

Chen Zhang, Meizhi Zhong, Qimeng Wang, Xuantao Lu, Zheyu Ye, Chengqiang Lu, Yan Gao, Yao Hu, Kehai Chen, Min Zhang, Dawei Song. Proceedings of the 2025 Conference of the Nations of the Americas Chapter of the Association for Computational Linguistics: Human Language Technologies (Volume 1: Long Papers). 2025.

With the exponential growth of data, traditional object detection methods are increasingly struggling to handle vast vocabulary object detection tasks effectively. We analyze two key limitations of classification-based detectors: positive gradient dilution, where rare positive categories receive insufficient learning signals, and hard negative gradient dilution, where discriminative gradients are overwhelmed by numerous easy negatives. To address these challenges, we propose CQ-DINO, a category query-based object detection framework that reformulates classification as a contrastive task between object queries and learnable category queries. Our method introduces image-guided query selection, which reduces the negative space by adaptively retrieving top-K relevant categories per image via cross-attention, thereby rebalancing gradient distributions and facilitating implicit hard example mining. Furthermore, CQ-DINO flexibly integrates explicit hierarchical category relationships in structured datasets (e.g., V3Det) or learns implicit category correlations via self-attention in generic datasets (e.g., COCO). Experiments demonstrate that CQ-DINO achieves superior performance on the challenging V3Det benchmark (surpassing previous methods by 2.1% AP) while maintaining competitiveness in COCO. Our work provides a scalable solution for real-world detection systems requiring wide category coverage.

Test-Time Scaling (TTS) improves the performance of Large Language Models (LLMs) by using additional inference-time computation to explore multiple reasoning paths through search. Yet how to allocate a fixed rollout budget most effectively during search remains underexplored, often resulting in inefficient use of compute at test time. To bridge this gap, we formulate test-time search as a resource allocation problem and derive the optimal allocation strategy that maximizes the probability of obtaining a correct solution under a fixed rollout budget. Within this formulation, we reveal a core limitation of existing search methods: solution-level allocation tends to favor reasoning directions with more candidates, leading to theoretically suboptimal and inefficient use of compute. To address this, we propose Direction-Oriented Resource Allocation (DORA), a provably optimal method that mitigates this bias by decoupling direction quality from candidate count and allocating resources at the direction level. To demonstrate DORA’s effectiveness, we conduct extensive experiments on challenging mathematical reasoning benchmarks including MATH500, AIME2024, and AIME2025. The empirical results show that DORA consistently outperforms strong baselines with comparable computational cost, achieving state-of-the-art accuracy. We hope our findings contribute to a broader understanding of optimal TTS for LLMs.

Diffusion models have demonstrated remarkable capabilities in generating high-quality images. Recent advancements in Layout-to-Image (L2I) generation have leveraged positional conditions and textual descriptions to facilitate precise and controllable image synthesis. Despite overall progress, current L2I methods still exhibit suboptimal performance. Therefore, we propose InstanceAssemble, a novel architecture that incorporates layout conditions via instance-assembling attention, enabling position control with bounding boxes (bbox) and multimodal content control including texts and additional visual content. Our method achieves flexible adaption to existing DiT-based T2I models through light-weighted LoRA modules. Additionally, we propose a Layout-to-Image benchmark, Denselayout, a comprehensive benchmark for layout-to-image generation, containing 5k images with 90k instances in total. We further introduce Layout Grounding Score (LGS), an interpretable evaluation metric to more precisely assess the accuracy of L2I generation. Experiments demonstrate that our InstanceAssemble method achieves state-of-the-art performance under complex layout conditions, while exhibiting strong compatibility with diverse style LoRA modules. The code and pretrained models are publicly available at \url{https://github.com/FireRedTeam/InstanceAssemble}.

Unlike reasoning, which often entails a deep sequence of deductive steps, complex real-world planning is characterized by the need to synthesize a broad spectrum of parallel and potentially conflicting information and constraints. For example, in travel planning scenarios, it requires the integration of diverse real-world information and user preferences. While LLMs show promise, existing methods with long-horizon thinking struggle with handling multifaceted constraints, leading to suboptimal solutions. Motivated by the challenges of real-world travel planning, this paper introduces the Multiple Aspects of Planning (MAoP), empowering LLMs with "wide-horizon thinking" to solve planning problems with multifaceted constraints. Instead of direct planning, MAoP leverages the strategist to conduct pre-planning from various aspects and provide the planning blueprint for planners, enabling strong inference-time scalability by scaling aspects to consider various constraints. In addition, existing benchmarks for multi-constraint planning are flawed because they assess constraints in isolation, ignoring causal dependencies within the constraints, e.g, travel planning, where past activities dictate future itinerary. To address this, we propose Travel-Sim, an agent-based benchmark assessing plans via real-world simulation, thereby inherently resolving these causal dependencies. This paper advances LLM capabilities in complex planning and offers novel insights for evaluating sophisticated scenarios through simulation.

LLM-as-Benchmark-Generator methods have been widely studied as a supplement to human annotators for scalable evaluation, while the potential biases within this paradigm remain underexplored. In this work, we systematically define and validate the phenomenon of inflated performance in models evaluated on their self-generated benchmarks, referred to as self-bias, and attribute it to sub-biases arising from question domain, language style, and wrong labels. On this basis, we propose Silencer, a general framework that leverages the heterogeneity between multiple generators at both the sample and benchmark levels to neutralize bias and generate high-quality, self-bias-silenced benchmark. Experimental results across various settings demonstrate that Silencer can suppress self-bias to near zero, significantly improve evaluation effectiveness of the generated benchmark (with an average improvement from 0.655 to 0.833 in Pearson correlation with high-quality human-annotated benchmark), while also exhibiting strong generalizability.

In real-world scenarios, providing user queries with visually enhanced responses can considerably benefit understanding and memory, underscoring the great value of interleaved image-text generation. Despite recent progress, like the visual autoregressive model that unifies text and image processing in a single transformer architecture, generating high-quality interleaved content remains challenging. Moreover, evaluations of these interleaved sequences largely remain underexplored, with existing benchmarks often limited by unimodal metrics that inadequately assess the intricacies of combined image-text outputs. To address these issues, we present RAG-IGBench, a thorough benchmark designed specifically to evaluate the task of Interleaved Generation based on Retrieval-Augmented Generation (RAG-IG) in open-domain question answering. RAG-IG integrates multimodal large language models (MLLMs) with retrieval mechanisms, enabling the models to access external image-text information for generating coherent multimodal content. Distinct from previous datasets, RAG-IGBench draws on the latest publicly available content from social platforms and introduces innovative evaluation metrics that measure the quality of text and images, as well as their consistency. Through extensive experiments with state-of-the-art MLLMs (both open-source and proprietary) on RAG-IGBench, we provide an in-depth analysis examining the capabilities and limitations of these models. Additionally, we validate our evaluation metrics by demonstrating their high correlation with human assessments. Models fine-tuned on RAG-IGBench's training set exhibit improved performance across multiple benchmarks, confirming both the quality and practical utility of our dataset. Our benchmark is available at https://github.com/zry13/RAG-IGBench.

Human–AI conversation frequently relies on quoting earlier text—“check it with the formula I just highlighted”—yet today’s large language models (LLMs) lack an explicit mechanism for locating and exploiting such spans. We formalize the challenge as span-conditioned generation, decomposing each turn into the dialogue history, a set of token-offset quotation spans, and an intent utterance. Building on this abstraction, we introduce a quotation-centric data pipeline that automatically synthesizes task-specific dialogues, verifies answer correctness through multi-stage consistency checks, and yields both a heterogeneous training corpus and the first benchmark covering five representative scenarios. To meet the benchmark’s zero-overhead and parameter-efficiency requirements, we propose QuAda, a lightweight training-based method that attaches two bottleneck projections to every attention head, dynamically amplifying or suppressing attention to quoted spans at inference time while leaving the prompt unchanged and updating < 2.8% of backbone weights. Experiments across models show that QuAda is suitable for all scenarios and generalizes to unseen topics, offering an effective, plug-and-play solution for quotation-aware dialogue.

Accurate watch time prediction is crucial for enhancing user engagement in streaming short-video platforms, although it is challenged by complex distribution characteristics across multi-granularity levels. Through systematic analysis of real-world industrial data, we uncover two critical challenges in watch time prediction from a distribution aspect: (1) coarse-grained skewness induced by a significant concentration of quick-skips1, (2) fine-grained diversity arising from various user-video interaction patterns. Consequently, we assume that the watch time follows the Exponential-Gaussian Mixture (EGM) distribution, where the exponential and Gaussian components respectively characterize the skewness and diversity. Accordingly, an Exponential-Gaussian Mixture Network (EGMN) is proposed for the parameterization of EGM distribution, which consists of two key modules: a hidden representation encoder and a mixture parameter generator. We conducted extensive offline experiments on public datasets and online A/B tests on the industrial short-video feeding scenario of Xiaohongshu App to validate the superiority of EGMN compared with existing state-of-the-art methods. Remarkably, comprehensive experimental results have proven that EGMN exhibits excellent distribution fitting ability across coarse-to-fine-grained levels. We open source related code on Github: https://github.com/BestActionNow/EGMN.

User-generated content (UGC) communities, especially those featuring multimodal content, improve user experiences by integrating visual and textual information into results (or items).The challenge of improving user experiences in complex systems with search and recommendation (S&R) services has drawn significant attention from both academia and industry these years.However, the lack of high-quality datasets has limited the research progress on multimodal S&R.To address the growing need for developing better S&R services, we present a novel multimodal information retrieval dataset in this paper, namely Qilin.The dataset is collected from Xiaohongshu, a popular social platform with over 300 million monthly active users and an average search penetration rate of over 70%.In contrast to existing datasets, Qilin offers a comprehensive collection of user sessions with heterogeneous results like image-text notes, video notes, commercial notes, and direct answers, facilitating the development of advanced multimodal neural retrieval models across diverse task settings.To better model user satisfaction and support the analysis of heterogeneous user behaviors, we also collect extensive APP-level contextual signals and genuine user feedback.Notably, Qilin contains user-favored answers and their referred results for search requests triggering the

Social chatbots have become essential companions in daily scenarios ranging from emotional support to personal interaction. However, conventional chatbots with passive response mechanisms usually rely on users to initiate or sustain dialogues by bringing up new topics, resulting in diminished engagement and shortened dialogue duration. In this paper, we present PaRT, a novel framework enabling context-aware proactive dialogues for social chatbots through personalized real-time retrieval and generation. Specifically, PaRT first integrates user profiles and dialogue context into a large language model (LLM), which is initially prompted to refine user queries and recognize underlying intents for the upcoming conversation. Guided by refined intents, the LLM generates personalized dialogue topics as targeted queries to retrieve relevant passages from RedNote. Finally, we prompt LLMs with summarized passages to generate knowledge-grounded and engagement-optimized responses. Our approach has been running stably in a real-world production environment for more than 30 days, achieving a 21.77% improvement in the average duration of dialogues.

Recently, audiences have increasingly watched diverse dramas and movies on streaming platforms, prompting platform administrators to enhance their understanding of video semantics. In particular, capturing and interpreting social relations among characters is critical for content-driven intelligent services and enhancing user experience. However, most existing research has solely approached social relations recognition as a classification problem, without justifiable interpretation of prediction results within the video context. To address this issue, we study cognitive science research and the Chain-of-Thought (CoT) strategy. Based on these foundations, we propose SaMo-CoT, an approach that leverages large language models (LLMs) in step-by-step reasoning. This approach simulates human cognitive processes by combining social scenarios in videos and incorporating empirical social interaction knowledge. In this way, we enable verbal rationales for determining social relations in video understanding. Furthermore, we present an innovative doubly-right social relation recognition framework that predicts both correct social relation labels and correct scenario rationales. Specifically, we translate verbal CoT into multimodal CoT by leveraging scenario-aware prompts and contrastive learning. Extensive experiments demonstrate significant improvements in classification accuracy and interpretability compared to traditional approaches.

Fine-grained visual categorization (FGVC) relies on hierarchical features extracted by deep convolutional neural networks (CNNs) to recognize closely alike objects. Particularly, shallow layer features containing rich spatial details are vital for specifying subtle differences between objects but are usually inadequately optimized due to gradient vanishing during backpropagation. In this article, hierarchical self-distillation (HSD) is introduced to generate well-optimized CNNs features for accurate fine-grained categorization. HSD inherits from the widely applied deep supervision and implements multiple intermediate losses for reinforced gradients. Besides that, we observe that the hard (one-hot) labels adopted for intermediate supervision hurt the performance of FGVC by enforcing overstrict supervision. As a solution, HSD seeks self-distillation where soft predictions generated by deeper layers of the network are hierarchically exploited to supervise shallow parts. Moreover, self-information entropy loss (SIELoss) is designed in HSD to adaptively soften intermediate predictions and facilitate better convergence. In addition, the gradient detached fusion (GDF) module is incorporated to produce an ensemble result with multiscale features via effective feature fusion. Extensive experiments on four challenging fine-grained datasets show that, with neglectable parameter increase, the proposed HSD framework and the GDF module both bring significant performance gains over different backbones, which also achieves state-of-the-art classification performance.

DEtection TRansformer (DETR)-based models have achieved remarkable performance. However, they are accompanied by a large computation overhead cost, which significantly prevents their applications on resource-limited devices. Prior arts attempt to reduce the computational burden of DETR using low-bit quantization, while these methods sacrifice a severe significant performance on weight-activation-attention low-bit quantization. We observe that the number of matching queries and positive samples affect much on the representation capacity of queries in DETR, while quantifying queries of DETR further reduces its representational capacity, thus leading to a severe performance drop. We introduce a new quantization strategy based on Auxiliary Queries for DETR (AQ-DETR), aiming to enhance the capacity of quantized queries. In addition, a layer-by-layer distillation is proposed to reduce the quantization error between quantized attention and full-precision counterpart. Through our extensive experiments on large-scale open datasets, the performance of the 4-bit quantization of DETR and Deformable DETR models is comparable to full-precision counterparts.

Brain signal visualization has emerged as an active research area, serving as a critical interface between the human visual system and computer vision models. Diffusion-based methods have recently shown promise in analyzing functional magnetic resonance imaging (fMRI) data, including the reconstruction of high-quality images consistent with original visual stimuli. Nonetheless, it remains a critical challenge to effectively harness the semantic and silhouette information extracted from brain signals. In this paper, we propose a novel approach, termed as Controllable Mind Visual Diffusion Model (CMVDM). Specifically, CMVDM first extracts semantic and silhouette information from fMRI data using attribute alignment and assistant networks. Then, a control model is introduced in conjunction with a residual block to fully exploit the extracted information for image synthesis, generating high-quality images that closely resemble the original visual stimuli in both semantic content and silhouette characteristics. Through extensive experimentation, we demonstrate that CMVDM outperforms existing state-of-the-art methods both qualitatively and quantitatively. Our code is available at https://github.com/zengbohan0217/CMVDM.

Xinglin Wang, Yiwei Li, Shaoxiong Feng, Peiwen Yuan, Boyuan Pan, Heda Wang, Yao Hu, Kan Li. Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2024.

Peiwen Yuan, Shaoxiong Feng, Yiwei Li, Xinglin Wang, Boyuan Pan, Heda Wang, Yao Hu, Kan Li. Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2024.

Recent advances in vision-language models like Stable Diffusion have shown remarkable power in creative image synthesis and editing. However, most existing text-to-image editing methods encounter two obstacles: First, the text prompt needs to be carefully crafted to achieve good results, which is not intuitive or user-friendly. Second, they are in-sensitive to local edits and can irreversibly affect non-edited regions, leaving obvious editing traces. To tackle these problems, we propose a Zero-shot instructiON-guided local image Editing approach, termed ZONE. We first convert the editing intent from the user-provided instruction (e.g., “make his tie blue”) into specific image editing regions through InstructPix2Pix. We then propose a Region-loll scheme for precise image layer extraction from an off-the-shelf segment model. We further develop an edge smoother based on FFT for seamless blending between the layer and the image. Our method allows for arbitrary manipulation of a specific region with a single instruction while preserving the rest. Extensive experiments demonstrate that our Z ONE achieves remarkable local editing results and user-friendliness, outperforming state-of-the-art methods. Code is available at https://github.com/ls1001006/ZONE.

Recent advancements in subject-driven image generation have led to zero-shot generation, yet precise selection and focus on crucial subject representations remain challenging. Addressing this, we introduce the SSR-Encoder, a novel architecture designed for selectively capturing any subject from single or multiple reference images. It responds to various query modalities including text and masks, without necessitating test-time fine-tuning. The SSR-Encoder combines a Token-to-Patch Aligner that aligns query inputs with image patches and a Detail-Preserving Subject Encoder for extracting and preserving fine features of the subjects, thereby generating subject embeddings. These embeddings, used in conjunction with original text embeddings, condition the generation process. Characterized by its model generalizability and efficiency, the SSR-Encoder adapts to a range of custom models and control modules. Enhanced by the Embedding Consistency Regularization Loss for improved training, our extensive experiments demonstrate its effectiveness in versatile and high-quality image generation, indicating its broad applicability. Project page: ssr-encoder.github.io

Peiwen Yuan, Shaoxiong Feng, Yiwei Li, Xinglin Wang, Yueqi Zhang, Chuyi Tan, Boyuan Pan, Heda Wang, Yao Hu, Kan Li. Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing. 2024.

Deep Recommender Systems (DRS) are essential for navigating the extensive data across various platforms in today's digital landscape. Current DRS models often treat all features equally and implement complex structures to enhance the capture of feature interactions. However, they may fail to recognize crucial user patterns due to not fully utilizing user-specific features for user modeling. Moreover, prevailing user modeling techniques concentrate exclusively on either the group or individual level, overlooking the potential insights from the unaddressed one. This oversight can miss shared group preferences or learn group patterns that conflict with individual preferences. To overcome these limitations, we introduce GPRec, a novel bi-level user modeling approach that substantially improves DRS. GPRec explicitly categorizes users into groups in a learnable manner and aligns them with corresponding group embeddings. We design the dual group embedding space to offer a diverse perspective on group preferences by contrasting positive and negative patterns. On the individual level, GPRec identifies personal preferences from ID-like features and refines the obtained individual representations to be independent of group ones, thereby providing a robust complement to the group-level modeling. We also present various strategies for the flexible integration of GPRec into various DRS models. Rigorous testing of GPRec on three public datasets has demonstrated significant improvements in recommendation quality. Additional experiments further explore crucial components of GPRec, its parameter sensitivity, and the group diversity. The implementation code is readily available online to facilitate future research and practical deployment: https://github.com/Applied-Machine-Learning-Lab/GPRec.

Video action segmentation aims to identify and localize actions. Existing models have achieved impressive performance with pre-extracted frame-level features, but this may limit zero-shot learning and cross-dataset inference, especially for new actions or scenes. To overcome this problem, we propose a novel end-to-end network designed for robust performance across both familiar and novel action segmentation scenarios. Our approach combines a plug-and-play visual prompt module enhancing CLIP features’ temporal understanding, and a learnable text prompt that enriches label semantics and refines the model’s focus, significantly boosting performance. Our results demonstrate that CLIP features can assist in action segmentation tasks, and prompts can improve task effectiveness. Furthermore, our findings show that CLIP features contain information that i3d features do not. We evaluate the proposed method on several video datasets, including Georgia Tech Egocentric Activities (GTEA), 50Salads, and Breakfast, and the results show that the proposed model outperforms existing SOTA models.

Recent years have witnessed the urgent request for multi-modal sarcasm detection in social media platforms. Though large efforts have been made with significant progress, prior arts may fail to fully integrate the commonsense knowledge, and struggle to model the incongruity within implicit meanings of multimodal cues. To address these limitations, we propose a novel Knowledge-Enhanced Multi-perspective Incongruity Perception network, named KEMIP. Specifically, we adopt generative language models to produce captions and commonsense for image and text respectively for comprehensive understanding. Subsequently, to exploit the essential cues from multiple perspectives, we customize an incongruity perception module, which utilizes several fusion networks to capture both literal and implicit inconsistencies. Afterwards, an ensemble weighting gate is employed to integrate the result of individual perspective. Experiments on a public multimodal sarcasm detection benchmark demonstrate the superiority of our proposed KEMIP framework.

While traditional distributionally robust optimization (DRO) aims to minimize the maximal risk over a set of distributions, Agarwal & Zhang (2022) recently proposed a variant that replaces risk with *excess risk*. Compared to DRO, the new formulation—minimax excess risk optimization (MERO) has the advantage of suppressing the effect of heterogeneous noise in different distributions. However, the choice of excess risk leads to a very challenging minimax optimization problem, and currently there exists only an inefficient algorithm for empirical MERO. In this paper, we develop efficient stochastic approximation approaches which directly target MERO. Specifically, we leverage techniques from stochastic convex optimization to estimate the minimal risk of every distribution, and solve MERO as a stochastic convex-concave optimization (SCCO) problem with biased gradients. The presence of bias makes existing theoretical guarantees of SCCO inapplicable, and fortunately, we demonstrate that the bias, caused by the estimation error of the minimal risk, is under-control. Thus, MERO can still be optimized with a nearly optimal convergence rate. Moreover, we investigate a practical scenario where the quantity of samples drawn from each distribution may differ, and propose a stochastic approach that delivers *distribution-dependent* convergence rates.

To deal with changing environments, adaptive regret has been proposed to minimize the regret over every interval. Previous studies have established a small-loss adaptive regret bound for general convex functions under the smoothness condition, offering the advantage of being much tighter than minimax rates for benign problems. However, it remains unclear whether similar bounds are attainable for other types of convex functions, such as exp-concave and strongly convex functions. In this paper, we first propose a novel algorithm that achieves a small-loss adaptive regret bound for exp-concave and smooth function. Subsequently, to address the limitation that existing algorithms can only handle one type of convex functions, we further design a universal algorithm capable of delivering small-loss adaptive regret bounds for general convex, exp-concave, and strongly convex functions simultaneously. That is challenging because the universal algorithm follows the meta-expert framework, and we need to ensure that upper bounds for both meta-regret and expert-regret are of small-loss types. Moreover, we provide a novel analysis demonstrating that our algorithms are also equipped with minimax adaptive regret bounds when functions are non-smooth.

Instruction data is crucial for improving the capability of Large Language Models (LLMs) to align with human-level performance. Recent research LIMA demonstrates that alignment is essentially a process where the model adapts instructions' interaction style or format to solve various tasks, leveraging pre-trained knowledge and skills. Therefore, for instructional data, the most important aspect is the task it represents, rather than the specific semantics and knowledge information. The latent representations of instructions play roles for some instruction-related tasks like data selection and demonstrations retrieval. However, they are always derived from text embeddings, encompass overall semantic information that influences the representation of task categories. In this work, we introduce a new concept, instruction embedding, and construct Instruction Embedding Benchmark (IEB) for its training and evaluation. Then, we propose a baseline Prompt-based Instruction Embedding (PIE) method to make the representations more attention on tasks. The evaluation of PIE, alongside other embedding methods on IEB with two designed tasks, demonstrates its superior performance in accurately identifying task categories. Moreover, the application of instruction embeddings in four downstream tasks showcases its effectiveness and suitability for instruction-related tasks.

A well-known dilemma in large vision-language models (e.g., GPT-4, LLaVA) is that while increasing the number of vision tokens generally enhances visual understanding, it also significantly raises memory and computational costs, especially in long-term, dense video frame streaming scenarios. Although learnable approaches like Q-Former and Perceiver Resampler have been developed to reduce the vision token burden, they overlook the context causally modeled by LLMs (i.e., key-value cache), potentially leading to missed visual cues when addressing user queries. In this paper, we introduce a novel approach to reduce vision compute by leveraging redundant vision tokens ``skipping layers'' rather than decreasing the number of vision tokens. Our method, VideoLLM-MoD, is inspired by mixture-of-depths LLMs and addresses the challenge of numerous vision tokens in long-term or streaming video. Specifically, for certain transformer layer, we learn to skip the computation for a high proportion (e.g., 80\%) of vision tokens, passing them directly to the next layer. This approach significantly enhances model efficiency, achieving approximately 42% time and 30% memory savings for the entire training. Moreover, our method reduces the computation in the context and avoid decreasing the vision tokens, thus preserving or even improving performance compared to the vanilla model. We conduct extensive experiments to demonstrate the effectiveness of VideoLLM-MoD, showing its state-of-the-art results on multiple benchmarks, including narration, forecasting, and summarization tasks in COIN, Ego4D, and Ego-Exo4D datasets. The code and checkpoints will be made available at github.com/showlab/VideoLLM-online.

Advancements in multimodal learning, particularly in video understanding and generation, require high-quality video-text datasets for improved model performance. Vript addresses this issue with a meticulously annotated corpus of 12K high-resolution videos, offering detailed, dense, and script-like captions for over 420K clips. Each clip has a caption of ~145 words, which is over 10x longer than most video-text datasets. Unlike captions only documenting static content in previous datasets, we enhance video captioning to video scripting by documenting not just the content, but also the camera operations, which include the shot types (medium shot, close-up, etc) and camera movements (panning, tilting, etc). By utilizing the Vript, we explore three training paradigms of aligning more text with the video modality rather than clip-caption pairs. This results in Vriptor, a top-performing video captioning model among open-source models, comparable to GPT-4V in performance. Vriptor is also a powerful model capable of end-to-end generation of dense and detailed captions for long videos. Moreover, we introduce Vript-Hard, a benchmark consisting of three video understanding tasks that are more challenging than existing benchmarks: Vript-HAL is the first benchmark evaluating action and object hallucinations in video LLMs, Vript-RR combines reasoning with retrieval resolving question ambiguity in long-video QAs, and Vript-ERO is a new task to evaluate the temporal understanding of events in long videos rather than actions in short videos in previous works. All code, models, and datasets are available in https://github.com/mutonix/Vript.

Multi-modal large language models (MLLMs), such as GPT-4, exhibit great comprehension capabilities on human instruction, as well as zero-shot ability on new downstream multi-modal tasks. To integrate the different modalities within a unified embedding space, previous MLLMs attempted to conduct visual instruction tuning with massive and high-quality image-text pair data, which requires substantial costs in data collection and training resources. In this article, we propose TOMGPT (Text-Only training Multi-modal GPT), a cost-effective MLLM tuned solely on easily accessible text data with much fewer resources. Along with pre-trained visual-linguistic coupled modality space (e.g., CLIP and ALIGN model), a text-only training strategy is devised to further project the aligned multi-modal latent space to that of LLM, endowing the LLM with visual comprehension capabilities in an efficient manner. Instead of enormous image-text training data required by previous MLLMs, we find that TOMGPT can be well-tuned with fewer yet diverse GPT-generated free-form text data, as we establish the semantic connection between LLM and pre-trained vision-language model. A quantitative evaluation is conducted on both MME and LVLM, which are recently released and extensively utilized MLLM benchmarks. The experiments reveal that TOMGPT achieved reliable performance compared to numerous models trained on a large amount of image-text pair data. Case studies are also presented, demonstrating TOMGPT’s broad understanding and dialogue capabilities across diverse image categories.

In this paper, we consider the problem of simultaneously detecting objects and inferring their visual attributes in an image, even for those with no manual annotations provided at the training stage, resembling an open-vocabulary scenario. To achieve this goal, we make the following contributions: (i) we start with a naive two-stage approach for open-vocabulary object detection and attribute classification, termed CLIP-Attr. The candidate objects are first proposed with an offline RPN and later classified for semantic category and attributes; (ii) we combine all available datasets and train with a federated strategy to finetune the CLIP model, aligning the visual representation with attributes, additionally, we investigate the efficacy of leveraging freely available online image-caption pairs under weakly supervised learning; (iii) in pursuit of efficiency, we train a Faster-RCNN type model end-to-end with knowledge distillation, that performs class-agnostic object proposals and classification on semantic categories and attributes with classifiers generated from a text encoder; Finally, (iv) we conduct extensive experiments on VAW, MS-COCO, LSA, and OVAD datasets, and show that recognition of semantic category and attributes is complementary for visual scene understanding, i.e., jointly training object detection and attributes prediction largely outperform existing approaches that treat the two tasks independently, demonstrating strong generalization ability to novel attributes and categories.

Video Instance Segmentation (VIS) aims at segmenting and categorizing objects in videos from a closed set of training categories, lacking the generalization ability to handle novel categories in real-world videos. To address this limitation, we make the following three contributions. First, we introduce the novel task of Open-Vocabulary Video Instance Segmentation, which aims to simultaneously segment, track, and classify objects in videos from open-set categories, including novel categories unseen during training. Second, to benchmark Open-Vocabulary VIS, we collect a Large-Vocabulary Video Instance Segmentation dataset (LV-VIS), that contains well-annotated objects from 1,196 diverse categories, significantly surpassing the category size of existing datasets by more than one order of magnitude. Third, we propose an efficient Memory-Induced Transformer architecture, OV2Seg, to first achieve Open-Vocabulary VIS in an end-to-end manner with near real-time inference speed. Extensive experiments on LV-VIS and four existing VIS datasets demonstrate the strong zero-shot generalization ability of OV2Seg on novel categories. The dataset and code are released here https://github.com/haochenheheda/LVVIS.

Abstract Can our video understanding systems perceive objects when a heavy occlusion exists in a scene? To answer this question, we collect a large-scale dataset called OVIS for occluded video instance segmentation, that is, to simultaneously detect, segment, and track instances in occluded scenes. OVIS consists of 296k high-quality instance masks from 25 semantic categories, where object occlusions usually occur. While our human vision systems can understand those occluded instances by contextual reasoning and association, our experiments suggest that current video understanding systems cannot. On the OVIS dataset, the highest AP achieved by state-of-the-art algorithms is only 16.3, which reveals that we are still at a nascent stage for understanding objects, instances, and videos in a real-world scenario. We also present a simple plug-and-play module that performs temporal feature calibration to complement missing object cues caused by occlusion. Built upon MaskTrack R-CNN and SipMask, we obtain a remarkable AP improvement on the OVIS dataset. The OVIS dataset and project code are available at http://songbai.site/ovis .

Temporal action detection (TAD) aims to determine the semantic label and the temporal interval of every action instance in an untrimmed video. It is a fundamental and challenging task in video understanding. Previous methods tackle this task with complicated pipelines. They often need to train multiple networks and involve hand-designed operations, such as non-maximal suppression and anchor generation, which limit the flexibility and prevent end-to-end learning. In this paper, we propose an end-to-end Transformer-based method for TAD, termed TadTR. Given a small set of learnable embeddings called action queries, TadTR adaptively extracts temporal context information from the video for each query and directly predicts action instances with the context. To adapt Transformer to TAD, we propose three improvements to enhance its locality awareness. The core is a temporal deformable attention module that selectively attends to a sparse set of key snippets in a video. A segment refinement mechanism and an actionness regression head are designed to refine the boundaries and confidence of the predicted instances, respectively. With such a simple pipeline, TadTR requires lower computation cost than previous detectors, while preserving remarkable performance. As a self-contained detector, it achieves state-of-the-art performance on THUMOS14 (56.7% mAP) and HACS Segments (32.09% mAP). Combined with an extra action classifier, it obtains 36.75% mAP on ActivityNet-1.3. Code is available at https://github.com/xlliu7/TadTR.

At this blooming age of social media and mobile platform, mass consumers are migrating from horizontal video to vertical contents delivered on hand-held devices. Accordingly, revitalizing the exposure of horizontal video becomes vital and urgent, which is hereby tackled by our automated horizontal-to-vertical (abbreviated as H2V) video conversion framework. Essentially, the {\it \textbf{H2V}} framework performs subject-preserving video cropping instantiated in the proposed Rank-SS module. Rank-SS incorporates object detection to discover candidate subjects, from which we select the primary subject-to-preserve leveraging location, appearance, and salient cues in a convolutional neural network. In addition to converting horizontal videos vertically by cropping around the selected subject, automatic shot detection and multi-object tracking are integrated into the {\it \textbf{H2V}} framework to accommodate long and complex videos. To develop {\it \textbf{H2V}} systems, we collect an {\it \textbf{H2V-142K}} dataset containing 125 videos (132K frames) and 9,500 cover images annotated with primary subject bounding boxes. On {\it \textbf{H2V-142K}} and public object detection datasets, our method demonstrates promising results on the subject selection comparing to the related solutions. Furthermore, our {\it \textbf{H2V}} framework is industrially deployed hosting millions of daily active users and exhibits favorable H2V conversion performance. By making this dataset as well as our approach publicly available, we wish to pave the way for more horizontal-to-vertical video conversion research. Our collected H2V-142K dataset is available atH2V-142K website.

Image-to-video adaptation leverages off-the-shelf learned models in labeled images to help classification in unlabeled videos, thus alleviating the high computation overhead of training a video classifier from scratch. This task is very challenging since there exist two types of domain shifts between images and videos: 1) spatial domain shift caused by static appearance variance between images and video frames, and 2) temporal domain shift caused by the absence of dynamic motion in images. Moreover, for different video classes, these two domain shifts have different effects on the domain gap and should not be treated equally during adaptation. In this paper, we propose a spatial-temporal causal inference framework for image-to-video adaptation. We first construct a spatial-temporal causal graph to infer the effects of the spatial and temporal domain shifts by performing counterfactual causality. We then learn causality-guided bidirectional heterogeneous mappings between images and videos to adaptively reduce the two domain shifts. Moreover, to relax the assumption that the label spaces of the image and video domains are the same by the existing methods, we incorporate class-wise alignment into the learning of image-video mappings to perform partial image-to-video adaptation where the image label space subsumes the video label space. Extensive experiments on several video datasets have validated the effectiveness of our proposed method.

We study a variant of stochastic bandits where the feedback model is specified by a graph. In this setting, after playing an arm, one can observe rewards of not only the played arm but also other arms that are adjacent to the played arm in the graph. Most of the existing work assumes the reward distributions are stationary over time, which, however, is often violated in common scenarios such as recommendation systems and online advertising. To address this limitation, we study stochastic bandits with graph feedback in non-stationary environments and propose algorithms with graph-dependent dynamic regret bounds. When the number of reward distribution changes L is known in advance, one of our algorithms achieves an Õ(√(αLT)) dynamic regret bound. We also develop an adaptive algorithm that can adapt to unknown L and attain an Õ(√(θLT)) dynamic regret. Here, α and θ are some graph-dependent quantities and T is the time horizon.

Current developments in temporal event or action localization usually target actions captured by a single camera. However, extensive events or actions in the wild may be captured as a sequence of shots by multiple cameras at different positions. In this paper, we propose a new and challenging task called multi-shot temporal event localization, and accordingly, collect a large-scale dataset called MUlti-Shot EventS (MUSES). MUSES has 31,477 event instances for a total of 716 video hours. The core nature of MUSES is the frequent shot cuts, for an average of 19 shots per instance and 176 shots per video, which induces large intra-instance variations. Our comprehensive evaluations show that the state-of-the-art method in temporal action localization only achieves an mAP of 13.1% at IoU=0.5. As a minor contribution, we present a simple baseline approach for handling the intra-instance variations, which reports an mAP of 18.9% on MUSES and 56.9% on THUMOS14 at IoU=0.5. To facilitate research in this direction, we release the dataset and the project code at https://songbai.site/muses/.

In this work we present SwiftNet for real-time semisupervised video object segmentation (one-shot VOS), which reports 77.8% $\mathcal{J}\& \mathcal{F}$ and 70 FPS on DAVIS 2017 validation dataset, leading all present solutions in overall accuracy and speed performance. We achieve this by elaborately compressing spatiotemporal redundancy in matching-based VOS via Pixel-Adaptive Memory (PAM). Temporally, PAM adaptively triggers memory updates on frames where objects display noteworthy inter-frame variations. Spatially, PAM selectively performs memory update and match on dynamic pixels while ignoring the static ones, significantly reducing redundant computations wasted on segmentation-irrelevant pixels. To promote efficient reference encoding, light-aggregation encoder is also introduced in SwiftNet deploying reversed sub-pixel. We hope SwiftNet could set a strong and efficient baseline for real-time VOS and facilitate its application in mobile vision. The source code of SwiftNet can be found at https://github.com/haochenheheda/SwiftNet.

Instance segmentation can detect where the objects are in an image, but hard to understand the relationship between them. We pay attention to a typical relationship, relative saliency. A closely related task, salient object detection, predicts a binary map highlighting a visually salient region while hard to distinguish multiple objects. Directly combining two tasks by post-processing also leads to poor performance. There is a lack of research on relative saliency at present, limiting the practical applications such as content-aware image cropping, video summary, and image labeling.In this paper, we study the Salient Object Ranking (SOR) task, which manages to assign a ranking order of each detected object according to its visual saliency. We propose the first end-to-end framework of the SOR task and solve it in a multi-task learning fashion. The framework handles instance segmentation and salient object ranking simultaneously. In this framework, the SOR branch is independent and flexible to cooperate with different detection methods, so that easy to use as a plugin. We also intro-duce a Position-Preserved Attention (PPA) module tailored for the SOR branch. It consists of the position embedding stage and feature interaction stage. Considering the importance of position in saliency comparison, we preserve absolute coordinates of objects in ROI pooling operation and then fuse positional information with semantic features in the first stage. In the feature interaction stage, we apply the attention mechanism to obtain proposals’ contextualized representations to predict their relative ranking orders. Extensive experiments have been conducted on the ASR dataset. Without bells and whistles, our proposed method outperforms the former state-of-the-art method significantly. The code will be released publicly available on https://github.com/EricFH/SOR.

Cross domain recommender system constitutes a powerful method to tackle the cold-start and sparsity problem by aggregating and transferring user preferences across multiple category domains. Therefore, it has great potential to improve click-through-rate prediction performance in online commerce platforms having many domains of products. While several cross domain sequential recommendation models have been proposed to leverage information from a source domain to improve CTR predictions in a target domain, they did not take into account bidirectional latent relations of user preferences across source-target domain pairs. As such, they cannot provide enhanced cross-domain CTR predictions for both domains simultaneously. In this paper, we propose a novel approach to cross-domain sequential recommendations based on the dual learning mechanism that simultaneously transfers information between two related domains in an iterative manner until the learning process stabilizes. In particular, the proposed Dual Attentive Sequential Learning (DASL) model consists of two novel components Dual Embedding and Dual Attention, which jointly establish the two-stage learning process: we first construct dual latent embeddings that extract user preferences in both domains simultaneously, and subsequently provide cross-domain recommendations by matching the extracted latent embeddings with candidate items through dual-attention learning mechanism. We conduct extensive offline experiments on three real-world datasets to demonstrate the superiority of our proposed model, which significantly and consistently outperforms several state-of-the-art baselines across all experimental settings. We also conduct an online A/B test at a major video streaming platform Alibaba-Youku, where our proposed model significantly improves business performance over the latest production system in the company.

In this paper, we propose a new task of deep interactive video inpainting and an application for users to interact with machines. To our best knowledge, this is the first deep learning-based interactive video inpainting framework that only uses a free form of user input as guidance (i.e. scribbles) instead of mask annotations, which has academic, entertainment, and commercial value.

Non-maximum suppression (NMS) is widely used in object detection pipelines for removing duplicated bounding boxes. The inconsistency between the confidence for NMS and the real localization confidence seriously affects detection performance. Prior works propose to predict Intersection-over-Union (IoU) between bounding boxes and corresponding ground-truths to improve NMS, while accurately predicting IoU is still a challenging problem. We argue that the complex definition of IoU and feature misalignment make it difficult to predict IoU accurately. In this paper, we propose a novel Decoupled IoU Regression (DIR) model to handle these problems. The proposed DIR decouples the traditional localization confidence metric IoU into two new metrics, Purity and Integrity. Purity reflects the proportion of the object area in the detected bounding box, and Integrity refers to the completeness of the detected object area. Separately predicting Purity and Integrity can divide the complex mapping between the bounding box and its IoU into two clearer mappings and model them independently. In addition, a simple but effective feature realignment approach is also introduced to make the IoU regressor work in a hindsight manner, which can make the target mapping more stable. The proposed DIR can be conveniently integrated with existing two-stage detectors and significantly improve their performance. Through a simple implementation of DIR with HTC, we obtain 51.3% AP on MS COCO benchmark, which outperforms previous methods and achieves state-of-the-art.

Recent years have witnessed the booming of online video platforms. Along this line, a graph to illustrate social relation among characters has been long expected to not only benefit the audiences for better understanding the story, but also support the fine-grained video analysis task in a semantic way. Unfortunately, though we humans could easily infer the social relations among characters, it is still an extremely challenging task for intelligent systems to automatically capture the social relation by absorbing multi-modal cues. Besides, they fail to describe the relations among multiple characters in a graph-generation perspective. To that end, inspired by the human inference ability on social relationship, we propose a novel Hierarchical- Cumulative Graph Convolutional Network (HC-GCN) to generate the social relation graph for multiple characters in the video. Specifically, we first integrate the short-term multi-modal cues, including visual, textual and audio information, to generate the frame-level graphs for part of characters via multimodal graph convolution technique. While dealing with the video-level aggregation task, we design an end-to-end framework to aggregate all frame-level subgraphs along the temporal trajectory, which results in a global video-level social graph with various social relationships among multiple characters. Extensive validations on two real-world large-scale datasets demonstrate the effectiveness of our proposed method compared with SOTA baselines.

Pedestrian detection is an initial step to perform outdoor scene analysis, which plays an essential role in many real-world applications. Although having enjoyed the merits of deep learning frameworks from the generic object detectors, pedestrian detection is still a very challenging task due to heavy occlusions, and highly crowded group. Generally, the conventional detectors are unable to differentiate individuals from each other effectively under such a dense environment. To tackle this critical problem, we propose an attribute-aware pedestrian detector to explicitly model people's semantic attributes in a high-level feature detection fashion. Besides the typical semantic features, center position, target's scale, and offset, we introduce a pedestrian-oriented attribute feature to encode the high-level semantic differences among the crowd. Moreover, a novel attribute-feature-based Non-Maximum Suppression (NMS) is proposed to distinguish the person from a highly overlapped group by adaptively rejecting the false-positive results in a very crowd settings. Furthermore, an enhanced ground truth target is designed to alleviate the difficulties caused by the attribute configuration, and to ease the class imbalance issue during training. Finally, we evaluate our proposed attribute-aware pedestrian detector on three benchmark datasets including CityPerson, CrowdHuman, and EuroCityPerson, and achieves the state-of-the-art results.

As a crucial task for video analysis, social relation recognition for characters not only provides semantically rich description of video content but also supports intelligent applications, e.g., video retrieval and visual question answering. Unfortunately, due to the semantic gap between visual and semantic features, traditional solutions may fail to reveal the accurate relations among characters. At the same time, the development of social media platforms has now promoted the emergence of crowdsourced comments, which may enhance the recognition task with semantic and descriptive cues. To that end, in this article, we propose a novel multimodal-based solution to deal with the character relation recognition task. Specifically, we capture the target character pairs via a search module and then design a multistream architecture for jointly embedding the visual and textual information, in which feature fusion and attention mechanism are adapted for better integrating the multimodal inputs. Finally, supervised learning is applied to classify character relations. Experiments on real-world data sets validate that our solution outperforms several competitive baselines.

In multi-view multi-label learning (MVML), each training example is represented by different feature vectors and associated with multiple labels simultaneously. Nonetheless, the labeling quality of training examples is tend to be affected by annotation noises. In this paper, the problem of multi-view partial multi-label learning (MVPML) is studied, where the set of associated labels are assumed to be candidate ones and only partially valid. To solve the MVPML problem, a two-stage graph-based disambiguation approach is proposed. Firstly, the ground-truth labels of each training example are estimated by disambiguating the candidate labels with fused similarity graph. After that, the predictive model for each label is learned from embedding features generated from disambiguation-guided clustering analysis. Extensive experimental studies clearly validate the effectiveness of the proposed approach in solving the MVPML problem.

Graph-based semi-supervised learning (GSSL) studies the problem where in addition to a set of data points with few available labels, there also exists a graph structure that describes the underlying relationship between data items. In practice, structure uncertainty often occurs in graphs when edges exist between data with different labels, which may further results in prediction uncertainty of labels. Considering that Gaussian process generalizes well with few labels and can naturally model uncertainty, in this paper, we propose an Uncertainty aware Graph Gaussian Process based approach (UaGGP) for GSSL. UaGGP exploits the prediction uncertainty and label smooth regularization to guide each other during learning. To further subdue the effect of irrelevant neighbors, UaGGP also aggregates the clean representation in the original space and the learned representation. Experiments on benchmarks demonstrate the effectiveness of the proposed approach.

Click-through rate (CTR) prediction is an essential task in industrial applications such as video recommendation. Recently, deep learning models have been proposed to learn the representation of users' overall interests, while ignoring the fact that interests may dynamically change over time. We argue that it is necessary to consider the continuous-time information in CTR models to track user interest trend from rich historical behaviors. In this paper, we propose a novel Deep Time-Stream framework (DTS) which introduces the time information by an ordinary differential equations (ODE). DTS continuously models the evolution of interests using a neural network, and thus is able to tackle the challenge of dynamically representing users' interests based on their historical behaviors. In addition, our framework can be seamlessly applied to any existing deep CTR models by leveraging the additional Time-Stream Module, while no changes are made to the original CTR models. Experiments on public dataset as well as real industry dataset with billions of samples demonstrate the effectiveness of proposed approaches, which achieve superior performance compared with existing methods.

We aim to design universal algorithms for online convex optimization, which can handle multiple common types of loss functions simultaneously. The previous state-of-the-art universal method has achieved the minimax optimality for general convex, exponentially concave and strongly convex loss functions. However, it remains an open problem whether smoothness can be exploited to further improve the theoretical guarantees. In this paper, we provide an affirmative answer by developing a novel algorithm, namely UFO, which achieves O(√L*), O(d log L*) and O(log L*) regret bounds for the three types of loss functions respectively under the assumption of smoothness, where L* is the cumulative loss of the best comparator in hindsight, and d is dimensionality. Thus, our regret bounds are much tighter when the comparator has a small loss, and ensure the minimax optimality in the worst case. In addition, it is worth pointing out that UFO is the first to achieve the O(log L*) regret bound for strongly convex and smooth functions, which is tighter than the existing small-loss bound by an O(d) factor.

In conventional multi-label learning framework, each example is assumed to be represented by a single feature vector and associated with multiple valid labels simultaneously. Nonetheless, real-world objects usually exhibit complicated properties which can have multi-view feature representation as well as false positive labeling. Accordingly, the problem of multi-view partial multi-label learning (MVPML) is studied in this paper, where each example is assumed to be presented by multiple feature vectors while associated with multiple candidate labels which are only partially valid. To learn from MVPML examples, a novel approach named FIMAN is proposed which makes use of multi-view feature representation to tackle the noisy labeling information. Firstly, an aggregate manifold structure over training examples is generated by adaptively fusing affinity information conveyed by feature vectors of different views. Then, candidate labels of each training example are disambiguated by preserving the feature-induced manifold structure in label space. Finally, the resulting predictive models are learned by fitting modeling outputs with the disambiguated labels. Extensive experiments on a number of real-world data sets show that FIMAN achieves highly competitive performance against state-of-the-art approaches in solving the MVPML problem.

Visual Grounding is the task of associating entities in a natural language sentence with objects in an image. In this paper, we formulate visual grounding as a graph matching problem to find node correspondences between a visual scene graph and a language scene graph. These two graphs are heterogeneous, representing structure layouts of the sentence and image, respectively. We learn unified contextual node representations of the two graphs by using a cross-modal graph convolutional network to reduce their discrepancy. The graph matching is thus relaxed as a linear assignment problem because the learned node representations characterize both node information and structure information. A permutation loss and a semantic cycle-consistency loss are further introduced to solve the linear assignment problem with or without ground-truth correspondences. Experimental results on two visual grounding tasks, i.e., referring expression comprehension and phrase localization, demonstrate the effectiveness of our method.

In this paper, we propose a scene-aware context reasoning method that exploits context information from visual features for unsupervised abnormal event detection in videos, which bridges the semantic gap between visual context and the meaning of abnormal events. In particular, we build na spatio-temporal context graph to model visual context information including appearances of objects, spatio-temporal relationships among objects and scene types. The context information is encoded into the nodes and edges of the graph, and their states are iteratively updated by using multiple RNNs with message passing for context reasoning. To infer the spatio-temporal context graph in various scenes, we develop a graph-based deep Gaussian mixture model for scene clustering in an unsupervised manner. We then compute frame-level anomaly scores based on the context information to discriminate abnormal events in various scenes. Evaluations on three challenging datasets, including the UCF-Crime, Avenue, and ShanghaiTech datasets, demonstrate the effectiveness of our method.

Classical recommender system methods typically face the filter bubble problem when users only receive recommendations of their familiar items, making them bored and dissatisfied. To address the filter bubble problem, unexpected recommendations have been proposed to recommend items significantly deviating from user’s prior expectations and thus surprising them by presenting ”fresh” and previously unexplored items to the users. In this paper, we describe a novel Personalized Unexpected Recommender System (PURS) model that incorporates unexpectedness into the recommendation process by providing multi-cluster modeling of user interests in the latent space and personalized unexpectedness via the self-attention mechanism and via selection of an appropriate unexpected activation function. Extensive offline experiments on three real-world datasets illustrate that the proposed PURS model significantly outperforms the state-of-the-art baseline approaches in terms of both accuracy and unexpectedness measures. In addition, we conduct an online A/B test at a major video platform Alibaba-Youku, where our model achieves over 3% increase in the average video view per user metric. The proposed model is in the process of being deployed by the company.

Since the label collecting is prohibitive and time-consuming, unsupervised methods are preferred in applications such as fraud detection. Meanwhile, such applications usually require modeling the intrinsic clusters in high-dimensional data, which usually displays heterogeneous statistical patterns as the patterns of different clusters may appear in different dimensions. Existing methods propose to model the data clusters on selected dimensions, yet globally omitting any dimension may damage the pattern of certain clusters. To address the above issues, we propose a novel unsupervised generative framework called FIRD, which utilizes adversarial distributions to fit and disentangle the heterogeneous statistical patterns. When applying to discrete spaces, FIRD effectively distinguishes the synchronized fraudsters from normal users. Besides, FIRD also provides superior performance on anomaly detection datasets compared with SOTA anomaly detection methods (over 5% average AUC improvement). The significant experiment results on various datasets verify that the proposed method can better model the heterogeneous statistical patterns in high-dimensional data and benefit downstream applications.

We study Lipschitz bandits, where a learner repeatedly plays one arm from an infinite arm set and then receives a stochastic reward whose expectation is a Lipschitz function of the chosen arm. Most of existing work assume the reward distributions are bounded or at least sub-Gaussian, and thus do not apply to heavy-tailed rewards arising in many real-world scenarios such as web advertising and financial markets. To address this limitation, in this paper we relax the assumption on rewards to allow arbitrary distributions that have finite $(1+\epsilon)$-th moments for some $\epsilon \in (0, 1]$, and propose algorithms that enjoy a sublinear regret of $\widetilde{O}(T^{(d_z\epsilon + 1)/(d_z \epsilon + \epsilon + 1)})$ where $T$ is the time horizon and $d_z$ is the zooming dimension. The key idea is to exploit the Lipschitz property of the expected reward function by adaptively discretizing the arm set, and employ upper confidence bound policies with robust mean estimators designed for heavy-tailed distributions. Furthermore, we provide a lower bound for Lipschitz bandits with heavy-tailed rewards, and show that our algorithms are optimal in terms of $T$. Finally, we conduct numerical experiments to demonstrate the effectiveness of our algorithms.

On many video websites, the recommendation is implemented as a prediction problem of video-user pairs, where the videos are represented by text features extracted from the metadata. However, the metadata is manually annotated by users and is usually missing for online videos. To train an effective recommender system with lower annotation cost, we propose an active learning approach to fully exploit the visual view of videos, while querying as few annotations as possible from the text view. On one hand, a joint model is proposed to learn the mapping from visual view to text view by simultaneously aligning the two views and minimizing the classification loss. On the other hand, a novel strategy based on prediction inconsistency and watching frequency is proposed to actively select the most important videos for metadata querying. Experiments on both classification datasets and real video recommendation tasks validate that the proposed approach can significantly reduce the annotation cost.

In this paper, we study the multi-objective bandits (MOB) problem, where a learner repeatedly selects one arm to play and then receives a reward vector consisting of multiple objectives. MOB has found many real-world applications as varied as online recommendation and network routing. On the other hand, these applications typically contain contextual information that can guide the learning process which, however, is ignored by most of existing work. To utilize this information, we associate each arm with a context vector and assume the reward follows the generalized linear model (GLM). We adopt the notion of Pareto regret to evaluate the learner's performance and develop a novel algorithm for minimizing it. The essential idea is to apply a variant of the online Newton step to estimate model parameters, based on which we utilize the upper confidence bound (UCB) policy to construct an approximation of the Pareto front, and then uniformly at random choose one arm from the approximate Pareto front. Theoretical analysis shows that the proposed algorithm achieves an \tilde O(d\sqrt{T}) Pareto regret, where T is the time horizon and d is the dimension of contexts, which matches the optimal result for single objective contextual bandits problem. Numerical experiments demonstrate the effectiveness of our method.

Multi-view multi-label learning serves an important framework to learn from objects with diverse representations and rich semantics. Existing multi-view multi-label learning techniques focus on exploiting shared subspace for fusing multi-view representations, where helpful view-specific information for discriminative modeling is usually ignored. In this paper, a novel multi-view multi-label learning approach named SIMM is proposed which leverages shared subspace exploitation and view-specific information extraction. For shared subspace exploitation, SIMM jointly minimizes confusion adversarial loss and multi-label loss to utilize shared information from all views. For view-specific information extraction, SIMM enforces an orthogonal constraint w.r.t. the shared subspace to utilize view-specific discriminative information. Extensive experiments on real-world data sets clearly show the favorable performance of SIMM against other state-of-the-art multi-view multi-label learning approaches.

Estimating missing entries in matrices has attracted much attention due to its wide range of applications like image inpainting and video denoising, which are usually considered as low-rank matrix completion problems theoretically. It is common to consider nuclear norm as a surrogate of the rank operator since it is the tightest convex lower bound of the rank operator under certain conditions. However, most approaches based on nuclear norm minimization involve a number of singular value decomposition (SVD) operations. Given a matrix X ∈ R m × n , the time complexity of the SVD operation is O ( mn 2 ), which brings prohibitive computational burden on large-scale matrices, limiting the further usage of these methods in real applications. Motivated by this observation, a series of atom-decomposition-based matrix completion methods have been studied. The key to these methods is to reconstruct the target matrix by pursuit methods in a greedy way, which only involves the computation of the top SVD and has great advantages in efficiency compared with the SVD-based matrix completion methods. However, due to gradually serious accumulation errors, atom-decomposition-based methods usually result in unsatisfactory reconstruction accuracy. In this article, we propose a new efficient and scalable atom decomposition algorithm for matrix completion called Adaptive Basis Selection Strategy ( ABSS ). Different from traditional greedy atom decomposition methods, a two-phase strategy is conducted to generate the basis separately via different strategies according to their different nature. At first, we globally prune the basis space to eliminate the unimportant basis as much as possible and locate the probable subspace containing the most informative basis. Then, another group of basis spaces are learned to improve the recovery accuracy based on local information. In this way, our proposed algorithm breaks through the accuracy bottleneck of traditional atom-decomposition-based matrix completion methods; meanwhile, it reserves the innate efficiency advantages over SVD-based matrix completion methods. We empirically evaluate the proposed algorithm ABSS on real visual image data and large-scale recommendation datasets. Results have shown that ABSS has much better reconstruction accuracy with comparable cost to atom-decomposition-based methods. At the same time, it outperforms the state-of-the-art SVD-based matrix completion algorithms by similar or better reconstruction accuracy with enormous advantages on efficiency.

In this paper, we focus on Stochastic Composite Optimization (SCO) for sparse learning that aims to learn a sparse solution. Although many SCO algorithms have been developed for sparse learning with an optimal convergence rate $O(1/T)$, they often fail to deliver sparse solutions at the end either because of the limited sparsity regularization during stochastic optimization or due to the limitation in online-to-batch conversion. To improve the sparsity of solutions obtained by SCO, we propose a simple but effective stochastic optimization scheme that adds a novel sparse online-to-batch conversion to the traditional SCO algorithms. The theoretical analysis shows that our scheme can find a solution with better sparse patterns without affecting the convergence rate. Experimental results on both synthetic and real-world data sets show that the proposed methods are more effective in recovering the sparse solution and have comparable convergence rate as the state-of-the-art SCO algorithms for sparse learning.

Cross media retrieval engines have gained massive popularity with rapid development of the Internet. Users may perform queries in a corpus consisting of audio, video, and textual information. To make such systems practically possible for large mount of multimedia data, two critical issues must be carefully considered: (a) reduce the storage as much as possible; (b) model the relationship of the heterogeneous media data. Recently academic community have proved that encoding the data into compact binary codes can drastically reduce the storage and computational cost. However, it is still unclear how to integrate multiple information sources properly into the binary code encoding scheme.

As pairwise constraints are usually easier to access than label information, pairwise constraint propagation attracts more and more attention in semi-supervised learning. Most existing pairwise constraint propagation methods are based on canonical graph propagation model, which heavily depends on the edge weights in the graph and cannot preserve local and global consistency simultaneously. In order to address this drawback, we cast cross-view pairwise constraint propagation into a problem of low rank matrix completion and propose a Matrix Completion method for cross-view Pairwise Constraint Propagation(MCPCP). With low rank requirement and graph regularization, our MCPCP can preserve local and global consistency simultaneously. We develop an algorithm based on alternating direction method of multipliers(ADMM) to solve the optimization problem. Finally, the effectiveness of MCPCP is demonstrated in cross-view multimedia retrieval.

Recently, the hashing techniques have been widely applied to approximate the nearest neighbor search problem in many real applications. The basic idea of these approaches is to generate binary codes for data points which can preserve the similarity between any two of them. Given a query, instead of performing a linear scan of the entire data base, the hashing method can perform a linear scan of the points whose hamming distance to the query is not greater than rh , where rh is a constant. However, in order to find the true nearest neighbors, both the locating time and the linear scan time are proportional to O(∑i=0(rh)(c || i)) ( c is the code length), which increase exponentially as rh increases. To address this limitation, we propose a novel algorithm named iterative expanding hashing in this paper, which builds an auxiliary index based on an offline constructed nearest neighbor table to avoid large rh . This auxiliary index can be easily combined with all the traditional hashing methods. Extensive experimental results over various real large-scale datasets demonstrate the superiority of the proposed approach.

Recently, hashing techniques have been widely applied to solve the approximate nearest neighbors search problem in many vision applications. Generally, these hashing approaches generate 2^c buckets, where c is the length of the hash code. A good hashing method should satisfy the following two requirements: 1) mapping the nearby data points into the same bucket or nearby (measured by the Hamming distance) buckets. 2) all the data points are evenly distributed among all the buckets. In this paper, we propose a novel algorithm named Complementary Projection Hashing (CPH) to find the optimal hashing functions which explicitly considers the above two requirements. Specifically, CPH aims at sequentially finding a series of hyper planes (hashing functions) which cross the sparse region of the data. At the same time, the data points are evenly distributed in the hyper cubes generated by these hyper planes. The experiments comparing with the state-of-the-art hashing methods demonstrate the effectiveness of the proposed method.

Recovering a large matrix from a small subset of its entries is a challenging problem arising in many real applications, such as image inpainting and recommender systems. Many existing approaches formulate this problem as a general low-rank matrix approximation problem. Since the rank operator is nonconvex and discontinuous, most of the recent theoretical studies use the nuclear norm as a convex relaxation. One major limitation of the existing approaches based on nuclear norm minimization is that all the singular values are simultaneously minimized, and thus the rank may not be well approximated in practice. In this paper, we propose to achieve a better approximation to the rank of matrix by truncated nuclear norm, which is given by the nuclear norm subtracted by the sum of the largest few singular values. In addition, we develop a novel matrix completion algorithm by minimizing the Truncated Nuclear Norm. We further develop three efficient iterative procedures, TNNR-ADMM, TNNR-APGL, and TNNR-ADMMAP, to solve the optimization problem. TNNR-ADMM utilizes the alternating direction method of multipliers (ADMM), while TNNR-AGPL applies the accelerated proximal gradient line search method (APGL) for the final optimization. For TNNR-ADMMAP, we make use of an adaptive penalty according to a novel update rule for ADMM to achieve a faster convergence rate. Our empirical study shows encouraging results of the proposed algorithms in comparison to the state-of-the-art matrix completion algorithms on both synthetic and real visual datasets.

Estimating missing values in visual data is a challenging problem in computer vision, which can be considered as a low rank matrix approximation problem. Most of the recent studies use the nuclear norm as a convex relaxation of the rank operator. However, by minimizing the nuclear norm, all the singular values are simultaneously minimized, and thus the rank can not be well approximated in practice. In this paper, we propose a novel matrix completion algorithm based on the Truncated Nuclear Norm Regularization (TNNR) by only minimizing the smallest N-r singular values, where N is the number of singular values and r is the rank of the matrix. In this way, the rank of the matrix can be better approximated than the nuclear norm. We further develop an efficient iterative procedure to solve the optimization problem by using the alternating direction method of multipliers and the accelerated proximal gradient line search method. Experimental results in a wide range of applications demonstrate the effectiveness of our proposed approach.

Recovering a large matrix from a small subset of its entries is a challenging problem arising in many real world applications, such as recommender system and image in-painting. These problems can be formulated as a general matrix completion problem. The Singular Value Thresholding (SVT) algorithm is a simple and efficient first-order matrix completion method to recover the missing values when the original data matrix is of low rank. SVT has been applied successfully in many applications. However, SVT is computationally expensive when the size of the data matrix is large, which significantly limits its applicability. In this paper, we propose an Accelerated Singular Value Thresholding (ASVT) algorithm which improves the convergence rate from O(1/N) for SVT to O(1/N2), where N is the number of iterations during optimization. Specifically, the dual problem of the nuclear norm minimization problem is derived and an adaptive line search scheme is introduced to solve this dual problem. Consequently, the optimal solution of the primary problem can be readily obtained from that of the dual problem. We have conducted a series of experiments on a synthetic dataset, a distance matrix dataset and a large movie rating dataset. The experimental results have demonstrated the efficiency and effectiveness of the proposed algorithm.