Fei Xiao

Yuqing Wang, Zhijie Lin, Ceyuan Yang et al.

Unified multimodal models (UMMs) aim to handle perception and generation in a single model. Yet existing UMMs still rely on a frozen, separately pretrained VAE for image generation, imposing a structural bottleneck. Naively removing it introduces a quality gap, as the model must learn both high-level structure and low-level details from raw pixels. In this paper, we propose Representation Forcing (RF), a technique that closes this gap by making representation prediction a native capability of the model. Concretely, RF forces the decoder to autoregressively predict visual representations as intermediate tokens before pixels; these tokens then stay in context to guide pixel diffusion within the same backbone. By turning representations from perception outputs into generation targets, RF eliminates the need for any external generative latent space. We find that RF benefits both understanding and generation. On image generation, our pixel-space model with RF matches state-of-the-art VAE-based unified models. On image understanding, pixel-space RF generally outperforms its VAE-based variant. Together, these results offer an effective step toward end-to-end, bottleneck-free UMMs.

Fei Xiao, Shaofeng Cai, Gang Chen et al.

Fraud detection presents a challenging task characterized by ever-evolving fraud patterns and scarce labeled data. Existing methods predominantly rely on graph-based or sequence-based approaches. While graph-based approaches connect users through shared entities to capture structural information, they remain vulnerable to fraudsters who can disrupt or manipulate these connections. In contrast, sequence-based approaches analyze users' behavioral patterns, offering robustness against tampering but overlooking the interactions between similar users. Inspired by cohort analysis in retention and healthcare, this paper introduces VecAug, a novel cohort-augmented learning framework that addresses these challenges by enhancing the representation learning of target users with personalized cohort information. To this end, we first propose a vector burn-in technique for automatic cohort identification, which retrieves a task-specific cohort for each target user. Then, to fully exploit the cohort information, we introduce an attentive cohort aggregation technique for augmenting target user representations. To improve the robustness of such cohort augmentation, we also propose a novel label-aware cohort neighbor separation mechanism to distance negative cohort neighbors and calibrate the aggregated cohort information. By integrating this cohort information with target user representations, VecAug enhances the modeling capacity and generalization capabilities of the model to be augmented. Our framework is flexible and can be seamlessly integrated with existing fraud detection models. We deploy our framework on e-commerce platforms and evaluate it on three fraud detection datasets, and results show that VecAug improves the detection performance of base models by up to 2.48\% in AUC and 22.5\% in R@P$_{0.9}$, outperforming state-of-the-art methods significantly.

Yuan Zhang, Fei Xiao, Tao Huang et al.

Large vision-language models (LVLMs) have recently achieved rapid progress, exhibiting great perception and reasoning abilities concerning visual information. However, when faced with prompts in different sizes of solution spaces, LVLMs fail to always give consistent answers regarding the same knowledge point. This inconsistency of answers between different solution spaces is prevalent in LVLMs and erodes trust. To this end, we provide a multi-modal benchmark ConBench, to intuitively analyze how LVLMs perform when the solution space of a prompt revolves around a knowledge point. Based on the ConBench tool, we are the first to reveal the tapestry and get the following findings: (1) In the discriminate realm, the larger the solution space of the prompt, the lower the accuracy of the answers. (2) Establish the relationship between the discriminative and generative realms: the accuracy of the discriminative question type exhibits a strong positive correlation with its Consistency with the caption. (3) Compared to open-source models, closed-source models exhibit a pronounced bias advantage in terms of Consistency. Eventually, we ameliorate the consistency of LVLMs by trigger-based diagnostic refinement, indirectly improving the performance of their caption. We hope this paper will accelerate the research community in better evaluating their models and encourage future advancements in the consistency domain. The project is available at https://github.com/foundation-multimodal-models/ConBench.

Mingming Zhang, Feiqing Zhuang, Na Li et al.

Automated bidding is central to modern digital advertising. Early rule-based methods lacked adaptability, while subsequent Reinforcement Learning approaches modeled bidding as a Markov Decision Process but struggled with long-term dependencies. Recent generative models show promise, yet they lack explicit mechanisms to balance exploration and safety, relying solely on action perturbations or trajectory guidance without a safety fallback. This results in inefficient exploration and elevated financial risk for advertising platforms. To address this gap, we propose GUIDE (Generative Auto-Bidding with Unified Modeling and Exploration), a framework that synergistically integrates directed exploration with a safe fallback mechanism. GUIDE employs a Decision Transformer (DT) to jointly model historical bidding actions and environmental state transitions. A Q-value module guides the DT's exploration via regularization constraints, while an Inverse Dynamics Module (IDM) leverages DT-predicted future states to infer robust, behaviorally consistent actions as a safe policy fallback. The Q-value module then adaptively selects the final action between these two options, balancing exploration and safety. Together, these components form an integrated "explore-safeguard-select" pipeline that unifies efficiency and safety. We conduct extensive experiments on public datasets, in simulated auction environments, and through large-scale online deployment on Taobao, a leading Chinese advertising platform. Results show GUIDE consistently outperforms state-of-the-art baselines across all scenarios. In real-world deployment, GUIDE achieves notable gains: +4.10% ad GMV, +1.40% ad clicks, +1.66% ad cost, and +3.52% ad ROI, demonstrating its effectiveness and strong industrial applicability.

Yifan Wang, Yixuan Wang, YiDan Liang et al.

New item growth is critical for maintaining a healthy ecosystem in large-scale e-commerce platforms. However, existing systems tend to prioritize presenting users with already popular items, a phenomenon often referred to as the "Matthew effect". In the context of search retrieval, current cold-start models suffer from the misalignment between training objectives and online business metrics, and they lack effective mechanisms to measure an item's growth potential. In this paper, we propose a Multi-Value-Aware retrieval framework tailored for e-commerce search, designed to better align with the cascaded online values across different stages of the search system while balancing immediate conversion and long-term item growth. Our framework GrowthGR consists of two key components: an Item Long-term Transaction Value Prediction (ItemLTV) module and a Multi-Value-Aware Generative Retrieval (MultiGR) module. First, in the ItemLTV module, we employ counterfactual inference to quantify the long-term value increment attributable to a single user interaction. Second, in the MultiGR module, building upon a semantic-ID-based generative retrieval architecture, we leverage structured samples with the search cascade signals and adopt a Multi-Value-Aware Policy Optimization (MoPO) training paradigm to align with multi-stage online values, while explicitly balancing short-term transactional value and long-term growth potential estimated by ItemLTV. We successfully deployed GrowthGR on Taobao's production platform, achieving a substantial 5.3% lift in new item GMV while delivering a non-trivial 0.3% gain in overall search GMV. Extensive online analysis and A/B testing demonstrate its positive impact on the overall ecosystem value.

Yinglei Ma, Fei Xiao

BJT-based 2D temperature-sensor arrays are factory-calibrated to +/-0.1 degC, but post-deployment thermal and mechanical stresses drift their per-sensor gain-offset parameters by an order of magnitude, and in-lab recalibration is impractical. We present RASC (Region-Aware Self-Calibration), a five-stage algorithm that decomposes the global ill-posed problem into local cluster-level problems, runs robust alternating estimation (trimmed-mean field reconstruction + Huber IRLS) inside each cluster, and reconciles overlapping estimates by linear consensus on the cluster-overlap graph with provable exponential convergence. On 7,632 frames from a deployed 16x16 array exhibiting ~5x factory-spec non-uniformity, RASC cuts the locally-non-smooth fixed-pattern residual by 71+/-5% (10-fold CV), restoring +/-0.1 degC accuracy while perturbing the calibrated field by only 0.041 degC RMSE; reduction concentrates at the edges (78% vs 55% interior). In simulations on 8x8 to 32x32 arrays, RASC matches an oracle centralized EKF within 0.10 degC with ~4x lower bandwidth.

Team Seawead, Ceyuan Yang, Zhijie Lin et al.

This technical report presents a cost-efficient strategy for training a video generation foundation model. We present a mid-sized research model with approximately 7 billion parameters (7B) called Seaweed-7B trained from scratch using 665,000 H100 GPU hours. Despite being trained with moderate computational resources, Seaweed-7B demonstrates highly competitive performance compared to contemporary video generation models of much larger size. Design choices are especially crucial in a resource-constrained setting. This technical report highlights the key design decisions that enhance the performance of the medium-sized diffusion model. Empirically, we make two observations: (1) Seaweed-7B achieves performance comparable to, or even surpasses, larger models trained on substantially greater GPU resources, and (2) our model, which exhibits strong generalization ability, can be effectively adapted across a wide range of downstream applications either by lightweight fine-tuning or continue training. See the project page at https://seaweed.video/

Yu Gao, Haoyuan Guo, Tuyen Hoang et al.

Notable breakthroughs in diffusion modeling have propelled rapid improvements in video generation, yet current foundational model still face critical challenges in simultaneously balancing prompt following, motion plausibility, and visual quality. In this report, we introduce Seedance 1.0, a high-performance and inference-efficient video foundation generation model that integrates several core technical improvements: (i) multi-source data curation augmented with precision and meaningful video captioning, enabling comprehensive learning across diverse scenarios; (ii) an efficient architecture design with proposed training paradigm, which allows for natively supporting multi-shot generation and jointly learning of both text-to-video and image-to-video tasks. (iii) carefully-optimized post-training approaches leveraging fine-grained supervised fine-tuning, and video-specific RLHF with multi-dimensional reward mechanisms for comprehensive performance improvements; (iv) excellent model acceleration achieving ~10x inference speedup through multi-stage distillation strategies and system-level optimizations. Seedance 1.0 can generate a 5-second video at 1080p resolution only with 41.4 seconds (NVIDIA-L20). Compared to state-of-the-art video generation models, Seedance 1.0 stands out with high-quality and fast video generation having superior spatiotemporal fluidity with structural stability, precise instruction adherence in complex multi-subject contexts, native multi-shot narrative coherence with consistent subject representation.

Ying Song, Yijing Wang, Hui Yang et al.

Evaluating platform-level interventions in search-based two-sided marketplaces is fundamentally challenged by systemic effects such as spillovers and network interference. While widely used for causal inference, the PSM (Propensity Score Matching) - DID (Difference-in-Differences) framework remains susceptible to selection bias and cross-unit interference from unaccounted spillovers. In this paper, we introduced Competitive Isolation PSM-DID, a novel causal framework that integrates propensity score matching with competitive isolation to enable platform-level effect measurement (e.g., order volume, GMV) instead of item-level metrics in search systems. Our approach provides theoretically guaranteed unbiased estimation under mutual exclusion conditions, with an open dataset released to support reproducible research on marketplace interference (github.com/xxxx). Extensive experiments demonstrate significant reductions in interference effects and estimation variance compared to baseline methods. Successful deployment in a large-scale marketplace confirms the framework's practical utility for platform-level causal inference.

Jianyi Wang, Zhijie Lin, Meng Wei et al.

Video restoration poses non-trivial challenges in maintaining fidelity while recovering temporally consistent details from unknown degradations in the wild. Despite recent advances in diffusion-based restoration, these methods often face limitations in generation capability and sampling efficiency. In this work, we present SeedVR, a diffusion transformer designed to handle real-world video restoration with arbitrary length and resolution. The core design of SeedVR lies in the shifted window attention that facilitates effective restoration on long video sequences. SeedVR further supports variable-sized windows near the boundary of both spatial and temporal dimensions, overcoming the resolution constraints of traditional window attention. Equipped with contemporary practices, including causal video autoencoder, mixed image and video training, and progressive training, SeedVR achieves highly-competitive performance on both synthetic and real-world benchmarks, as well as AI-generated videos. Extensive experiments demonstrate SeedVR's superiority over existing methods for generic video restoration.

Ceyuan Yang, Zhijie Lin, Yang Zhao et al.

We present Omni, a unified multimodal model natively trained on diverse modalities, including text, images, videos, 3D geometry, and hidden representations. We find that such training enables Context Unrolling, where the model explicitly reasons across multiple modal representations before producing predictions. This process enables the model to aggregate complementary information across heterogeneous modalities, facilitating a more faithful approximation of the shared multimodal knowledge manifold and improving downstream reasoning fidelity. As a result, Omni achieves strong performance on both multimodal generation and understanding benchmarks, while demonstrating advanced multimodal reasoning capabilities, including in-context generation of text, image, video, and 3D geometry.

Team Seedance, Heyi Chen, Siyan Chen et al.

Recent strides in video generation have paved the way for unified audio-visual generation. In this work, we present Seedance 1.5 pro, a foundational model engineered specifically for native, joint audio-video generation. Leveraging a dual-branch Diffusion Transformer architecture, the model integrates a cross-modal joint module with a specialized multi-stage data pipeline, achieving exceptional audio-visual synchronization and superior generation quality. To ensure practical utility, we implement meticulous post-training optimizations, including Supervised Fine-Tuning (SFT) on high-quality datasets and Reinforcement Learning from Human Feedback (RLHF) with multi-dimensional reward models. Furthermore, we introduce an acceleration framework that boosts inference speed by over 10X. Seedance 1.5 pro distinguishes itself through precise multilingual and dialect lip-syncing, dynamic cinematic camera control, and enhanced narrative coherence, positioning it as a robust engine for professional-grade content creation. Seedance 1.5 pro is now accessible on Volcano Engine at https://console.volcengine.com/ark/region:ark+cn-beijing/experience/vision?type=GenVideo.

Guangting Zheng, Qinyu Zhao, Tao Yang et al.

Directly modeling the explicit likelihood of the raw data distribution is key topic in the machine learning area, which achieves the scaling successes in Large Language Models by autoregressive modeling. However, continuous AR modeling over visual pixel data suffer from extremely long sequences and high-dimensional spaces. In this paper, we present FARMER, a novel end-to-end generative framework that unifies Normalizing Flows (NF) and Autoregressive (AR) models for tractable likelihood estimation and high-quality image synthesis directly from raw pixels. FARMER employs an invertible autoregressive flow to transform images into latent sequences, whose distribution is modeled implicitly by an autoregressive model. To address the redundancy and complexity in pixel-level modeling, we propose a self-supervised dimension reduction scheme that partitions NF latent channels into informative and redundant groups, enabling more effective and efficient AR modeling. Furthermore, we design a one-step distillation scheme to significantly accelerate inference speed and introduce a resampling-based classifier-free guidance algorithm to boost image generation quality. Extensive experiments demonstrate that FARMER achieves competitive performance compared to existing pixel-based generative models while providing exact likelihoods and scalable training.

Fei Xiao, Liang Pang, Yanyan Lan et al.

Unsupervised style transfer models are mainly based on an inductive learning approach, which represents the style as embeddings, decoder parameters, or discriminator parameters and directly applies these general rules to the test cases. However, the lacking of parallel corpus hinders the ability of these inductive learning methods on this task. As a result, it is likely to cause severe inconsistent style expressions, like `the salad is rude`. To tackle this problem, we propose a novel transductive learning approach in this paper, based on a retrieval-based context-aware style representation. Specifically, an attentional encoder-decoder with a retriever framework is utilized. It involves top-K relevant sentences in the target style in the transfer process. In this way, we can learn a context-aware style embedding to alleviate the above inconsistency problem. In this paper, both sparse (BM25) and dense retrieval functions (MIPS) are used, and two objective functions are designed to facilitate joint learning. Experimental results show that our method outperforms several strong baselines. The proposed transductive learning approach is general and effective to the task of unsupervised style transfer, and we will apply it to the other two typical methods in the future.

Zhou Xing, Fei Xiao

Predictions of driver's intentions and their behaviors using the road is of great importance for planning and decision making processes of autonomous driving vehicles. In particular, relatively short-term driving intentions are the fundamental units that constitute more sophisticated driving goals, behaviors, such as overtaking the slow vehicle in front, exit or merge onto a high way, etc. While it is not uncommon that most of the time human driver can rationalize, in advance, various on-road behaviors, intentions, as well as the associated risks, aggressiveness, reciprocity characteristics, etc., such reasoning skills can be challenging and difficult for an autonomous driving system to learn. In this article, we demonstrate a disciplined methodology that can be used to build and train a predictive drive system, therefore to learn the on-road characteristics aforementioned.

Yusen Zhan, Qing Da, Fei Xiao et al.

In industrial large-scale search systems, such as Taobao.com search for commodities, the quality of the ranking result is getting continually improved by introducing more factors from complex procedures, e.g., deep neural networks for extracting image factors. Meanwhile, the increasing of the factors demands more computation resource and raises the system response latency. It has been observed that a search instance usually requires only a small set of effective factors, instead of all factors. Therefore, removing ineffective factors significantly improves the system efficiency. This paper studies the \emph{Contextual Factor Selection} (CFS), which selects only a subset of effective factors for every search instance, for a well balance between the search quality and the response latency. We inject CFS into the search engine ranking score to accelerate the engine, considering both ranking effectiveness and efficiency. The learning of the CFS model involves a combinatorial optimization, which is transformed as a sequential decision-making problem. Solving the problem by reinforcement learning, we propose the RankCFS, which has been assessed in an off-line environment as well as a real-world on-line environment (Taobao.com). The empirical results show that, the proposed CFS approach outperforms several existing supervised/unsupervised methods for feature selection in the off-line environment, and also achieves significant real-world performance improvement, in term of service latency, in daily test as well as Singles' Day Shopping Festival in $2017$.

Shichen Liu, Fei Xiao, Wenwu Ou et al.

In the 'Big Data' era, many real-world applications like search involve the ranking problem for a large number of items. It is important to obtain effective ranking results and at the same time obtain the results efficiently in a timely manner for providing good user experience and saving computational costs. Valuable prior research has been conducted for learning to efficiently rank like the cascade ranking (learning) model, which uses a sequence of ranking functions to progressively filter some items and rank the remaining items. However, most existing research of learning to efficiently rank in search is studied in a relatively small computing environments with simulated user queries. This paper presents novel research and thorough study of designing and deploying a Cascade model in a Large-scale Operational E-commerce Search application (CLOES), which deals with hundreds of millions of user queries per day with hundreds of servers. The challenge of the real-world application provides new insights for research: 1). Real-world search applications often involve multiple factors of preferences or constraints with respect to user experience and computational costs such as search accuracy, search latency, size of search results and total CPU cost, while most existing search solutions only address one or two factors; 2). Effectiveness of e-commerce search involves multiple types of user behaviors such as click and purchase, while most existing cascade ranking in search only models the click behavior. Based on these observations, a novel cascade ranking model is designed and deployed in an operational e-commerce search application. An extensive set of experiments demonstrate the advantage of the proposed work to address multiple factors of effectiveness, efficiency and user experience in the real-world application.

Zhou Xing, Eddy Baik, Yan Jiao et al.

While both the data volume and heterogeneity of the digital music content is huge, it has become increasingly important and convenient to build a recommendation or search system to facilitate surfacing these content to the user or consumer community. Most of the recommendation models fall into two primary species, collaborative filtering based and content based approaches. Variants of instantiations of collaborative filtering approach suffer from the common issues of so called "cold start" and "long tail" problems where there is not much user interaction data to reveal user opinions or affinities on the content and also the distortion towards the popular content. Content-based approaches are sometimes limited by the richness of the available content data resulting in a heavily biased and coarse recommendation result. In recent years, the deep neural network has enjoyed a great success in large-scale image and video recognitions. In this paper, we propose and experiment using deep convolutional neural network to imitate how human brain processes hierarchical structures in the auditory signals, such as music, speech, etc., at various timescales. This approach can be used to discover the latent factor models of the music based upon acoustic hyper-images that are extracted from the raw audio waves of music. These latent embeddings can be used either as features to feed to subsequent models, such as collaborative filtering, or to build similarity metrics between songs, or to classify music based on the labels for training such as genre, mood, sentiment, etc.

Na Fu, Hoong Chuin Lau, Pradeep R. Varakantham et al.

Scheduling problems in manufacturing, logistics and project management have frequently been modeled using the framework of Resource Constrained Project Scheduling Problems with minimum and maximum time lags (RCPSP/max). Due to the importance of these problems, providing scalable solution schedules for RCPSP/max problems is a topic of extensive research. However, all existing methods for solving RCPSP/max assume that durations of activities are known with certainty, an assumption that does not hold in real world scheduling problems where unexpected external events such as manpower availability, weather changes, etc. lead to delays or advances in completion of activities. Thus, in this paper, our focus is on providing a scalable method for solving RCPSP/max problems with durational uncertainty. To that end, we introduce the robust local search method consisting of three key ideas: (a) Introducing and studying the properties of two decision rule approximations used to compute start times of activities with respect to dynamic realizations of the durational uncertainty; (b) Deriving the expression for robust makespan of an execution strategy based on decision rule approximations; and (c) A robust local search mechanism to efficiently compute activity execution strategies that are robust against durational uncertainty. Furthermore, we also provide enhancements to local search that exploit temporal dependencies between activities. Our experimental results illustrate that robust local search is able to provide robust execution strategies efficiently.