Xiaowen Shi

Xiaowen Shi, Fan Yang, Ze Wang et al.

Re-ranking draws increased attention on both academics and industries, which rearranges the ranking list by modeling the mutual influence among items to better meet users' demands. Many existing re-ranking methods directly take the initial ranking list as input, and generate the optimal permutation through a well-designed context-wise model, which brings the evaluation-before-reranking problem. Meanwhile, evaluating all candidate permutations brings unacceptable computational costs in practice. Thus, to better balance efficiency and effectiveness, online systems usually use a two-stage architecture which uses some heuristic methods such as beam-search to generate a suitable amount of candidate permutations firstly, which are then fed into the evaluation model to get the optimal permutation. However, existing methods in both stages can be improved through the following aspects. As for generation stage, heuristic methods only use point-wise prediction scores and lack an effective judgment. As for evaluation stage, most existing context-wise evaluation models only consider the item context and lack more fine-grained feature context modeling. This paper presents a novel end-to-end re-ranking framework named PIER to tackle the above challenges which still follows the two-stage architecture and contains two mainly modules named FPSM and OCPM. We apply SimHash in FPSM to select top-K candidates from the full permutation based on user's permutation-level interest in an efficient way. Then we design a novel omnidirectional attention mechanism in OCPM to capture the context information in the permutation. Finally, we jointly train these two modules end-to-end by introducing a comparative learning loss. Offline experiment results demonstrate that PIER outperforms baseline models on both public and industrial datasets, and we have successfully deployed PIER on Meituan food delivery platform.

Guogang Liao, Xiaowen Shi, Ze Wang et al. · tsinghua

A mixed list of ads and organic items is usually displayed in feed and how to allocate the limited slots to maximize the overall revenue is a key problem. Meanwhile, modeling user preference with historical behavior is essential in recommendation and advertising (e.g., CTR prediction and ads allocation). Most previous works for user behavior modeling only model user's historical point-level positive feedback (i.e., click), which neglect the page-level information of feedback and other types of feedback. To this end, we propose Deep Page-level Interest Network (DPIN) to model the page-level user preference and exploit multiple types of feedback. Specifically, we introduce four different types of page-level feedback as input, and capture user preference for item arrangement under different receptive fields through the multi-channel interaction module. Through extensive offline and online experiments on Meituan food delivery platform, we demonstrate that DPIN can effectively model the page-level user preference and increase the revenue for the platform.

Ze Wang, Guogang Liao, Xiaowen Shi et al. · tsinghua

Ads allocation, which involves allocating ads and organic items to limited slots in feed with the purpose of maximizing platform revenue, has become a research hotspot. Notice that, e-commerce platforms usually have multiple entrances for different categories and some entrances have few visits. Data from these entrances has low coverage, which makes it difficult for the agent to learn. To address this challenge, we propose Similarity-based Hybrid Transfer for Ads Allocation (SHTAA), which effectively transfers samples as well as knowledge from data-rich entrance to data-poor entrance. Specifically, we define an uncertainty-aware similarity for MDP to estimate the similarity of MDP for different entrances. Based on this similarity, we design a hybrid transfer method, including instance transfer and strategy transfer, to efficiently transfer samples and knowledge from one entrance to another. Both offline and online experiments on Meituan food delivery platform demonstrate that the proposed method could achieve better performance for data-poor entrance and increase the revenue for the platform.

Ze Wang, Guogang Liao, Xiaowen Shi et al. · tsinghua

With the recent prevalence of reinforcement learning (RL), there have been tremendous interests in utilizing RL for ads allocation in recommendation platforms (e.g., e-commerce and news feed sites). To achieve better allocation, the input of recent RL-based ads allocation methods is upgraded from point-wise single item to list-wise item arrangement. However, this also results in a high-dimensional space of state-action pairs, making it difficult to learn list-wise representations with good generalization ability. This further hinders the exploration of RL agents and causes poor sample efficiency. To address this problem, we propose a novel RL-based approach for ads allocation which learns better list-wise representations by leveraging task-specific signals on Meituan food delivery platform. Specifically, we propose three different auxiliary tasks based on reconstruction, prediction, and contrastive learning respectively according to prior domain knowledge on ads allocation. We conduct extensive experiments on Meituan food delivery platform to evaluate the effectiveness of the proposed auxiliary tasks. Both offline and online experimental results show that the proposed method can learn better list-wise representations and achieve higher revenue for the platform compared to the state-of-the-art baselines.

Xiaowen Shi, Ze Wang, Yuanying Cai et al.

Nowadays, the mainstream approach in position allocation system is to utilize a reinforcement learning model to allocate appropriate locations for items in various channels and then mix them into the feed. There are two types of data employed to train reinforcement learning (RL) model for position allocation, named strategy data and random data. Strategy data is collected from the current online model, it suffers from an imbalanced distribution of state-action pairs, resulting in severe overestimation problems during training. On the other hand, random data offers a more uniform distribution of state-action pairs, but is challenging to obtain in industrial scenarios as it could negatively impact platform revenue and user experience due to random exploration. As the two types of data have different distributions, designing an effective strategy to leverage both types of data to enhance the efficacy of the RL model training has become a highly challenging problem. In this study, we propose a framework named Multi-Distribution Data Learning (MDDL) to address the challenge of effectively utilizing both strategy and random data for training RL models on mixed multi-distribution data. Specifically, MDDL incorporates a novel imitation learning signal to mitigate overestimation problems in strategy data and maximizes the RL signal for random data to facilitate effective learning. In our experiments, we evaluated the proposed MDDL framework in a real-world position allocation system and demonstrated its superior performance compared to the previous baseline. MDDL has been fully deployed on the Meituan food delivery platform and currently serves over 300 million users.

Meituan LongCat Team, Anchun Gui, Bei Li et al.

We present LongCat-Flash-Thinking, an efficient 560-billion-parameter open-source Mixture-of-Experts (MoE) reasoning model. Its advanced capabilities are cultivated through a meticulously crafted training process, beginning with long Chain-of-Thought (CoT) data cold-start and culminating in large-scale Reinforcement Learning (RL). We first employ a well-designed cold-start training strategy, which significantly enhances the reasoning potential and equips the model with specialized skills in both formal and agentic reasoning. Then, a core innovation is our domain-parallel training scheme, which decouples optimization across distinct domains (e.g., STEM, Code, Agentic) and subsequently fuses the resulting expert models into a single, nearly Pareto-optimal model. This entire process is powered by our Dynamic ORchestration for Asynchronous rollout (DORA) system, a large-scale RL framework that delivers a greater than threefold training speedup over synchronous methods on tens of thousands of accelerators. As a result, LongCat-Flash-Thinking achieves state-of-the-art performance among open-source models on a suite of complex reasoning tasks. The model exhibits exceptional efficiency in agentic reasoning, reducing average token consumption by 64.5% (from 19, 653 to 6, 965) on AIME-25, without degrading task accuracy. We release LongCat-Flash-Thinking to promote further advances in reasoning systems and agentic AI research.

Zhaohui Yang, Chenghua He, Xiaowen Shi et al.

Many studies focus on data annotation techniques for training effective PRMs. However, current methods encounter a significant issue when applied to long CoT reasoning processes: they tend to focus solely on the first incorrect step and all preceding steps, assuming that all subsequent steps are incorrect. These methods overlook the unique self-correction and reflection mechanisms inherent in long CoT, where correct reasoning steps may still occur after initial reasoning mistakes. To address this issue, we propose a novel data annotation method for PRMs specifically designed to score the long CoT reasoning process. Given that under the reflection pattern, correct and incorrect steps often alternate, we introduce the concepts of Error Propagation and Error Cessation, enhancing PRMs' ability to identify both effective self-correction behaviors and reasoning based on erroneous steps. Leveraging an LLM-based judger for annotation, we collect 1.7 million data samples to train a 7B PRM and evaluate it at both solution and step levels. Experimental results demonstrate that compared to existing open-source PRMs and PRMs trained on open-source datasets, our PRM achieves superior performance across various metrics, including search guidance, BoN, and F1 scores. Compared to widely used MC-based annotation methods, our annotation approach not only achieves higher data efficiency but also delivers superior performance. Detailed analysis is also conducted to demonstrate the stability and generalizability of our method.

Fengchuang Xing, Xiaowen Shi, Yuan-Gen Wang et al.

Unveiling the real appearance of retouched faces to prevent malicious users from deceptive advertising and economic fraud has been an increasing concern in the era of digital economics. This article makes the first attempt to investigate the face retouching reversal (FRR) problem. We first collect an FRR dataset, named deepFRR, which contains 50,000 StyleGAN-generated high-resolution (1024*1024) facial images and their corresponding retouched ones by a commercial online API. To our best knowledge, deepFRR is the first FRR dataset tailored for training the deep FRR models. Then, we propose a novel diffusion-based FRR approach (FRRffusion) for the FRR task. Our FRRffusion consists of a coarse-to-fine two-stage network: A diffusion-based Facial Morpho-Architectonic Restorer (FMAR) is constructed to generate the basic contours of low-resolution faces in the first stage, while a Transformer-based Hyperrealistic Facial Detail Generator (HFDG) is designed to create high-resolution facial details in the second stage. Tested on deepFRR, our FRRffusion surpasses the GP-UNIT and Stable Diffusion methods by a large margin in four widespread quantitative metrics. Especially, the de-retouched images by our FRRffusion are visually much closer to the raw face images than both the retouched face images and those restored by the GP-UNIT and Stable Diffusion methods in terms of qualitative evaluation with 85 subjects. These results sufficiently validate the efficacy of our work, bridging the recently-standing gap between the FRR and generic image restoration tasks. The dataset and code are available at https://github.com/GZHU-DVL/FRRffusion.

Xiaowen Shi, Yuan-Gen Wang

Underwater image enhancement (UIE) is challenging since image degradation in aquatic environments is complicated and changing over time. Existing mainstream methods rely on either physical-model or data-driven, suffering from performance bottlenecks due to changes in imaging conditions or training instability. In this article, we make the first attempt to adapt the diffusion model to the UIE task and propose a Content-Preserving Diffusion Model (CPDM) to address the above challenges. CPDM first leverages a diffusion model as its fundamental model for stable training and then designs a content-preserving framework to deal with changes in imaging conditions. Specifically, we construct a conditional input module by adopting both the raw image and the difference between the raw and noisy images as the input, which can enhance the model's adaptability by considering the changes involving the raw images in underwater environments. To preserve the essential content of the raw images, we construct a content compensation module for content-aware training by extracting low-level features from the raw images. Extensive experimental results validate the effectiveness of our CPDM, surpassing the state-of-the-art methods in terms of both subjective and objective metrics.

Meng Zhou, Bei Li, Jiahao Liu et al.

Reinforcement learning (RL) has significantly improved the reasoning ability of large language models. However, current reward models underperform in challenging reasoning scenarios and predominant RL training paradigms rely on rule-based or reference-based rewards, which impose two critical limitations: 1) the dependence on finely annotated reference answer to attain rewards; and 2) the requirement for constrained output format. These limitations fundamentally hinder further RL data scaling and sustained enhancement of model reasoning performance. To address these limitations, we propose a comprehensive framework for evaluating and improving the performance of reward models in complex reasoning scenarios. We first present a reasoning-oriented benchmark (Libra Bench), systematically constructed from a diverse collection of challenging mathematical problems and advanced reasoning models, to address the limitations of existing reward model benchmarks in reasoning scenarios. We further introduce a novel approach for improving the generative reward model via learning-to-think methodologies. Based on the proposed approach, we develop Libra-RM series, a collection of generative reward models with reasoning capabilities that achieve state-of-the-art results on various benchmarks. Comprehensive downstream experiments are conducted and the experimental results demonstrate the correlation between our Libra Bench and downstream application, and the potential of Libra-RM to further improve reasoning models with unlabeled data.

Chao Zhou, Xiaowen Shi, Yuan-Gen Wang

Recent studies have revealed that vision transformers (ViTs) face similar security risks from adversarial attacks as deep convolutional neural networks (CNNs). However, directly applying attack methodology on CNNs to ViTs has been demonstrated to be ineffective since the ViTs typically work on patch-wise encoding. This article explores the vulnerability of ViTs against adversarial attacks under a black-box scenario, and proposes a novel query-efficient hard-label adversarial attack method called AdvViT. Specifically, considering that ViTs are highly sensitive to patch modification, we propose to optimize the adversarial perturbation on the individual patches. To reduce the dimension of perturbation search space, we modify only a handful of low-frequency components of each patch. Moreover, we design a weight mask matrix for all patches to further optimize the perturbation on different regions of a whole image. We test six mainstream ViT backbones on the ImageNet-1k dataset. Experimental results show that compared with the state-of-the-art attacks on CNNs, our AdvViT achieves much lower $L_2$-norm distortion under the same query budget, sufficiently validating the vulnerability of ViTs against adversarial attacks.

Guogang Liao, Ze Wang, Xiaoxu Wu et al.

E-commerce platforms usually display a mixed list of ads and organic items in feed. One key problem is to allocate the limited slots in the feed to maximize the overall revenue as well as improve user experience, which requires a good model for user preference. Instead of modeling the influence of individual items on user behaviors, the arrangement signal models the influence of the arrangement of items and may lead to a better allocation strategy. However, most of previous strategies fail to model such a signal and therefore result in suboptimal performance. In addition, the percentage of ads exposed (PAE) is an important indicator in ads allocation. Excessive PAE hurts user experience while too low PAE reduces platform revenue. Therefore, how to constrain the PAE within a certain range while keeping personalized recommendation under the PAE constraint is a challenge. In this paper, we propose Cross Deep Q Network (Cross DQN) to extract the crucial arrangement signal by crossing the embeddings of different items and modeling the crossed sequence by multi-channel attention. Besides, we propose an auxiliary loss for batch-level constraint on PAE to tackle the above-mentioned challenge. Our model results in higher revenue and better user experience than state-of-the-art baselines in offline experiments. Moreover, our model demonstrates a significant improvement in the online A/B test and has been fully deployed on Meituan feed to serve more than 300 millions of customers.

Xiaowen Shi, Xin Li, Caili Wu et al.

Automatic analysis of highly crowded people has attracted extensive attention from computer vision research. Previous approaches for crowd counting have already achieved promising performance across various benchmarks. However, to deal with the real situation, we hope the model run as fast as possible while keeping accuracy. In this paper, we propose a compact convolutional neural network for crowd counting which learns a more efficient model with a small number of parameters. With three parallel filters executing the convolutional operation on the input image simultaneously at the front of the network, our model could achieve nearly real-time speed and save more computing resources. Experiments on two benchmarks show that our proposed method not only takes a balance between performance and efficiency which is more suitable for actual scenes but also is superior to existing light-weight models in speed.