Zhen Yang

Tencent Hunyuan Team, Ao Liu, Botong Zhou et al. · tencent-ai

As Large Language Models (LLMs) rapidly advance, we introduce Hunyuan-TurboS, a novel large hybrid Transformer-Mamba Mixture of Experts (MoE) model. It synergistically combines Mamba's long-sequence processing efficiency with Transformer's superior contextual understanding. Hunyuan-TurboS features an adaptive long-short chain-of-thought (CoT) mechanism, dynamically switching between rapid responses for simple queries and deep "thinking" modes for complex problems, optimizing computational resources. Architecturally, this 56B activated (560B total) parameter model employs 128 layers (Mamba2, Attention, FFN) with an innovative AMF/MF block pattern. Faster Mamba2 ensures linear complexity, Grouped-Query Attention minimizes KV cache, and FFNs use an MoE structure. Pre-trained on 16T high-quality tokens, it supports a 256K context length and is the first industry-deployed large-scale Mamba model. Our comprehensive post-training strategy enhances capabilities via Supervised Fine-Tuning (3M instructions), a novel Adaptive Long-short CoT Fusion method, Multi-round Deliberation Learning for iterative improvement, and a two-stage Large-scale Reinforcement Learning process targeting STEM and general instruction-following. Evaluations show strong performance: overall top 7 rank on LMSYS Chatbot Arena with a score of 1356, outperforming leading models like Gemini-2.0-Flash-001 (1352) and o4-mini-2025-04-16 (1345). TurboS also achieves an average of 77.9% across 23 automated benchmarks. Hunyuan-TurboS balances high performance and efficiency, offering substantial capabilities at lower inference costs than many reasoning models, establishing a new paradigm for efficient large-scale pre-trained models.

Ganggui Ding, Canyu Zhao, Wen Wang et al.

Benefiting from large-scale pre-trained text-to-image (T2I) generative models, impressive progress has been achieved in customized image generation, which aims to generate user-specified concepts. Existing approaches have extensively focused on single-concept customization and still encounter challenges when it comes to complex scenarios that involve combining multiple concepts. These approaches often require retraining/fine-tuning using a few images, leading to time-consuming training processes and impeding their swift implementation. Furthermore, the reliance on multiple images to represent a singular concept increases the difficulty of customization. To this end, we propose FreeCustom, a novel tuning-free method to generate customized images of multi-concept composition based on reference concepts, using only one image per concept as input. Specifically, we introduce a new multi-reference self-attention (MRSA) mechanism and a weighted mask strategy that enables the generated image to access and focus more on the reference concepts. In addition, MRSA leverages our key finding that input concepts are better preserved when providing images with context interactions. Experiments show that our method's produced images are consistent with the given concepts and better aligned with the input text. Our method outperforms or performs on par with other training-based methods in terms of multi-concept composition and single-concept customization, but is simpler. Codes can be found at https://github.com/aim-uofa/FreeCustom.

Zhen Yang, Wenhui Wang, Tao Qi et al.

Accurately recommending candidate news articles to users has always been the core challenge of news recommendation system. News recommendations often require modeling of user interest to match candidate news. Recent efforts have primarily focused on extracting local subgraph information in a global click graph constructed by the clicked news sequence of all users. Howerer, the computational complexity of extracting global click graph information has hindered the ability to utilize far-reaching linkage which is hidden between two distant nodes in global click graph collaboratively among similar users. To overcome the problem above, we propose a Global-view Long Chain Interests Modeling for news recommendation (GLoCIM), which combines neighbor interest with long chain interest distilled from a global click graph, leveraging the collaboration among similar users to enhance news recommendation. We therefore design a long chain selection algorithm and long chain interest encoder to obtain global-view long chain interest from the global click graph. We design a gated network to integrate long chain interest with neighbor interest to achieve the collaborative interest among similar users. Subsequently we aggregate it with local news category-enhanced representation to generate final user representation. Then candidate news representation can be formed to match user representation to achieve news recommendation. Experimental results on real-world datasets validate the effectiveness of our method to improve the performance of news recommendation.

Yixing Li, Ruobing Xie, Zhen Yang et al.

Transformers are the cornerstone of modern large language models, but their quadratic computational complexity limits efficiency in long-sequence processing. Recent advancements in Mamba, a state space model (SSM) with linear complexity, offer promising efficiency gains but suffer from unstable contextual learning and multitask generalization. This paper proposes TransMamba, a novel framework that unifies Transformer and Mamba through shared parameter matrices (e.g., QKV and CBx), and thus could dynamically switch between attention and SSM mechanisms at different token lengths and layers. We design the Memory converter to bridge Transformer and Mamba by converting attention outputs into SSM-compatible states, ensuring seamless information flow at TransPoints where the transformation happens. The TransPoint scheduling is also thoroughly explored for further improvements. We conducted extensive experiments demonstrating that TransMamba achieves superior training efficiency and performance compared to baselines, and validated the deeper consistency between Transformer and Mamba paradigms, offering a scalable solution for next-generation sequence modeling.

Yunliang Qi, Meng Lou, Yimin Liu et al.

Remote sensing image super-resolution (RSISR) is a crucial task in remote sensing image processing, aiming to reconstruct high-resolution (HR) images from their low-resolution (LR) counterparts. Despite the growing number of RSISR methods proposed in recent years, a systematic and comprehensive review of these methods is still lacking. This paper presents a thorough review of RSISR algorithms, covering methodologies, datasets, and evaluation metrics. We provide an in-depth analysis of RSISR methods, categorizing them into supervised, unsupervised, and quality evaluation approaches, to help researchers understand current trends and challenges. Our review also discusses the strengths, limitations, and inherent challenges of these techniques. Notably, our analysis reveals significant limitations in existing methods, particularly in preserving fine-grained textures and geometric structures under large-scale degradation. Based on these findings, we outline future research directions, highlighting the need for domain-specific architectures and robust evaluation protocols to bridge the gap between synthetic and real-world RSISR scenarios.

Guangyao Zhai, Xin Kong, Jinhao Cui et al.

Most end-to-end Multi-Object Tracking (MOT) methods face the problems of low accuracy and poor generalization ability. Although traditional filter-based methods can achieve better results, they are difficult to be endowed with optimal hyperparameters and often fail in varying scenarios. To alleviate these drawbacks, we propose a LiDAR-based 3D MOT framework named FlowMOT, which integrates point-wise motion information with the traditional matching algorithm, enhancing the robustness of the motion prediction. We firstly utilize a scene flow estimation network to obtain implicit motion information between two adjacent frames and calculate the predicted detection for each old tracklet in the previous frame. Then we use Hungarian algorithm to generate optimal matching relations with the ID propagation strategy to finish the tracking task. Experiments on KITTI MOT dataset show that our approach outperforms recent end-to-end methods and achieves competitive performance with the state-of-the-art filter-based method. In addition, ours can work steadily in the various-speed scenarios where the filter-based methods may fail.

Zhen Yang, Chi Zhang, Huiming Guo et al.

LiDAR based 3D object detectors typically need a large amount of detailed-labeled point cloud data for training, but these detailed labels are commonly expensive to acquire. In this paper, we propose a manual-label free 3D detection algorithm that leverages the CARLA simulator to generate a large amount of self-labeled training samples and introduces a novel Domain Adaptive VoxelNet (DA-VoxelNet) that can cross the distribution gap from the synthetic data to the real scenario. The self-labeled training samples are generated by a set of high quality 3D models embedded in a CARLA simulator and a proposed LiDAR-guided sampling algorithm. Then a DA-VoxelNet that integrates both a sample-level DA module and an anchor-level DA module is proposed to enable the detector trained by the synthetic data to adapt to real scenario. Experimental results show that the proposed unsupervised DA 3D detector on KITTI evaluation set can achieve 76.66% and 56.64% mAP on BEV mode and 3D mode respectively. The results reveal a promising perspective of training a LIDAR-based 3D detector without any hand-tagged label.

Yueming Yin, Zhen Yang, Haifeng Hu et al.

Unsupervised domain adaptation enables intelligent models to transfer knowledge from a labeled source domain to a similar but unlabeled target domain. Recent study reveals that knowledge can be transferred from one source domain to another unknown target domain, called Universal Domain Adaptation (UDA). However, in the real-world application, there are often more than one source domain to be exploited for domain adaptation. In this paper, we formally propose a more general domain adaptation setting, universal multi-source domain adaptation (UMDA), where the label sets of multiple source domains can be different and the label set of target domain is completely unknown. The main challenges in UMDA are to identify the common label set between each source domain and target domain, and to keep the model scalable as the number of source domains increases. To address these challenges, we propose a universal multi-source adaptation network (UMAN) to solve the domain adaptation problem without increasing the complexity of the model in various UMDA settings. In UMAN, we estimate the reliability of each known class in the common label set via the prediction margin, which helps adversarial training to better align the distributions of multiple source domains and target domain in the common label set. Moreover, the theoretical guarantee for UMAN is also provided. Massive experimental results show that existing UDA and multi-source DA (MDA) methods cannot be directly applied to UMDA and the proposed UMAN achieves the state-of-the-art performance in various UMDA settings.

Yueming Yin, Zhen Yang, Xiaofu Wu et al.

As a more practical setting for unsupervised domain adaptation, Universal Domain Adaptation (UDA) is recently introduced, where the target label set is unknown. One of the big challenges in UDA is how to determine the common label set shared by source and target domains, as there is simply no labeling available in the target domain. In this paper, we employ a probabilistic approach for locating the common label set, where each source class may come from the common label set with a probability. In particular, we propose a novel approach for evaluating the probability of each source class from the common label set, where this probability is computed by the prediction margin accumulated over the whole target domain. Then, we propose a simple universal adaptation network (S-UAN) by incorporating the probabilistic structure for the common label set. Finally, we analyse the generalization bound focusing on the common label set and explore the properties on the target risk for UDA. Extensive experiments indicate that S-UAN works well in different UDA settings and outperforms the state-of-the-art methods by large margins.

Yueming Yin, Zhen Yang, Haifeng Hu et al.

Domain alignment (DA) has been widely used in unsupervised domain adaptation. Many existing DA methods assume that a low source risk, together with the alignment of distributions of source and target, means a low target risk. In this paper, we show that this does not always hold. We thus propose a novel metric-learning-assisted domain adaptation (MLA-DA) method, which employs a novel triplet loss for helping better feature alignment. We explore the relationship between the second largest probability of a target sample's prediction and its distance to the decision boundary. Based on the relationship, we propose a novel mechanism to adaptively adjust the margin in the triplet loss according to target predictions. Experimental results show that the use of proposed triplet loss can achieve clearly better results. We also demonstrate the performance improvement of MLA-DA on all four standard benchmarks compared with the state-of-the-art unsupervised domain adaptation methods. Furthermore, MLA-DA shows stable performance in robust experiments.

Xiaofu Wu, Suofei hang, Quan Zhou et al.

Entropy minimization has been widely used in unsupervised domain adaptation (UDA). However, existing works reveal that entropy minimization only may result into collapsed trivial solutions. In this paper, we propose to avoid trivial solutions by further introducing diversity maximization. In order to achieve the possible minimum target risk for UDA, we show that diversity maximization should be elaborately balanced with entropy minimization, the degree of which can be finely controlled with the use of deep embedded validation in an unsupervised manner. The proposed minimal-entropy diversity maximization (MEDM) can be directly implemented by stochastic gradient descent without use of adversarial learning. Empirical evidence demonstrates that MEDM outperforms the state-of-the-art methods on four popular domain adaptation datasets.