Yang Du

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.

Yang Du, Yuqi Liu, Qin Jin

Cross-modal (e.g. image-text, video-text) retrieval is an important task in information retrieval and multimodal vision-language understanding field. Temporal understanding makes video-text retrieval more challenging than image-text retrieval. However, we find that the widely used video-text benchmarks have shortcomings in comprehensively assessing abilities of models, especially in temporal understanding, causing large-scale image-text pre-trained models can already achieve comparable zero-shot performance with video-text pre-trained models. In this paper, we introduce RTime, a novel temporal-emphasized video-text retrieval dataset. We first obtain videos of actions or events with significant temporality, and then reverse these videos to create harder negative samples. We then recruit annotators to judge the significance and reversibility of candidate videos, and write captions for qualified videos. We further adopt GPT-4 to extend more captions based on human-written captions. Our RTime dataset currently consists of 21k videos with 10 captions per video, totalling about 122 hours. Based on RTime, we propose three retrieval benchmark tasks: RTime-Origin, RTime-Hard, and RTime-Binary. We further enhance the use of harder-negatives in model training, and benchmark a variety of video-text models on RTime. Extensive experiment analysis proves that RTime indeed poses new and higher challenges to video-text retrieval. We release our RTime dataset https://github.com/qyr0403/Reversed-in-Time to further advance video-text retrieval and multimodal understanding research.

Yang Du, Siyuan Dai, Yonghao Song et al.

Neural visual decoding is a central problem in brain computer interface research, aiming to reconstruct human visual perception and to elucidate the structure of neural representations. However, existing approaches overlook a fundamental granularity mismatch between human and machine vision, where deep vision models emphasize semantic invariance by suppressing local texture information, whereas neural signals preserve an intricate mixture of low-level visual attributes and high-level semantic content. To address this mismatch, we propose Shallow Alignment, a novel contrastive learning strategy that aligns neural signals with intermediate representations of visual encoders rather than their final outputs, thereby striking a better balance between low-level texture details and high-level semantic features. Extensive experiments across multiple benchmarks demonstrate that Shallow Alignment significantly outperforms standard final-layer alignment, with performance gains ranging from 22% to 58% across diverse vision backbones. Notably, our approach effectively unlocks the scaling law in neural visual decoding, enabling decoding performance to scale predictably with the capacity of pre-trained vision backbones. We further conduct systematic empirical analyses to shed light on the mechanisms underlying the observed performance gains.

Bo Peng, Pi Bu, Keyu Pan et al.

Recent advances in vision-language models (VLMs) have shown promise for human-level embodied intelligence. However, existing benchmarks for VLM-driven embodied agents often rely on high-level commands or discretized action spaces, which are non-native settings that differ markedly from real-world control. In addition, current benchmarks focus primarily on high-level tasks and lack joint evaluation and analysis at both low and high levels. To address these limitations, we present NativeEmbodied, a challenging benchmark for VLM-driven embodied agents that uses a unified, native low-level action space. Built on diverse simulated scenes, NativeEmbodied includes three representative high-level tasks in complex scenarios to evaluate overall performance. For more detailed analysis, we further decouple the skills required by complex tasks and construct four types of low-level tasks, each targeting a fundamental embodied skill. This joint evaluation across task and skill granularities enables fine-grained assessment of embodied agents. Experiments with state-of-the-art VLMs reveal clear deficiencies in several fundamental embodied skills, and further analysis shows that these bottlenecks significantly limit performance on high-level tasks. NativeEmbodied highlights key challenges for current VLM-driven embodied agents and provides insights to guide future research.

Yang Du, Zhuoran Lin, Kaiqiang Song et al.

Recent advances in text-to-video (T2V) generation highlight the critical role of high-quality video-text pairs in training models capable of producing coherent and instruction-aligned videos. However, strategies for optimizing video captions specifically for T2V training remain underexplored. In this paper, we introduce VC4VG (Video Captioning for Video Generation), a comprehensive caption optimization framework tailored to the needs of T2V models. We begin by analyzing caption content from a T2V perspective, decomposing the essential elements required for video reconstruction into multiple dimensions, and proposing a principled caption design methodology. To support evaluation, we construct VC4VG-Bench, a new benchmark featuring fine-grained, multi-dimensional, and necessity-graded metrics aligned with T2V-specific requirements. Extensive T2V fine-tuning experiments demonstrate a strong correlation between improved caption quality and video generation performance, validating the effectiveness of our approach. We release all benchmark tools and code at https://github.com/alimama-creative/VC4VG to support further research.

Mao Zheng, Zheng Li, Bingxin Qu et al.

In this report, we introduce Hunyuan-MT-7B, our first open-source multilingual translation model, which supports bidirectional translation across 33 major languages and places a special emphasis on translation between Mandarin and several ethnic minority languages as well as dialects. Furthermore, to serve and address diverse translation scenarios and enhance model performance at test time, we introduce Hunyuan-MT-Chimera-7B, a translation model inspired by the slow thinking mode. This model integrates multiple outputs generated by the Hunyuan-MT-7B model under varying parameter settings, thereby achieving performance superior to that of conventional slow-thinking models based on Chain-of-Thought (CoT). The development of our models follows a holistic training process specifically engineered for multilingual translation, which begins with general and MT-oriented pre-training to build foundational capabilities, proceeds to Supervised Fine-Tuning (SFT) for task-specific adaptation, and culminates in advanced alignment through Reinforcement Learning (RL) and weak-to-strong RL. Through comprehensive experimentation, we demonstrate that both Hunyuan-MT-7B and Hunyuan-MT-Chimera-7B significantly outperform all translation-specific models of comparable parameter size and most of the SOTA large models, particularly on the task of translation between Mandarin and minority languages as well as dialects. In the WMT2025 shared task (General Machine Translation), our models demonstrate state-of-the-art performance, ranking first in 30 out of 31 language pairs. This result highlights the robustness of our models across a diverse linguistic spectrum, encompassing high-resource languages such as Chinese, English, and Japanese, as well as low-resource languages including Czech, Marathi, Estonian, and Icelandic.

Xuqing Geng, Lei Su, Zhongwei Bian et al.

Magnetic Particle Imaging (MPI) is a novel medical imaging modality. One of the established methods for MPI reconstruction is based on the System Matrix (SM). However, the calibration of the SM is often time-consuming and requires repeated measurements whenever the system parameters change. Current methodologies utilize deep learning-based super-resolution (SR) techniques to expedite SM calibration; nevertheless, these strategies do not fully exploit physical prior knowledge associated with the SM, such as symmetric positional priors. Consequently, we integrated positional priors into existing frameworks for SM calibration. Underpinned by theoretical justification, we empirically validated the efficacy of incorporating positional priors through experiments involving both 2D and 3D SM SR methods.

Ye Wang, Ziheng Wang, Boshen Xu et al.

Temporal Video Grounding (TVG), the task of locating specific video segments based on language queries, is a core challenge in long-form video understanding. While recent Large Vision-Language Models (LVLMs) have shown early promise in tackling TVG through supervised fine-tuning (SFT), their abilities to generalize remain limited. To address this, we propose a novel post-training framework that enhances the generalization capabilities of LVLMs via reinforcement learning (RL). Specifically, our contributions span three key directions: (1) Time-R1: we introduce a reasoning-guided post-training framework via RL with verifiable reward to enhance the capabilities of LVLMs on the TVG task. (2) TimeRFT: we explore data-efficient post-training strategies on our curated RL-friendly dataset, which trains the model to progressively comprehend difficult samples, leading to better generalization. (3) TVGBench: we carefully construct a small yet comprehensive benchmark for LVLM evaluation, assessing 11 types of queries and featuring balanced distributions across both videos and queries. Extensive experiments demonstrate that Time-R1 achieves state-of-the-art performance across multiple downstream datasets using only 2.5K training data, while improving its general video understanding capabilities.

Zheng Li, Yang Du, Mao Zheng et al.

Extensive research has been conducted to explore the capability of Large Language Models (LLMs) for table reasoning and has significantly improved the performance on existing benchmarks. However, tables and user questions in real-world applications are more complex and diverse, presenting an unignorable gap compared to the existing benchmarks. To fill the gap, we propose a \textbf{M}ult\textbf{i}-scale spreadsheet benchmark with \textbf{M}eta \textbf{o}perations for \textbf{Table} reasoning, named as MiMoTable. Specifically, MiMoTable incorporates two key features. First, the tables in MiMoTable are all spreadsheets used in real-world scenarios, which cover seven domains and contain different types. Second, we define a new criterion with six categories of meta operations for measuring the difficulty of each question in MiMoTable, simultaneously as a new perspective for measuring the difficulty of the existing benchmarks. Experimental results show that Claude-3.5-Sonnet achieves the best performance with 77.4\% accuracy, indicating that there is still significant room to improve for LLMs on MiMoTable. Furthermore, we grade the difficulty of existing benchmarks according to our new criteria. Experiments have shown that the performance of LLMs decreases as the difficulty of benchmarks increases, thereby proving the effectiveness of our proposed new criterion.

Yanan Guo, Wenhui Dong, Jun Song et al.

Recent advancements in video understanding within visual large language models (VLLMs) have led to notable progress. However, the complexity of video data and contextual processing limitations still hinder long-video comprehension. A common approach is video feature compression to reduce token input to large language models, yet many methods either fail to prioritize essential features, leading to redundant inter-frame information, or introduce computationally expensive modules.To address these issues, we propose FiLA(Fine-grained Vision Language Model)-Video, a novel framework that leverages a lightweight dynamic-weight multi-frame fusion strategy, which adaptively integrates multiple frames into a single representation while preserving key video information and reducing computational costs. To enhance frame selection for fusion, we introduce a keyframe selection strategy, effectively identifying informative frames from a larger pool for improved summarization. Additionally, we present a simple yet effective long-video training data generation strategy, boosting model performance without extensive manual annotation. Experimental results demonstrate that FiLA-Video achieves superior efficiency and accuracy in long-video comprehension compared to existing methods.

Yuqi Liu, Qin Jin, Tianyuan Qu et al.

Understanding accurate atomic temporal event is essential for video comprehension. However, current video-language benchmarks often fall short to evaluate Large Multi-modal Models' (LMMs) temporal event understanding capabilities, as they can be effectively addressed using image-language models. In this paper, we introduce RTime-QA, a novel benchmark specifically designed to assess the atomic temporal event understanding ability of LMMs. RTime-QA comprises 822 high-quality, carefully-curated video-text questions, each meticulously annotated by human experts. Each question features a video depicting an atomic temporal event, paired with both correct answers and temporal negative descriptions, specifically designed to evaluate temporal understanding. To advance LMMs' temporal event understanding ability, we further introduce RTime-IT, a 14k instruction-tuning dataset that employs a similar annotation process as RTime-QA. Extensive experimental analysis demonstrates that RTime-QA presents a significant challenge for LMMs: the state-of-the-art model Qwen2-VL achieves only 34.6 on strict-ACC metric, substantially lagging behind human performance. Furthermore, our experiments reveal that RTime-IT effectively enhance LMMs' capacity in temporal understanding. By fine-tuning on RTime-IT, our Qwen2-VL achieves 65.9 on RTime-QA.

Lei Su, Yang Du

We focus on tertiary lymphoid structure (TLS) semantic segmentation in whole slide image (WSI). Unlike TLS binary segmentation, TLS semantic segmentation identifies boundaries and maturity, which requires integrating contextual information to discover discriminative features. Due to the extensive scale of WSI (e.g., 100,000 \times 100,000 pixels), the segmentation of TLS is usually carried out through a patch-based strategy. However, this prevents the model from accessing information outside of the patches, limiting the performance. To address this issue, we propose GCUNet, a GNN-based contextual learning network for TLS semantic segmentation. Given an image patch (target) to be segmented, GCUNet first progressively aggregates long-range and fine-grained context outside the target. Then, a Detail and Context Fusion block (DCFusion) is designed to integrate the context and detail of the target to predict the segmentation mask. We build four TLS semantic segmentation datasets, called TCGA-COAD, TCGA-LUSC, TCGA-BLCA and INHOUSE-PAAD, and make the former three datasets (comprising 826 WSIs and 15,276 TLSs) publicly available to promote the TLS semantic segmentation. Experiments on these datasets demonstrate the superiority of GCUNet, achieving at least 7.41% improvement in mF1 compared with SOTA.

Jie Wang, Zhaoxia Yin, Jing Jiang et al.

Fooling deep neural networks (DNNs) with the black-box optimization has become a popular adversarial attack fashion, as the structural prior knowledge of DNNs is always unknown. Nevertheless, recent black-box adversarial attacks may struggle to balance their attack ability and visual quality of the generated adversarial examples (AEs) in tackling high-resolution images. In this paper, we propose an attention-guided black-box adversarial attack based on the large-scale multiobjective evolutionary optimization, termed as LMOA. By considering the spatial semantic information of images, we firstly take advantage of the attention map to determine the perturbed pixels. Instead of attacking the entire image, reducing the perturbed pixels with the attention mechanism can help to avoid the notorious curse of dimensionality and thereby improves the performance of attacking. Secondly, a large-scale multiobjective evolutionary algorithm is employed to traverse the reduced pixels in the salient region. Benefiting from its characteristics, the generated AEs have the potential to fool target DNNs while being imperceptible by the human vision. Extensive experimental results have verified the effectiveness of the proposed LMOA on the ImageNet dataset. More importantly, it is more competitive to generate high-resolution AEs with better visual quality compared with the existing black-box adversarial attacks.

Zhaoxia Yin, Hongnian Guo, Yang Du

As a branch of reversible data hiding (RDH), reversible data hiding in JEPG is particularly important. Because JPEG images are widely used, it is great significance to study reversible data hiding algorithm for JEPG images. The existing JEPG reversible data methods can be divided into two categories, one is based on Discrete Cosine Transform (DCT) coefficients modification, the other is based on Huffman table modification, the methods based on DCT coefficient modification result in large file expansion and visual quality distortion, while the methods based on entropy coding domain modification have low capacity and they may lead to large file expansion. In order to effectively solve the problems in these two kinds of methods, this paper proposes a reversible data hiding in JPEG images methods based on multi-domain modification. In this method, the secret data is divided into two parts by payload distribution algorithm, part of the secret data is first embedded in the DCT coefficient domain, and then the remaining secret data is embedded in the entropy coding domain. Experimental results demonstrate that most JPEG image files with this scheme have smaller file size increment and higher payload than previous RDH schemes.

Yang Du, Zhaoxia Yin

Code mapping (CM) is an efficient technique for reversible data hiding (RDH) in JPEG images, which embeds data by constructing a mapping relationship between the used and unused codes in the JPEG bitstream. This study presents a new framework for designing a CM-based RDH method. First, a new code mapping strategy is proposed to suppress file size expansion and improve applicability. Based on our proposed strategy, the mapped codes are redefined by creating a new Huffman table rather than selecting them from the unused codes in the original Huffman table. The critical issue of designing the CM-based RDH method, that is, constructing code mapping, is converted into a combinatorial optimization problem. This study proposes a novel CM-based RDH method that utilizes a genetic algorithm (GA). The experimental results demonstrate that the proposed method achieves a high embedding capacity with no signal distortion while suppressing file size expansion.

Yang Du, Zhaoxia Yin, Xinpeng Zhang

JPEG is the most popular image format, which is widely used in our daily life. Therefore, reversible data hiding (RDH) for JPEG images is important. Most of the RDH schemes for JPEG images will cause significant distortions and large file size increments in the marked JPEG image. As a special case of RDH, the lossless data hiding (LDH) technique can keep the visual quality of the marked images no degradation. In this paper, a novel high capacity LDH scheme is proposed. In the JPEG bitstream, not all the variable length codes (VLC) are used to encode image data. By constructing the mapping between the used and unused VLCs, the secret data can be embedded by replacing the used VLC with the unused VLC. Different from the previous schemes, our mapping strategy allows the lengths of unused and used VLCs in a mapping set to be unequal. We present some basic insights into the construction of the mapping relationship. Experimental results show that most of the JPEG images using the proposed scheme obtain smaller file size increments than previous RDH schemes. Furthermore, the proposed scheme can obtain high embedding capacity while keeping the marked JPEG image with no distortion.

Yang Du, Chunfeng Yuan, Bing Li et al.

Local features at neighboring spatial positions in feature maps have high correlation since their receptive fields are often overlapped. Self-attention usually uses the weighted sum (or other functions) with internal elements of each local feature to obtain its weight score, which ignores interactions among local features. To address this, we propose an effective interaction-aware self-attention model inspired by PCA to learn attention maps. Furthermore, since different layers in a deep network capture feature maps of different scales, we use these feature maps to construct a spatial pyramid and then utilize multi-scale information to obtain more accurate attention scores, which are used to weight the local features in all spatial positions of feature maps to calculate attention maps. Moreover, our spatial pyramid attention is unrestricted to the number of its input feature maps so it is easily extended to a spatio-temporal version. Finally, our model is embedded in general CNNs to form end-to-end attention networks for action classification. Experimental results show that our method achieves the state-of-the-art results on the UCF101, HMDB51 and untrimmed Charades.

Minglan Li, Yang Gao, Hui Wen et al.

Argument Component Boundary Detection (ACBD) is an important sub-task in argumentation mining; it aims at identifying the word sequences that constitute argument components, and is usually considered as the first sub-task in the argumentation mining pipeline. Existing ACBD methods heavily depend on task-specific knowledge, and require considerable human efforts on feature-engineering. To tackle these problems, in this work, we formulate ACBD as a sequence labeling problem and propose a variety of Recurrent Neural Network (RNN) based methods, which do not use domain specific or handcrafted features beyond the relative position of the sentence in the document. In particular, we propose a novel joint RNN model that can predict whether sentences are argumentative or not, and use the predicted results to more precisely detect the argument component boundaries. We evaluate our techniques on two corpora from two different genres; results suggest that our joint RNN model obtain the state-of-the-art performance on both datasets.

Nayyar A. Zaidi, Yang Du, Geoffrey I. Webb

Learning algorithms that learn linear models often have high representation bias on real-world problems. In this paper, we show that this representation bias can be greatly reduced by discretization. Discretization is a common procedure in machine learning that is used to convert a quantitative attribute into a qualitative one. It is often motivated by the limitation of some learners to qualitative data. Discretization loses information, as fewer distinctions between instances are possible using discretized data relative to undiscretized data. In consequence, where discretization is not essential, it might appear desirable to avoid it. However, it has been shown that discretization often substantially reduces the error of the linear generative Bayesian classifier naive Bayes. This motivates a systematic study of the effectiveness of discretizing quantitative attributes for other linear classifiers. In this work, we study the effect of discretization on the performance of linear classifiers optimizing three distinct discriminative objective functions --- logistic regression (optimizing negative log-likelihood), support vector classifiers (optimizing hinge loss) and a zero-hidden layer artificial neural network (optimizing mean-square-error). We show that discretization can greatly increase the accuracy of these linear discriminative learners by reducing their representation bias, especially on big datasets. We substantiate our claims with an empirical study on $42$ benchmark datasets.