Yin Zhao

Ailin Huang, Ang Li, Aobo Kong et al.

We introduce Step 3.5 Flash, a sparse Mixture-of-Experts (MoE) model that bridges frontier-level agentic intelligence and computational efficiency. We focus on what matters most when building agents: sharp reasoning and fast, reliable execution. Step 3.5 Flash pairs a 196B-parameter foundation with 11B active parameters for efficient inference. It is optimized with interleaved 3:1 sliding-window/full attention and Multi-Token Prediction (MTP-3) to reduce the latency and cost of multi-round agentic interactions. To reach frontier-level intelligence, we design a scalable reinforcement learning framework that combines verifiable signals with preference feedback, while remaining stable under large-scale off-policy training, enabling consistent self-improvement across mathematics, code, and tool use. Step 3.5 Flash demonstrates strong performance across agent, coding, and math tasks, achieving 85.4% on IMO-AnswerBench, 86.4% on LiveCodeBench-v6 (2024.08-2025.05), 88.2% on tau2-Bench, 69.0% on BrowseComp (with context management), and 51.0% on Terminal-Bench 2.0, comparable to frontier models such as GPT-5.2 xHigh and Gemini 3.0 Pro. By redefining the efficiency frontier, Step 3.5 Flash provides a high-density foundation for deploying sophisticated agents in real-world industrial environments.

Wenxin Xu, Hexin Jiang, Xuefeng Liang et al.

Subjective time-series regression (STR) tasks have gained increasing attention recently. However, most existing methods overlook the label distribution bias in STR data, which results in biased models. Emerging studies on imbalanced regression tasks, such as age estimation and depth estimation, hypothesize that the prior label distribution of the dataset is uniform. However, we observe that the label distributions of training and test sets in STR tasks are likely to be neither uniform nor identical. This distinct feature calls for new approaches that estimate more reasonable distributions to train a fair model. In this work, we propose Utopia Label Distribution Approximation (ULDA) for time-series data, which makes the training label distribution closer to real-world but unknown (utopia) label distribution. This would enhance the model's fairness. Specifically, ULDA first convolves the training label distribution by a Gaussian kernel. After convolution, the required sample quantity at each regression label may change. We further devise the Time-slice Normal Sampling (TNS) to generate new samples when the required sample quantity is greater than the initial sample quantity, and the Convolutional Weighted Loss (CWL) to lower the sample weight when the required sample quantity is less than the initial quantity. These two modules not only assist the model training on the approximated utopia label distribution, but also maintain the sample continuity in temporal context space. To the best of our knowledge, ULDA is the first method to address the label distribution bias in time-series data. Extensive experiments demonstrate that ULDA lifts the state-of-the-art performance on two STR tasks and three benchmark datasets.

Jiayi Wang, Lichun Zhang, Xiaoqi Zhuang et al.

Video technology is advancing toward Ultra High Definition (UHD) and High Dynamic Range (HDR), which intensifies the need for higher compression efficiency for these high-specification videos. Beyond advances in traditional codecs, neural video codecs (NVCs) have attracted significant research attention and have evolved rapidly over the past few years. The coding artifacts of NVCs often exhibit content-varying and generative characteristics, which differ from those of conventional codecs and are challenging for traditional video quality assessment (VQA) methods to capture. Therefore, VQA metrics are required to generalize across different codecs, content types, and dynamic ranges to better support video codec research and evaluation. In this paper, we propose FDIM, a feature-distance-based generic video quality metric for both traditional and neural video codecs across SDR and HDR formats. FDIM employs a hybrid architecture that integrates deep and hand-crafted features. The deep feature component learns multi-scale representations to capture distortions ranging from structural and textural fidelity degradation to high-level semantic deviations, while the hand-crafted feature component provides stable complementary cues to improve overall generalization. We trained FDIM on a large-scale subjective quality assessment dataset (DCVQA) consisting of over 16k video sequences encoded by traditional block-based hybrid video codecs and end-to-end perceptually optimized neural video codecs. Extensive experiments on ten SDR/HDR VQA datasets containing diverse, previously unseen codecs demonstrate that FDIM achieves strong generalization and high correlation with subjective assessment. The source code for FDIM and the DCVQA validation set will be released at https://github.com/MCL-ZJU/FDIM.

Congmin Zheng, Xiaoyun Mo, Xinbei Ma et al.

Reinforcement Learning (RL) has emerged as a mainstream paradigm for training Mobile GUI Agents, yet it struggles with the temporal credit assignment problem inherent in long-horizon tasks. A primary challenge lies in the trade-off between reward fidelity and density: outcome reward offers high fidelity but suffers from signal sparsity, while process reward provides dense supervision but remains prone to bias and reward hacking. To resolve this conflict, we propose the Adaptive Milestone Reward (ADMIRE) mechanism. ADMIRE constructs a verifiable, adaptive reward system by anchoring trajectory to milestones, which are dynamically distilled from successful explorations. Crucially, ADMIRE integrates an asymmetric credit assignment strategy that denoises successful trajectories and scaffolds failed trajectories. Extensive experiments demonstrate that ADMIRE consistently yields over 10% absolute improvement in success rate across different base models on AndroidWorld. Moreover, the method exhibits robust generalizability, achieving strong performance across diverse RL algorithms and heterogeneous environments such as web navigation and embodied tasks.

Yuanyi Song, Heyuan Huang, Qiqiang Lin et al.

The rapid advancement of multimodal large language models has enabled agents to operate mobile devices by directly interacting with graphical user interfaces, opening new possibilities for mobile automation. However, real-world mobile tasks are often complex and allow for multiple valid solutions. This contradicts current mobile agent evaluation standards: offline static benchmarks can only validate a single predefined "golden path", while online dynamic testing is constrained by the complexity and non-reproducibility of real devices, making both approaches inadequate for comprehensively assessing agent capabilities. To bridge the gap between offline and online evaluation and enhance testing stability, this paper introduces a novel graph-structured benchmarking framework. By modeling the finite states observed during real-device interactions, it achieves static simulation of dynamic behaviors. Building on this, we develop ColorBench, a benchmark focused on complex long-horizon tasks. It supports evaluation of multiple valid solutions, subtask completion rate statistics, and atomic-level capability analysis. ColorBench contains 175 tasks (74 single-app, 101 cross-app) with an average length of over 13 steps. Each task includes at least two correct paths and several typical error paths, enabling quasi-dynamic interaction. By evaluating ColorBench across various baselines, we discover limitations of existing models and propose improvement directions and feasible technical pathways to enhance agents' performance on complex, long-horizon problems based on experimental results. Code and data are available at: https://github.com/MadeAgents/ColorBench.

Yin Zhao, Minquan Wang, Longjun Cai

Eliminating the covariate shift cross domains is one of the common methods to deal with the issue of domain shift in visual unsupervised domain adaptation. However, current alignment methods, especially the prototype based or sample-level based methods neglect the structural properties of the underlying distribution and even break the condition of covariate shift. To relieve the limitations and conflicts, we introduce a novel concept named (virtual) mirror, which represents the equivalent sample in another domain. The equivalent sample pairs, named mirror pairs reflect the natural correspondence of the empirical distributions. Then a mirror loss, which aligns the mirror pairs cross domains, is constructed to enhance the alignment of the domains. The proposed method does not distort the internal structure of the underlying distribution. We also provide theoretical proof that the mirror samples and mirror loss have better asymptotic properties in reducing the domain shift. By applying the virtual mirror and mirror loss to the generic unsupervised domain adaptation model, we achieved consistent superior performance on several mainstream benchmarks. Code is available at https://github.com/CTI-VISION/Mirror-Sample

Ying Zhou, Xuefeng Liang, Han Chen et al.

Multimodal learning has exhibited a significant advantage in affective analysis tasks owing to the comprehensive information of various modalities, particularly the complementary information. Thus, many emerging studies focus on disentangling the modality-invariant and modality-specific representations from input data and then fusing them for prediction. However, our study shows that modality-specific representations may contain information that is irrelevant or conflicting with the tasks, which downgrades the effectiveness of learned multimodal representations. We revisit the disentanglement issue, and propose a novel triple disentanglement approach, TriDiRA, which disentangles the modality-invariant, effective modality-specific and ineffective modality-specific representations from input data. By fusing only the modality-invariant and effective modality-specific representations, TriDiRA can significantly alleviate the impact of irrelevant and conflicting information across modalities during model training. Extensive experiments conducted on four benchmark datasets demonstrate the effectiveness and generalization of our triple disentanglement, which outperforms SOTA methods.

Panqi Jia, A. Burakhan Koyuncu, Jue Mao et al.

The research on neural network (NN) based image compression has shown superior performance compared to classical compression frameworks. Unlike the hand-engineered transforms in the classical frameworks, NN-based models learn the non-linear transforms providing more compact bit representations, and achieve faster coding speed on parallel devices over their classical counterparts. Those properties evoked the attention of both scientific and industrial communities, resulting in the standardization activity JPEG-AI. The verification model for the standardization process of JPEG-AI is already in development and has surpassed the advanced VVC intra codec. To generate reconstructed images with the desired bits per pixel and assess the BD-rate performance of both the JPEG-AI verification model and VVC intra, bit rate matching is employed. However, the current state of the JPEG-AI verification model experiences significant slowdowns during bit rate matching, resulting in suboptimal performance due to an unsuitable model. The proposed methodology offers a gradual algorithmic optimization for matching bit rates, resulting in a fourfold acceleration and over 1% improvement in BD-rate at the base operation point. At the high operation point, the acceleration increases up to sixfold.

Panqi Jia, Jue Mao, Esin Koyuncu et al.

Currently, there is a high demand for neural network-based image compression codecs. These codecs employ non-linear transforms to create compact bit representations and facilitate faster coding speeds on devices compared to the hand-crafted transforms used in classical frameworks. The scientific and industrial communities are highly interested in these properties, leading to the standardization effort of JPEG-AI. The JPEG-AI verification model has been released and is currently under development for standardization. Utilizing neural networks, it can outperform the classic codec VVC intra by over 10% BD-rate operating at base operation point. Researchers attribute this success to the flexible bit distribution in the spatial domain, in contrast to VVC intra's anchor that is generated with a constant quality point. However, our study reveals that VVC intra displays a more adaptable bit distribution structure through the implementation of various block sizes. As a result of our observations, we have proposed a spatial bit allocation method to optimize the JPEG-AI verification model's bit distribution and enhance the visual quality. Furthermore, by applying the VVC bit distribution strategy, the objective performance of JPEG-AI verification mode can be further improved, resulting in a maximum gain of 0.45 dB in PSNR-Y.

Ning Li, Qiqiang Lin, Zheng Wu et al.

With the advancements in hardware, software, and large language model technologies, the interaction between humans and operating systems has evolved from the command-line interface to the rapidly emerging AI agent interactions. Building an operating system (OS) agent capable of executing user instructions and faithfully following user desires is becoming a reality. In this technical report, we present ColorAgent, an OS agent designed to engage in long-horizon, robust interactions with the environment while also enabling personalized and proactive user interaction. To enable long-horizon interactions with the environment, we enhance the model's capabilities through step-wise reinforcement learning and self-evolving training, while also developing a tailored multi-agent framework that ensures generality, consistency, and robustness. In terms of user interaction, we explore personalized user intent recognition and proactive engagement, positioning the OS agent not merely as an automation tool but as a warm, collaborative partner. We evaluate ColorAgent on the AndroidWorld and AndroidLab benchmarks, achieving success rates of 77.2% and 50.7%, respectively, establishing a new state of the art. Nonetheless, we note that current benchmarks are insufficient for a comprehensive evaluation of OS agents and propose further exploring directions in future work, particularly in the areas of evaluation paradigms, agent collaboration, and security.

Xun Gao, Yin Zhao, Jie Zhang et al.

Recognizing the emotional state of people is a basic but challenging task in video understanding. In this paper, we propose a new task in this field, named Pairwise Emotional Relationship Recognition (PERR). This task aims to recognize the emotional relationship between the two interactive characters in a given video clip. It is different from the traditional emotion and social relation recognition task. Varieties of information, consisting of character appearance, behaviors, facial emotions, dialogues, background music as well as subtitles contribute differently to the final results, which makes the task more challenging but meaningful in developing more advanced multi-modal models. To facilitate the task, we develop a new dataset called Emotional RelAtionship of inTeractiOn (ERATO) based on dramas and movies. ERATO is a large-scale multi-modal dataset for PERR task, which has 31,182 video clips, lasting about 203 video hours. Different from the existing datasets, ERATO contains interaction-centric videos with multi-shots, varied video length, and multiple modalities including visual, audio and text. As a minor contribution, we propose a baseline model composed of Synchronous Modal-Temporal Attention (SMTA) unit to fuse the multi-modal information for the PERR task. In contrast to other prevailing attention mechanisms, our proposed SMTA can steadily improve the performance by about 1\%. We expect the ERATO as well as our proposed SMTA to open up a new way for PERR task in video understanding and further improve the research of multi-modal fusion methodology.

Zhigao Fang, Jiaqi Zhang, Lu Yu et al.

We conduct a subjective experiment to compare the performance of traditional image coding methods and learning-based image coding methods. HEVC and VVC, the state-of-the-art traditional coding methods, are used as the representative traditional methods. The learning-based methods used contain not only CNN-based methods, but also a GAN-based method, all of which are advanced or typical. Single Stimuli (SS), which is also called Absolute Category Rating (ACR), is adopted as the methodology of the experiment to obtain perceptual quality of images. Additionally, we utilize some typical and frequently used objective quality metrics to evaluate the coding methods in the experiment as comparison. The experiment shows that CNN-based and GAN-based methods can perform better than traditional methods in low bit-rates. In high bit-rates, however, it is hard to verify whether CNN-based methods are superior to traditional methods. Because the GAN method does not provide models with high target bit-rates, we cannot exactly tell the performance of the GAN method in high bit-rates. Furthermore, some popular objective quality metrics have not shown the ability well to measure quality of images generated by learning-based coding methods, especially the GAN-based one.

Chaoping Tu, Yin Zhao, Longjun Cai

Person re-identification (re-ID) is a challenging task in real-world. Besides the typical application in surveillance system, re-ID also has significant values to improve the recall rate of people identification in content video (TV or Movies). However, the occlusion, shot angle variations and complicated background make it far away from application, especially in content video. In this paper we propose an entropy based semantic feature alignment model, which takes advantages of the detailed information of the human semantic feature. Considering the uncertainty of semantic segmentation, we introduce a semantic alignment with an entropy-based mask which can reduce the negative effects of mask segmentation errors. We construct a new re-ID dataset based on content videos with many cases of occlusion and body part missing, which will be released in future. Extensive studies on both existing datasets and the new dataset demonstrate the superior performance of the proposed model.

Jianbin Lin, Daixin Wang, Lu Guan et al.

Nowadays designing a real recommendation system has been a critical problem for both academic and industry. However, due to the huge number of users and items, the diversity and dynamic property of the user interest, how to design a scalable recommendation system, which is able to efficiently produce effective and diverse recommendation results on billion-scale scenarios, is still a challenging and open problem for existing methods. In this paper, given the user-item interaction graph, we propose RNE, a data-efficient Recommendation-based Network Embedding method, to give personalized and diverse items to users. Specifically, we propose a diversity- and dynamics-aware neighbor sampling method for network embedding. On the one hand, the method is able to preserve the local structure between the users and items while modeling the diversity and dynamic property of the user interest to boost the recommendation quality. On the other hand the sampling method can reduce the complexity of the whole method theoretically to make it possible for billion-scale recommendation. We also implement the designed algorithm in a distributed way to further improves its scalability. Experimentally, we deploy RNE on a recommendation scenario of Taobao, the largest E-commerce platform in China, and train it on a billion-scale user-item graph. As is shown on several online metrics on A/B testing, RNE is able to achieve both high-quality and diverse results compared with CF-based methods. We also conduct the offline experiments on Pinterest dataset comparing with several state-of-the-art recommendation methods and network embedding methods. The results demonstrate that our method is able to produce a good result while runs much faster than the baseline methods.

Jie Zhang, Yin Zhao, Longjun Cai et al.

Predicting the emotional impact of videos using machine learning is a challenging task considering the varieties of modalities, the complicated temporal contex of the video as well as the time dependency of the emotional states. Feature extraction, multi-modal fusion and temporal context fusion are crucial stages for predicting valence and arousal values in the emotional impact, but have not been successfully exploited. In this paper, we propose a comprehensive framework with novel designs of modal structure and multi-modal fusion strategy. We select the most suitable modalities for valence and arousal tasks respectively and each modal feature is extracted using the modality-specific pre-trained deep model on large generic dataset. Two-time-scale structures, one for the intra-clip and the other for the inter-clip, are proposed to capture the temporal dependency of video content and emotion states. To combine the complementary information from multiple modalities, an effective and efficient residual-based progressive training strategy is proposed. Each modality is step-wisely combined into the multi-modal model, responsible for completing the missing parts of features. With all those improvements above, our proposed prediction framework achieves better performance on the LIRIS-ACCEDE dataset with a large margin compared to the state-of-the-art.

Qiong Wu, Yong Liu, Chunyan Miao et al.

With the rapid development of recommender systems, accuracy is no longer the only golden criterion for evaluating whether the recommendation results are satisfying or not. In recent years, diversity has gained tremendous attention in recommender systems research, which has been recognized to be an important factor for improving user satisfaction. On the one hand, diversified recommendation helps increase the chance of answering ephemeral user needs. On the other hand, diversifying recommendation results can help the business improve product visibility and explore potential user interests. In this paper, we are going to review the recent advances in diversified recommendation. Specifically, we first review the various definitions of diversity and generate a taxonomy to shed light on how diversity have been modeled or measured in recommender systems. After that, we summarize the major optimization approaches to diversified recommendation from a taxonomic view. Last but not the least, we project into the future and point out trending research directions on this topic.

Yong Liu, Yinan Zhang, Qiong Wu et al.

Interactive recommendation that models the explicit interactions between users and the recommender system has attracted a lot of research attentions in recent years. Most previous interactive recommendation systems only focus on optimizing recommendation accuracy while overlooking other important aspects of recommendation quality, such as the diversity of recommendation results. In this paper, we propose a novel recommendation model, named \underline{D}iversity-promoting \underline{D}eep \underline{R}einforcement \underline{L}earning (D$^2$RL), which encourages the diversity of recommendation results in interaction recommendations. More specifically, we adopt a Determinantal Point Process (DPP) model to generate diverse, while relevant item recommendations. A personalized DPP kernel matrix is maintained for each user, which is constructed from two parts: a fixed similarity matrix capturing item-item similarity, and the relevance of items dynamically learnt through an actor-critic reinforcement learning framework. We performed extensive offline experiments as well as simulated online experiments with real world datasets to demonstrate the effectiveness of the proposed model.