Haotian Zhang

Haotian Zhang, Pengchuan Zhang, Xiaowei Hu et al. · microsoft-research

We present GLIPv2, a grounded VL understanding model, that serves both localization tasks (e.g., object detection, instance segmentation) and Vision-Language (VL) understanding tasks (e.g., VQA, image captioning). GLIPv2 elegantly unifies localization pre-training and Vision-Language Pre-training (VLP) with three pre-training tasks: phrase grounding as a VL reformulation of the detection task, region-word contrastive learning as a novel region-word level contrastive learning task, and the masked language modeling. This unification not only simplifies the previous multi-stage VLP procedure but also achieves mutual benefits between localization and understanding tasks. Experimental results show that a single GLIPv2 model (all model weights are shared) achieves near SoTA performance on various localization and understanding tasks. The model also shows (1) strong zero-shot and few-shot adaption performance on open-vocabulary object detection tasks and (2) superior grounding capability on VL understanding tasks. Code will be released at https://github.com/microsoft/GLIP.

Aditi, Niket Agarwal, Arslan Ali et al.

We introduce Cosmos 3, a family of omnimodal world models designed to jointly process and generate language, image, video, audio, and action sequences within a unified mixture-of-transformers architecture. By supporting highly flexible input-output configurations, Cosmos 3 seamlessly unifies critical modalities for Physical AI -- effectively subsuming vision-language models, video generators, world simulators, and world-action models into a single framework. Our evaluation demonstrates that Cosmos 3 establishes a new state-of-the-art across a diverse suite of understanding and generation tasks, demonstrating omnimodal world models as scalable, general-purpose backbones for embodied agents. Our post-trained Cosmos 3 models were ranked as the best open-source Text-to-Image and Image-to-Video models by Artificial Analysis, and the best policy model by RoboArena at the time the technical report was written. To accelerate open research and deployment in Physical AI, we make our code, model checkpoints, curated synthetic datasets, and evaluation benchmark available under the Linux Foundation's OpenMDW-1.1 https://openmdw.ai/license/1-1/ License at https://github.com/nvidia/cosmos}{github.com/nvidia/cosmos and https://huggingface.co/collections/nvidia/cosmos3 . The project website is available at https://research.nvidia.com/labs/cosmos-lab/cosmos3 .

Zhengfeng Lai, Haotian Zhang, Bowen Zhang et al.

Large-scale web-crawled datasets are fundamental for the success of pre-training vision-language models, such as CLIP. However, the inherent noise and potential irrelevance of web-crawled AltTexts pose challenges in achieving precise image-text alignment. Existing methods utilizing large language models (LLMs) for caption rewriting have shown promise on small, curated datasets like CC3M and CC12M. This study introduces a scalable pipeline for noisy caption rewriting. Unlike recent LLM rewriting techniques, we emphasize the incorporation of visual concepts into captions, termed as Visual-enriched Captions (VeCap). To ensure data diversity, we propose a novel mixed training scheme that optimizes the utilization of AltTexts alongside newly generated VeCap. We showcase the adaptation of this method for training CLIP on large-scale web-crawled datasets, termed VeCLIP. Employing this cost-effective pipeline, we effortlessly scale our dataset up to 300 million samples named VeCap dataset. Our results show significant advantages in image-text alignment and overall model performance. For example, VeCLIP achieves up to +25.2% gain in COCO and Flickr30k retrieval tasks under the 12M setting. For data efficiency, VeCLIP achieves +3% gain while only using 14% of the data employed in the vanilla CLIP and 11% in ALIGN. We also note the VeCap data is complementary with other well curated datasets good for zero-shot classification tasks. When combining VeCap and DFN, our model can achieve strong performance on both of image-text retrieval and zero-shot classification tasks, e.g. 83.1% accuracy@1 on ImageNet zero-shot for a H/14 model. We release the pre-trained models at https://github.com/apple/ml-veclip.

Haoxuan You, Haotian Zhang, Zhe Gan et al.

We introduce Ferret, a new Multimodal Large Language Model (MLLM) capable of understanding spatial referring of any shape or granularity within an image and accurately grounding open-vocabulary descriptions. To unify referring and grounding in the LLM paradigm, Ferret employs a novel and powerful hybrid region representation that integrates discrete coordinates and continuous features jointly to represent a region in the image. To extract the continuous features of versatile regions, we propose a spatial-aware visual sampler, adept at handling varying sparsity across different shapes. Consequently, Ferret can accept diverse region inputs, such as points, bounding boxes, and free-form shapes. To bolster the desired capability of Ferret, we curate GRIT, a comprehensive refer-and-ground instruction tuning dataset including 1.1M samples that contain rich hierarchical spatial knowledge, with 95K hard negative data to promote model robustness. The resulting model not only achieves superior performance in classical referring and grounding tasks, but also greatly outperforms existing MLLMs in region-based and localization-demanded multimodal chatting. Our evaluations also reveal a significantly improved capability of describing image details and a remarkable alleviation in object hallucination. Code and data will be available at https://github.com/apple/ml-ferret

Tianyi Xie, Haotian Zhang, Jinhyung Park et al.

Scaling humanoid loco-manipulation requires robot-compatible demonstrations across diverse objects, whole-body motions, and scene geometries, but teleoperation and motion capture are difficult to scale because each collection depends on physical setups, instrumented actors, and robot operation. We present GRAIL, a digital generation pipeline that remains fully virtual until deployment: it composes 3D assets, simulator-ready scenes, and priors from video foundation models (VFMs) to synthesize interactions without rebuilding physical environments or teleoperating the robot. Rather than reconstructing unconstrained in-the-wild videos, GRAIL starts from fully specified 3D configurations in which object geometry, camera parameters, metric scale, environment depth, and a robot-proportioned character are known before video generation and reused during reconstruction. This privileged setup better conditions 4D recovery, allowing model-based object tracking, human motion estimation, and interaction-aware optimization to reconstruct metric 4D human-object interaction (HOI) trajectories with reduced depth ambiguity and morphology mismatch. We retarget the recovered motions to a humanoid robot and train complementary task-general trackers: an object-aware latent adaptor for manipulation and a scene-aware tracker for terrain traversal. GRAIL produces over 20,000 sequences spanning pick-up, object manipulation, sitting, and terrain traversal. Using only GRAIL-generated data, we train egocentric visual policies through a sim-to-real pipeline and deploy them on a Unitree G1 humanoid, achieving 84\% real-world success on diverse object pick-up and 90\% success on stair-climbing.

Keyan Chen, Chenyang Liu, Hao Chen et al.

Leveraging the extensive training data from SA-1B, the Segment Anything Model (SAM) demonstrates remarkable generalization and zero-shot capabilities. However, as a category-agnostic instance segmentation method, SAM heavily relies on prior manual guidance, including points, boxes, and coarse-grained masks. Furthermore, its performance in remote sensing image segmentation tasks remains largely unexplored and unproven. In this paper, we aim to develop an automated instance segmentation approach for remote sensing images, based on the foundational SAM model and incorporating semantic category information. Drawing inspiration from prompt learning, we propose a method to learn the generation of appropriate prompts for SAM. This enables SAM to produce semantically discernible segmentation results for remote sensing images, a concept we have termed RSPrompter. We also propose several ongoing derivatives for instance segmentation tasks, drawing on recent advancements within the SAM community, and compare their performance with RSPrompter. Extensive experimental results, derived from the WHU building, NWPU VHR-10, and SSDD datasets, validate the effectiveness of our proposed method. The code for our method is publicly available at kychen.me/RSPrompter.

Xiao Zhu, Xinyu Zhou, Boyu Zhu et al.

Reinforcement Learning from Verifiable Rewards (RLVR) has driven recent progress in code large language models by leveraging execution-based feedback from unit tests, but its scalability is fundamentally constrained by the availability and reliability of high-quality test cases. We propose CodeScaler, a reward model designed to scale both reinforcement learning training and test-time inference for code generation. CodeScaler is trained on carefully curated preference data derived from verified code problems and incorporates syntax-aware code extraction and validity-preserving reward shaping to ensure stable and robust optimization. Across four coding benchmarks, CodeScaler consistently outperforms execution-based RL by +1.55 points on Qwen3-8B-Base and +4.23 points on Qwen3-14B-Base. By further scaling to 44K problems with additional synthetic data, CodeScaler yields +14.64 points improvement over the base model without requiring any test cases. At inference time, CodeScaler serves as an effective test-time scaling method, achieving performance comparable to unit test approaches while providing a 10-fold reduction in latency. Moreover, CodeScaler surpasses existing reward models on RM-Bench not only in the code domain (+3.3 points), but also in general and reasoning domains (+2.7 points on average).

Chuanbo Tang, Xihua Sheng, Zhuoyuan Li et al.

Video compression relies heavily on exploiting the temporal redundancy between video frames, which is usually achieved by estimating and using the motion information. The motion information is represented as optical flows in most of the existing deep video compression networks. Indeed, these networks often adopt pre-trained optical flow estimation networks for motion estimation. The optical flows, however, may be less suitable for video compression due to the following two factors. First, the optical flow estimation networks were trained to perform inter-frame prediction as accurately as possible, but the optical flows themselves may cost too many bits to encode. Second, the optical flow estimation networks were trained on synthetic data, and may not generalize well enough to real-world videos. We address the twofold limitations by enhancing the optical flows in two stages: offline and online. In the offline stage, we fine-tune a trained optical flow estimation network with the motion information provided by a traditional (non-deep) video compression scheme, e.g. H.266/VVC, as we believe the motion information of H.266/VVC achieves a better rate-distortion trade-off. In the online stage, we further optimize the latent features of the optical flows with a gradient descent-based algorithm for the video to be compressed, so as to enhance the adaptivity of the optical flows. We conduct experiments on a state-of-the-art deep video compression scheme, DCVC. Experimental results demonstrate that the proposed offline and online enhancement together achieves on average 12.8% bitrate saving on the tested videos, without increasing the model or computational complexity of the decoder side.

Heath J. LeBlanc, Haotian Zhang, Shreyas Sundaram et al.

This paper addresses the problem of resilient consensus in the presence of misbehaving nodes. Although it is typical to assume knowledge of at least some nonlocal information when studying secure and fault-tolerant consensus algorithms, this assumption is not suitable for large-scale dynamic networks. To remedy this, we emphasize the use of local strategies to deal with resilience to security breaches. We study a consensus protocol that uses only local information and we consider worst-case security breaches, where the compromised nodes have full knowledge of the network and the intentions of the other nodes. We provide necessary and sufficient conditions for the normal nodes to reach consensus despite the influence of the malicious nodes under different threat assumptions. These conditions are stated in terms of a novel graph-theoretic property referred to as network robustness.

Haotian Zhang, Shreyas Sundaram

We consider the problem of diffusing information in networks that contain malicious nodes. We assume that each normal node in the network has no knowledge of the network topology other than an upper bound on the number of malicious nodes in its neighborhood. We introduce a topological property known as r-robustness of a graph, and show that this property provides improved bounds on tolerating malicious behavior, in comparison to traditional concepts such as connectivity and minimum degree. We use this topological property to analyze the canonical problems of distributed consensus and broadcasting, and provide sufficient conditions for these operations to succeed. Finally, we provide a construction for r-robust graphs and show that the common preferential-attachment model for scale-free networks produces a robust graph.

Shuochen Liu, Junyi Zhu, Long Shu et al.

Empowering large language models with long-term memory is crucial for building agents that adapt to users' evolving needs. However, prior evaluations typically interleave preference-related dialogues with irrelevant conversations, reducing the task to needle-in-a-haystack retrieval while ignoring relationships between events that drive the evolution of user preferences. Such settings overlook a fundamental characteristic of real-world personalization: preferences emerge gradually and accumulate across interactions within noisy contexts. To bridge this gap, we introduce PERMA, a benchmark designed to evaluate persona consistency over time beyond static preference recall. Additionally, we incorporate (1) text variability and (2) linguistic alignment to simulate erratic user inputs and individual idiolects in real-world data. PERMA consists of temporally ordered interaction events spanning multiple sessions and domains, with preference-related queries inserted over time. We design both multiple-choice and interactive tasks to probe the model's understanding of persona along the interaction timeline. Experiments demonstrate that by linking related interactions, advanced memory systems can extract more precise preferences and reduce token consumption, outperforming traditional semantic retrieval of raw dialogues. Nevertheless, they still struggle to maintain a coherent persona across temporal depth and cross-domain interference, highlighting the need for more robust personalized memory management in agents. Our code and data are open-sourced at https://github.com/PolarisLiu1/PERMA.

Haotian Zhang, Chenyang Bu, Fei Liu et al.

Knowledge tracing (KT) is a fundamental task in educational data mining that mainly focuses on students' dynamic cognitive states of skills. The question-answering process of students can be regarded as a thinking process that considers the following two problems. One problem is which skills are needed to answer the question, and the other is how to use these skills in order. If a student wants to answer a question correctly, the student should not only master the set of skills involved in the question but also think and obtain the associative path on the skills graph. The nodes in the associative path refer to the skills needed and the path shows the order of using them. The associative path is referred to as the skill mode. Thus, obtaining the skill modes is the key to answering questions successfully. However, most existing KT models only focus on a set of skills, without considering the skill modes. We propose a KT model, called APGKT, that exploits skill modes. Specifically, we extract the subgraph topology of the skills involved in the question and combine the difficulty level of the skills to obtain the skill modes via encoding; then, through multi-layer recurrent neural networks, we obtain a student's higher-order cognitive states of skills, which is used to predict the student's future answering performance. Experiments on five benchmark datasets validate the effectiveness of the proposed model.

Heath J. LeBlanc, Haotian Zhang, Shreyas Sundaram et al.

In this paper, we study the continuous-time consensus problem in the presence of adversaries. The networked multi-agent system is modeled as a switched system, where the normal agents have integrator dynamics and the switching signal determines the topology of the network. We consider several models of omniscient adversaries under the assumption that at most a fraction of any normal agent's neighbors may be adversaries. Under this fractional assumption on the interaction between normal and adversary agents, we show that a novel graph theoretic metric, called fractional robustness, is useful for analyzing the network topologies under which the normal agents achieve consensus.

Haotian Zhang, Elaheh Fata, Shreyas Sundaram

We study a graph-theoretic property known as robustness, which plays a key role in certain classes of dynamics on networks (such as resilient consensus, contagion and bootstrap percolation). This property is stronger than other graph properties such as connectivity and minimum degree in that one can construct graphs with high connectivity and minimum degree but low robustness. However, we show that the notions of connectivity and robustness coincide on common random graph models for complex networks (Erdos-Renyi, geometric random, and preferential attachment graphs). More specifically, the properties share the same threshold function in the Erdos-Renyi model, and have the same values in one-dimensional geometric graphs and preferential attachment networks. This indicates that a variety of purely local diffusion dynamics will be effective at spreading information in such networks. Although graphs generated according to the above constructions are inherently robust, we also show that it is coNP-complete to determine whether any given graph is robust to a specified extent.

James Hong, Haotian Zhang, Michaël Gharbi et al.

We introduce the task of spotting temporally precise, fine-grained events in video (detecting the precise moment in time events occur). Precise spotting requires models to reason globally about the full-time scale of actions and locally to identify subtle frame-to-frame appearance and motion differences that identify events during these actions. Surprisingly, we find that top performing solutions to prior video understanding tasks such as action detection and segmentation do not simultaneously meet both requirements. In response, we propose E2E-Spot, a compact, end-to-end model that performs well on the precise spotting task and can be trained quickly on a single GPU. We demonstrate that E2E-Spot significantly outperforms recent baselines adapted from the video action detection, segmentation, and spotting literature to the precise spotting task. Finally, we contribute new annotations and splits to several fine-grained sports action datasets to make these datasets suitable for future work on precise spotting.

Wentao Yuan, Qingtian Zhu, Xiangyue Liu et al.

Recently, Implicit Neural Representations (INRs) parameterized by neural networks have emerged as a powerful and promising tool to represent different kinds of signals due to its continuous, differentiable properties, showing superiorities to classical discretized representations. However, the training of neural networks for INRs only utilizes input-output pairs, and the derivatives of the target output with respect to the input, which can be accessed in some cases, are usually ignored. In this paper, we propose a training paradigm for INRs whose target output is image pixels, to encode image derivatives in addition to image values in the neural network. Specifically, we use finite differences to approximate image derivatives. We show how the training paradigm can be leveraged to solve typical INRs problems, i.e., image regression and inverse rendering, and demonstrate this training paradigm can improve the data-efficiency and generalization capabilities of INRs. The code of our method is available at \url{https://github.com/megvii-research/Sobolev_INRs}.

Jiefeng Li, Ye Yuan, Davis Rempe et al.

Estimating global human motion from moving cameras is challenging due to the entanglement of human and camera motions. To mitigate the ambiguity, existing methods leverage learned human motion priors, which however often result in oversmoothed motions with misaligned 2D projections. To tackle this problem, we propose COIN, a control-inpainting motion diffusion prior that enables fine-grained control to disentangle human and camera motions. Although pre-trained motion diffusion models encode rich motion priors, we find it non-trivial to leverage such knowledge to guide global motion estimation from RGB videos. COIN introduces a novel control-inpainting score distillation sampling method to ensure well-aligned, consistent, and high-quality motion from the diffusion prior within a joint optimization framework. Furthermore, we introduce a new human-scene relation loss to alleviate the scale ambiguity by enforcing consistency among the humans, camera, and scene. Experiments on three challenging benchmarks demonstrate the effectiveness of COIN, which outperforms the state-of-the-art methods in terms of global human motion estimation and camera motion estimation. As an illustrative example, COIN outperforms the state-of-the-art method by 33% in world joint position error (W-MPJPE) on the RICH dataset.

Haotian Zhang, Mingfei Gao, Zhe Gan et al.

We present MM1.5, a new family of multimodal large language models (MLLMs) designed to enhance capabilities in text-rich image understanding, visual referring and grounding, and multi-image reasoning. Building upon the MM1 architecture, MM1.5 adopts a data-centric approach to model training, systematically exploring the impact of diverse data mixtures across the entire model training lifecycle. This includes high-quality OCR data and synthetic captions for continual pre-training, as well as an optimized visual instruction-tuning data mixture for supervised fine-tuning. Our models range from 1B to 30B parameters, encompassing both dense and mixture-of-experts (MoE) variants, and demonstrate that careful data curation and training strategies can yield strong performance even at small scales (1B and 3B). Additionally, we introduce two specialized variants: MM1.5-Video, designed for video understanding, and MM1.5-UI, tailored for mobile UI understanding. Through extensive empirical studies and ablations, we provide detailed insights into the training processes and decisions that inform our final designs, offering valuable guidance for future research in MLLM development.

Jonathan Ivey, Shivani Kumar, Jiayu Liu et al.

Studying and building datasets for dialogue tasks is both expensive and time-consuming due to the need to recruit, train, and collect data from study participants. In response, much recent work has sought to use large language models (LLMs) to simulate both human-human and human-LLM interactions, as they have been shown to generate convincingly human-like text in many settings. However, to what extent do LLM-based simulations \textit{actually} reflect human dialogues? In this work, we answer this question by generating a large-scale dataset of 100,000 paired LLM-LLM and human-LLM dialogues from the WildChat dataset and quantifying how well the LLM simulations align with their human counterparts. Overall, we find relatively low alignment between simulations and human interactions, demonstrating a systematic divergence along the multiple textual properties, including style and content. Further, in comparisons of English, Chinese, and Russian dialogues, we find that models perform similarly. Our results suggest that LLMs generally perform better when the human themself writes in a way that is more similar to the LLM's own style.

Zhuoyuan Li, Junqi Liao, Chuanbo Tang et al.

Image/video coding has been a remarkable research area for both academia and industry for many years. Testing datasets, especially high-quality image/video datasets are desirable for the justified evaluation of coding-related research, practical applications, and standardization activities. We put forward a test dataset namely USTC-TD, which has been successfully adopted in the practical end-to-end image/video coding challenge of the IEEE International Conference on Visual Communications and Image Processing (VCIP) in 2022 and 2023. USTC-TD contains 40 images at 4K spatial resolution and 10 video sequences at 1080p spatial resolution, featuring various content due to the diverse environmental factors (e.g. scene type, texture, motion, view) and the designed imaging factors (e.g. illumination, lens, shadow). We quantitatively evaluate USTC-TD on different image/video features (spatial, temporal, color, lightness), and compare it with the previous image/video test datasets, which verifies its excellent compensation for the shortcomings of existing datasets. We also evaluate both classic standardized and recently learned image/video coding schemes on USTC-TD using objective quality metrics (PSNR, MS-SSIM, VMAF) and subjective quality metric (MOS), providing an extensive benchmark for these evaluated schemes. Based on the characteristics and specific design of the proposed test dataset, we analyze the benchmark performance and shed light on the future research and development of image/video coding. All the data are released online: https://esakak.github.io/USTC-TD.

Hongyu Wang, Yuhan Jing, Yibing Shi et al.

Proper parameter configuration is a prerequisite for the success of Evolutionary Algorithms (EAs). While various adaptive strategies have been proposed, it remains an open question whether all control dimensions contribute equally to algorithmic scalability. To investigate this, we categorize control variables into numerical parameters (e.g., crossover and mutation rates) and structural parameters (e.g., population size and operator switching), hypothesizing that they play distinct roles. This paper presents an empirical study utilizing a dual-level Deep Reinforcement Learning (DRL) framework to decouple and analyze the impact of these two dimensions on the Traveling Salesman Problem (TSP). We employ a Recurrent PPO agent to dynamically regulate these parameters, treating the DRL model as a probe to reveal evolutionary dynamics. Experimental results confirm the effectiveness of this approach: the learned policies outperform static baselines, reducing the optimality gap by approximately 45% on the largest tested instance (rl5915). Building on this validated framework, our ablation analysis reveals a fundamental insight: while numerical tuning offers local refinement, structural plasticity is the decisive factor in preventing stagnation and facilitating escape from local optima. These findings suggest that future automated algorithm design should prioritize dynamic structural reconfiguration over fine-grained probability adjustment. To facilitate reproducibility, the source code is available at https://github.com/StarDream1314/DRLGA-TSP

Yinan Xia, Haotian Zhang, Huiming Wang

Large Reasoning Models (LRMs) have shown exceptional reasoning capabilities, but they also suffer from the issue of overthinking, often generating excessively long and redundant answers. For problems that exceed the model's capabilities, LRMs tend to exhibit the overconfidence phenomenon, generating overly short but incorrect answers, which may contribute to suboptimal performance. To address these issues, we propose Difficulty-Differentiated Policy Optimization (DDPO), an efficient reinforcement learning algorithm that optimizes simple and complex tasks separately based on the overconfidence phenomenon. Specifically, it reduces the output length for simple tasks without compromising accuracy, while for complex tasks, it expands the exploration space to improve performance. We further derive the theoretical conditions for maximizing expected accuracy, which require the length distribution to closely approximate the optimal length and be as concentrated as possible. Based on these conditions, we propose using the difficulty-level average as a well-founded reference for length optimization. Extensive experiments on both in-domain and out-of-domain benchmarks validate the superiority and effectiveness of DDPO. Compared to GRPO, DDPO reduces the average answer length by 12% while improving accuracy by 1.85% across multiple benchmarks, achieving a better trade-off between accuracy and length. The code is available at https://github.com/Yinan-Xia/DDPO.

Davis Rempe, Mathis Petrovich, Ye Yuan et al.

High-quality human motion data is becoming increasingly important for applications in robotics, simulation, and entertainment. Recent generative models offer a potential data source, enabling human motion synthesis through intuitive inputs like text prompts or kinematic constraints on poses. However, the small scale of public mocap datasets has limited the motion quality, control accuracy, and generalization of these models. In this work, we introduce Kimodo, an expressive and controllable kinematic motion diffusion model trained on 700 hours of optical motion capture data. Our model generates high-quality motions while being easily controlled through text and a comprehensive suite of kinematic constraints including full-body keyframes, sparse joint positions/rotations, 2D waypoints, and dense 2D paths. This is enabled through a carefully designed motion representation and two-stage denoiser architecture that decomposes root and body prediction to minimize motion artifacts while allowing for flexible constraint conditioning. Experiments on the large-scale mocap dataset justify key design decisions and analyze how the scaling of dataset size and model size affect performance.

Tianshi Cao, Jiawei Ren, Yuxuan Zhang et al.

Closed-loop simulation is a core component of autonomous vehicle (AV) development, enabling scalable testing, training, and safety validation before real-world deployment. Neural scene reconstruction converts driving logs into interactive 3D environments for simulation, but it does not produce complete 3D object assets required for agent manipulation and large-viewpoint novel-view synthesis. To address this challenge, we present Asset Harvester, an image-to-3D model and end-to-end pipeline that converts sparse, in-the-wild object observations from real driving logs into complete, simulation-ready assets. Rather than relying on a single model component, we developed a system-level design for real-world AV data that combines large-scale curation of object-centric training tuples, geometry-aware preprocessing across heterogeneous sensors, and a robust training recipe that couples sparse-view-conditioned multiview generation with 3D Gaussian lifting. Within this system, SparseViewDiT is explicitly designed to address limited-angle views and other real-world data challenges. Together with hybrid data curation, augmentation, and self-distillation, this system enables scalable conversion of sparse AV object observations into reusable 3D assets.

Keen You, Haotian Zhang, Eldon Schoop et al.

Recent advancements in multimodal large language models (MLLMs) have been noteworthy, yet, these general-domain MLLMs often fall short in their ability to comprehend and interact effectively with user interface (UI) screens. In this paper, we present Ferret-UI, a new MLLM tailored for enhanced understanding of mobile UI screens, equipped with referring, grounding, and reasoning capabilities. Given that UI screens typically exhibit a more elongated aspect ratio and contain smaller objects of interest (e.g., icons, texts) than natural images, we incorporate "any resolution" on top of Ferret to magnify details and leverage enhanced visual features. Specifically, each screen is divided into 2 sub-images based on the original aspect ratio (i.e., horizontal division for portrait screens and vertical division for landscape screens). Both sub-images are encoded separately before being sent to LLMs. We meticulously gather training samples from an extensive range of elementary UI tasks, such as icon recognition, find text, and widget listing. These samples are formatted for instruction-following with region annotations to facilitate precise referring and grounding. To augment the model's reasoning ability, we further compile a dataset for advanced tasks, including detailed description, perception/interaction conversations, and function inference. After training on the curated datasets, Ferret-UI exhibits outstanding comprehension of UI screens and the capability to execute open-ended instructions. For model evaluation, we establish a comprehensive benchmark encompassing all the aforementioned tasks. Ferret-UI excels not only beyond most open-source UI MLLMs, but also surpasses GPT-4V on all the elementary UI tasks.

Yikang Ding, Qingtian Zhu, Xiangyue Liu et al.

Supervised multi-view stereo (MVS) methods have achieved remarkable progress in terms of reconstruction quality, but suffer from the challenge of collecting large-scale ground-truth depth. In this paper, we propose a novel self-supervised training pipeline for MVS based on knowledge distillation, termed KD-MVS, which mainly consists of self-supervised teacher training and distillation-based student training. Specifically, the teacher model is trained in a self-supervised fashion using both photometric and featuremetric consistency. Then we distill the knowledge of the teacher model to the student model through probabilistic knowledge transferring. With the supervision of validated knowledge, the student model is able to outperform its teacher by a large margin. Extensive experiments performed on multiple datasets show our method can even outperform supervised methods.

Chenyang Liu, Keyan Chen, Bowen Chen et al.

Remote Sensing Image Change Captioning (RSICC) aims to describe surface changes between multi-temporal remote sensing images in language, including the changed object categories, locations, and dynamics of changing objects (e.g., added or disappeared). This poses challenges to spatial and temporal modeling of bi-temporal features. Despite previous methods progressing in the spatial change perception, there are still weaknesses in joint spatial-temporal modeling. To address this, in this paper, we propose a novel RSCaMa model, which achieves efficient joint spatial-temporal modeling through multiple CaMa layers, enabling iterative refinement of bi-temporal features. To achieve efficient spatial modeling, we introduce the recently popular Mamba (a state space model) with a global receptive field and linear complexity into the RSICC task and propose the Spatial Difference-aware SSM (SD-SSM), overcoming limitations of previous CNN- and Transformer-based methods in the receptive field and computational complexity. SD-SSM enhances the model's ability to capture spatial changes sharply. In terms of efficient temporal modeling, considering the potential correlation between the temporal scanning characteristics of Mamba and the temporality of the RSICC, we propose the Temporal-Traversing SSM (TT-SSM), which scans bi-temporal features in a temporal cross-wise manner, enhancing the model's temporal understanding and information interaction. Experiments validate the effectiveness of the efficient joint spatial-temporal modeling and demonstrate the outstanding performance of RSCaMa and the potential of the Mamba in the RSICC task. Additionally, we systematically compare three different language decoders, including Mamba, GPT-style decoder, and Transformer decoder, providing valuable insights for future RSICC research. The code will be available at \emph{\url{https://github.com/Chen-Yang-Liu/RSCaMa}}

Chenyang Liu, Keyan Chen, Haotian Zhang et al.

Monitoring changes in the Earth's surface is crucial for understanding natural processes and human impacts, necessitating precise and comprehensive interpretation methodologies. Remote sensing satellite imagery offers a unique perspective for monitoring these changes, leading to the emergence of remote sensing image change interpretation (RSICI) as a significant research focus. Current RSICI technology encompasses change detection and change captioning, each with its limitations in providing comprehensive interpretation. To address this, we propose an interactive Change-Agent, which can follow user instructions to achieve comprehensive change interpretation and insightful analysis, such as change detection and change captioning, change object counting, change cause analysis, etc. The Change-Agent integrates a multi-level change interpretation (MCI) model as the eyes and a large language model (LLM) as the brain. The MCI model contains two branches of pixel-level change detection and semantic-level change captioning, in which the BI-temporal Iterative Interaction (BI3) layer is proposed to enhance the model's discriminative feature representation capabilities. To support the training of the MCI model, we build the LEVIR-MCI dataset with a large number of change masks and captions of changes. Experiments demonstrate the SOTA performance of the MCI model in achieving both change detection and change description simultaneously, and highlight the promising application value of our Change-Agent in facilitating comprehensive interpretation of surface changes, which opens up a new avenue for intelligent remote sensing applications. To facilitate future research, we will make our dataset and codebase of the MCI model and Change-Agent publicly available at https://github.com/Chen-Yang-Liu/Change-Agent

Haotian Zhang, Jintu Zhang, Huifeng Zhao et al.

Conformation Generation is a fundamental problem in drug discovery and cheminformatics. And organic molecule conformation generation, particularly in vacuum and protein pocket environments, is most relevant to drug design. Recently, with the development of geometric neural networks, the data-driven schemes have been successfully applied in this field, both for molecular conformation generation (in vacuum) and binding pose generation (in protein pocket). The former beats the traditional ETKDG method, while the latter achieves similar accuracy compared with the widely used molecular docking software. Although these methods have shown promising results, some researchers have recently questioned whether deep learning (DL) methods perform better in molecular conformation generation via a parameter-free method. To our surprise, what they have designed is some kind analogous to the famous infinite monkey theorem, the monkeys that are even equipped with physics education. To discuss the feasibility of their proving, we constructed a real infinite stochastic monkey for molecular conformation generation, showing that even with a more stochastic sampler for geometry generation, the coverage of the benchmark QM-computed conformations are higher than those of most DL-based methods. By extending their physical monkey algorithm for binding pose prediction, we also discover that the successful docking rate also achieves near-best performance among existing DL-based docking models. Thus, though their conclusions are right, their proof process needs more concern.

Haotian Zhang, Li Li, Dong Liu

Learned image compression possesses a unique challenge when incorporating non-differentiable quantization into the gradient-based training of the networks. Several quantization surrogates have been proposed to fulfill the training, but they were not systematically justified from a theoretical perspective. We fill this gap by contrasting uniform scalar quantization, the most widely used category with rounding being its simplest case, and its training surrogates. In principle, we find two factors crucial: one is the discrepancy between the surrogate and rounding, leading to train-test mismatch; the other is gradient estimation risk due to the surrogate, which consists of bias and variance of the gradient estimation. Our analyses and simulations imply that there is a tradeoff between the train-test mismatch and the gradient estimation risk, and the tradeoff varies across different network structures. Motivated by these analyses, we present a method based on stochastic uniform annealing, which has an adjustable temperature coefficient to control the tradeoff. Moreover, our analyses enlighten us as to two subtle tricks: one is to set an appropriate lower bound for the variance parameter of the estimated quantized latent distribution, which effectively reduces the train-test mismatch; the other is to use zero-center quantization with partial stop-gradient, which reduces the gradient estimation variance and thus stabilize the training. Our method with the tricks is verified to outperform the existing practices of quantization surrogates on a variety of representative image compression networks.

Yusu Qian, Haotian Zhang, Yinfei Yang et al.

The remarkable advancements in Multimodal Large Language Models (MLLMs) have not rendered them immune to challenges, particularly in the context of handling deceptive information in prompts, thus producing hallucinated responses under such conditions. To quantitatively assess this vulnerability, we present MAD-Bench, a carefully curated benchmark that contains 1000 test samples divided into 5 categories, such as non-existent objects, count of objects, and spatial relationship. We provide a comprehensive analysis of popular MLLMs, ranging from GPT-4v, Reka, Gemini-Pro, to open-sourced models, such as LLaVA-NeXT and MiniCPM-Llama3. Empirically, we observe significant performance gaps between GPT-4o and other models; and previous robust instruction-tuned models are not effective on this new benchmark. While GPT-4o achieves 82.82% accuracy on MAD-Bench, the accuracy of any other model in our experiments ranges from 9% to 50%. We further propose a remedy that adds an additional paragraph to the deceptive prompts to encourage models to think twice before answering the question. Surprisingly, this simple method can even double the accuracy; however, the absolute numbers are still too low to be satisfactory. We hope MAD-Bench can serve as a valuable benchmark to stimulate further research to enhance model resilience against deceptive prompts.

Haotian Zhang, Linxiaoyi Wan

To design a drug given a biological molecule by using deep learning methods, there are many successful models published recently. People commonly used generative models to design new molecules given certain protein. LiGAN was regarded as the baseline of deep learning model which was developed on convolutional neural networks. Recently, GraphBP showed its ability to predict innovative "real" chemicals that the binding affinity outperformed with traditional molecular docking methods by using a flow-based generative model with a graph neural network and multilayer perception. However, all those methods regarded proteins as rigid bodies and only include a very small part of proteins related to binding. However, the dynamics of proteins are essential for drug binding. Based on GraphBP, we proposed to generate more solid work derived from protein data bank. The results will be evaluated by validity and binding affinity by using a computational chemistry algorithm.

Shaojie Wang, Jinghui Wang, Yinghan Cui et al.

In agentic LLM scenarios, an agent's interaction process during a single rollout often exhibits branching behaviors. Due to memory retrieval and concurrent tool executions at certain decision points, the token trajectory of one task evolves into a tree-like structure rather than a linear sequence. However, current training pipelines decompose such tree-structured trajectories into separate linear segments, treating each branch as an independent sequence. As a result, shared prefixes across these branches are repeatedly recomputed during both forward and backward passes. To address this inefficiency, we propose Tree Training, a paradigm that computes each shared prefix only once and reuses its intermediate results across related branches during both forward and backward passes, substantially improving computation efficiency in large-scale agentic training. This is achieved via (i) Tree Packing, which efficiently reuses shared computations across trajectories, and (ii) Gradient Restoration, which ensures correct gradient propagation across reused prefixes. Experiments on multiple open-source models demonstrate up to 3.9x reduction in total training time, enabling more efficient agentic LLM SFT and RL training.

Haiyi Mao, Minxue Jia, Jason Xiaotian Dou et al.

Metacells are disjoint and homogeneous groups of single-cell profiles, representing discrete and highly granular cell states. Existing metacell algorithms tend to use only one modality to infer metacells, even though single-cell multi-omics datasets profile multiple molecular modalities within the same cell. Here, we present \textbf{C}ross-M\textbf{O}dal \textbf{E}mbedding for \textbf{M}etaCell Identification (COEM), which utilizes an embedded space leveraging the information of both scATAC-seq and scRNA-seq to perform aggregation, balancing the trade-off between fine resolution and sufficient sequencing coverage. COEM outperforms the state-of-the-art method SEACells by efficiently identifying accurate and well-separated metacells across datasets with continuous and discrete cell types. Furthermore, COEM significantly improves peak-to-gene association analyses, and facilitates complex gene regulatory inference tasks.

Haozheng Luo, Haoran Dai, Shaoyang Zhang et al.

We propose OASIS, an outlier- and sink-aware technique built on inter-layer null signaling. As AttnResidual architectures introduce an additional depth-wise normalization channel, they improve inter-layer routing flexibility but also exacerbate attention sinks, activation outliers, and the resulting degradation in inference stability and quantization robustness. OASIS addresses this issue by introducing a Softmax1-based null space and coupling token-level null evidence to depth routing through an inter-layer null signal, thereby reducing sink-dominated routing and improving structural robustness. Theoretically, we show that the dual-normalization design of AttnResidual intensifies sink formation and quantization brittleness. Experimentally, we compare OASIS against five baselines on three real-world datasets and observe consistent improvements in both attention sink and post-quantization performance. Notably, OASIS achieves an average reduction of 9.26% in maximum infinity norm and 2.60% in average kurtosis across the evaluated settings, while lowering perplexity by 75.85% under W8A8 and improving GSM8K Pass@1 by 12.42% under W4A4.

Tianyu Zhang, Haotian Zhang, Yuqi Li et al.

Learned image compression (LIC) has achieved state-of-the-art rate-distortion performance, deemed promising for next-generation image compression techniques. However, pre-trained LIC models usually suffer from significant performance degradation when applied to out-of-training-domain images, implying their poor generalization capabilities. To tackle this problem, we propose a few-shot domain adaptation method for LIC by integrating plug-and-play adapters into pre-trained models. Drawing inspiration from the analogy between latent channels and frequency components, we examine domain gaps in LIC and observe that out-of-training-domain images disrupt pre-trained channel-wise decomposition. Consequently, we introduce a method for channel-wise re-allocation using convolution-based adapters and low-rank adapters, which are lightweight and compatible to mainstream LIC schemes. Extensive experiments across multiple domains and multiple representative LIC schemes demonstrate that our method significantly enhances pre-trained models, achieving comparable performance to H.266/VVC intra coding with merely 25 target-domain samples. Additionally, our method matches the performance of full-model finetune while transmitting fewer than $2\%$ of the parameters.

Haotian Zhang, Feiyue Long, Yixin Yu et al.

Multi-view image compression (MIC) aims to achieve high compression efficiency by exploiting inter-image correlations, playing a crucial role in 3D applications. As a subfield of MIC, distributed multi-view image compression (DMIC) offers performance comparable to MIC while eliminating the need for inter-view information at the encoder side. However, existing methods in DMIC typically treat all images equally, overlooking the varying degrees of correlation between different views during decoding, which leads to suboptimal coding performance. To address this limitation, we propose a novel $\textbf{OmniParallax Attention Mechanism}$ (OPAM), which is a general mechanism for explicitly modeling correlations and aligned features between arbitrary pairs of information sources. Building upon OPAM, we propose a Parallax Multi Information Fusion Module (PMIFM) to adaptively integrate information from different sources. PMIFM is incorporated into both the joint decoder and the entropy model to construct our end-to-end DMIC framework, $\textbf{ParaHydra}$. Extensive experiments demonstrate that $\textbf{ParaHydra}$ is $\textbf{the first DMIC method}$ to significantly surpass state-of-the-art MIC codecs, while maintaining low computational overhead. Performance gains become more pronounced as the number of input views increases. Compared with LDMIC, $\textbf{ParaHydra}$ achieves bitrate savings of $\textbf{19.72%}$ on WildTrack(3) and up to $\textbf{24.18%}$ on WildTrack(6), while significantly improving coding efficiency (as much as $\textbf{65}\times$ in decoding and $\textbf{34}\times$ in encoding).

Shuochen Liu, Pengfei Luo, Chao Zhang et al.

Aiming to identify precise evidence sources from visual documents, visual evidence attribution for visual document retrieval-augmented generation (VD-RAG) ensures reliable and verifiable predictions from vision-language models (VLMs) in multimodal question answering. Most existing methods adopt end-to-end training to facilitate intuitive answer verification. However, they lack fine-grained supervision and progressive traceability throughout the reasoning process. In this paper, we introduce the Chain-of-Evidence (CoE) paradigm for VD-RAG. CoE unifies Chain-of-Thought (CoT) reasoning and visual evidence attribution by grounding reference elements in reasoning steps to specific regions with bounding boxes and page indexes. To enable VLMs to generate such evidence-grounded reasoning, we propose Look As You Think (LAT), a reinforcement learning framework that trains models to produce verifiable reasoning paths with consistent attribution. During training, LAT evaluates the attribution consistency of each evidence region and provides rewards only when the CoE trajectory yields correct answers, encouraging process-level self-verification. Experiments on vanilla Qwen2.5-VL-7B-Instruct with Paper- and Wiki-VISA benchmarks show that LAT consistently improves the vanilla model in both single- and multi-image settings, yielding average gains of 8.23% in soft exact match (EM) and 47.0% in IoU@0.5. Meanwhile, LAT not only outperforms the supervised fine-tuning baseline, which is trained to directly produce answers with attribution, but also exhibits stronger generalization across domains.

Jingxuan Xu, Ken Deng, Weihao Li et al.

Evaluating large language models (LLMs) for software engineering has been limited by narrow task coverage, language bias, and insufficient alignment with real-world developer workflows. Existing benchmarks often focus on algorithmic problems or Python-centric bug fixing, leaving critical dimensions of software engineering underexplored. To address these gaps, we introduce SWE-Compass1, a comprehensive benchmark that unifies heterogeneous code-related evaluations into a structured and production-aligned framework. SWE-Compass spans 8 task types, 8 programming scenarios, and 10 programming languages, with 2000 high-quality instances curated from authentic GitHub pull requests and refined through systematic filtering and validation. We benchmark ten state-of-the-art LLMs under two agentic frameworks, SWE-Agent and Claude Code, revealing a clear hierarchy of difficulty across task types, languages, and scenarios. Moreover, by aligning evaluation with real-world developer practices, SWE-Compass provides a rigorous and reproducible foundation for diagnosing and advancing agentic coding capabilities in large language models.

Yuqi Li, Haotian Zhang, Li Li et al.

Context modeling is essential in learned image compression for accurately estimating the distribution of latents. While recent advanced methods have expanded context modeling capacity, they still struggle to efficiently exploit long-range dependency and diverse context information across different coding steps. In this paper, we introduce a novel Hierarchical Progressive Context Model (HPCM) for more efficient context information acquisition. Specifically, HPCM employs a hierarchical coding schedule to sequentially model the contextual dependencies among latents at multiple scales, which enables more efficient long-range context modeling. Furthermore, we propose a progressive context fusion mechanism that incorporates contextual information from previous coding steps into the current step, effectively exploiting diverse contextual information. Experimental results demonstrate that our method achieves state-of-the-art rate-distortion performance and strikes a better balance between compression performance and computational complexity. The code is available at https://github.com/lyq133/LIC-HPCM.

Zizheng Zhan, Ken Deng, Huaixi Tang et al.

We present Kwaipilot-AutoThink (KAT), an open-source 40B large language model developed to address the overthinking problem in reasoning-intensive tasks, where an automatic thinking training paradigm is proposed to dynamically switch between reasoning and non-reasoning modes based on task complexity. Specifically, first, we construct the dual-regime dataset based on a novel tagging pipeline and a multi-agent synthesis strategy, and then we apply Multi-Token Prediction (MTP)-enhanced knowledge distillation, enabling efficient and fine-grained reasoning transfer with minimal pretraining cost. Besides, we implement a cold-start initialization strategy that introduces mode-selection priors using majority-vote signals and intent-aware prompting. Finally, we propose Step-SRPO, a reinforcement learning algorithm that incorporates intermediate supervision into the GRPO framework, offering structured guidance over both reasoning-mode selection and response accuracy. Extensive experiments across multiple benchmarks demonstrate that KAT consistently matches or even outperforms current state-of-the-art models, including DeepSeek-R1-0528 and Qwen3-235B-A22B, across a wide range of reasoning-intensive tasks while reducing token usage. Notably, KAT outperforms all open-source models and even surpasses o3-mini on the leakage-controlled LiveCodeBench Pro. Beyond academic evaluation, KAT has been successfully deployed in Kwaipilot (i.e., Kuaishou's internal coding assistant), where it improves real-world development workflows with high accuracy, efficiency, and controllable reasoning behaviors. Moreover, we are actively training a 200B Mixture-of-Experts (MoE) model with 40B active parameters, and early results already show significant gains, further demonstrating the scalability of the AutoThink paradigm.

Yanghao Li, Rui Qian, Bowen Pan et al.

Unified multimodal Large Language Models (LLMs) that can both understand and generate visual content hold immense potential. However, existing open-source models often suffer from a performance trade-off between these capabilities. We present Manzano, a simple and scalable unified framework that substantially reduces this tension by coupling a hybrid image tokenizer with a well-curated training recipe. A single shared vision encoder feeds two lightweight adapters that produce continuous embeddings for image-to-text understanding and discrete tokens for text-to-image generation within a common semantic space. A unified autoregressive LLM predicts high-level semantics in the form of text and image tokens, with an auxiliary diffusion decoder subsequently translating the image tokens into pixels. The architecture, together with a unified training recipe over understanding and generation data, enables scalable joint learning of both capabilities. Manzano achieves state-of-the-art results among unified models, and is competitive with specialist models, particularly on text-rich evaluation. Our studies show minimal task conflicts and consistent gains from scaling model size, validating our design choice of a hybrid tokenizer.

Xiangyi Wei, Fei Wang, Haotian Zhang et al.

Chemical laboratory automation has long been constrained by rigid workflows and poor adaptability to the long-tail distribution of experimental tasks. While most automated platforms perform well on a narrow set of standardized procedures, real laboratories involve diverse, infrequent, and evolving operations that fall outside predefined protocols. This mismatch prevents existing systems from generalizing to novel reaction conditions, uncommon instrument configurations, and unexpected procedural variations. We present a multi-agent robotic platform designed to address this long-tail challenge through collaborative task decomposition, dynamic scheduling, and adaptive control. The system integrates chemical perception for real-time reaction monitoring with feedback-driven execution, enabling it to adjust actions based on evolving experimental states rather than fixed scripts. Validation via acid-base titration demonstrates autonomous progress tracking, adaptive dispensing control, and reliable end-to-end experiment execution. By improving generalization across diverse laboratory scenarios, this platform provides a practical pathway toward intelligent, flexible, and scalable laboratory automation.

Xuechen Mu, Zhenyu Huang, Kewei Li et al.

Recent advancements in feature representation and dimension reduction have highlighted their crucial role in enhancing the efficacy of predictive modeling. This work introduces TemporalPaD, a novel end-to-end deep learning framework designed for temporal pattern datasets. TemporalPaD integrates reinforcement learning (RL) with neural networks to achieve concurrent feature representation and feature reduction. The framework consists of three cooperative modules: a Policy Module, a Representation Module, and a Classification Module, structured based on the Actor-Critic (AC) framework. The Policy Module, responsible for dimensionality reduction through RL, functions as the actor, while the Representation Module for feature extraction and the Classification Module collectively serve as the critic. We comprehensively evaluate TemporalPaD using 29 UCI datasets, a well-known benchmark for validating feature reduction algorithms, through 10 independent tests and 10-fold cross-validation. Additionally, given that TemporalPaD is specifically designed for time series data, we apply it to a real-world DNA classification problem involving enhancer category and enhancer strength. The results demonstrate that TemporalPaD is an efficient and effective framework for achieving feature reduction, applicable to both structured data and sequence datasets. The source code of the proposed TemporalPaD is freely available as supplementary material to this article and at http://www.healthinformaticslab.org/supp/.

Marco Terral, Haotian Zhang, Tianyang Zhang et al.

We introduce the task of SVG extraction, which consists in translating specific visual inputs from an image into scalable vector graphics. Existing multimodal models achieve strong results when generating SVGs from clean renderings or textual descriptions, but they fall short in real-world scenarios where natural images introduce noise, clutter, and domain shifts. A central challenge in this direction is the lack of suitable benchmarks. To address this need, we introduce the WildSVG Benchmark, formed by two complementary datasets: Natural WildSVG, built from real images containing company logos paired with their SVG annotations, and Synthetic WildSVG, which blends complex SVG renderings into real scenes to simulate difficult conditions. Together, these resources provide the first foundation for systematic benchmarking SVG extraction. We benchmark state-of-the-art multimodal models and find that current approaches perform well below what is needed for reliable SVG extraction in real scenarios. Nonetheless, iterative refinement methods point to a promising path forward, and model capabilities are steadily improving

Juan Rodriguez, Haotian Zhang, Abhay Puri et al.

We introduce VectorGym, a comprehensive benchmark suite for Scalable Vector Graphics (SVG) that spans generation from text and sketches, complex editing, and visual understanding. VectorGym addresses the lack of realistic, challenging benchmarks aligned with professional design workflows. Our benchmark comprises four tasks with expert human-authored annotations: the novel Sketch2SVG task (VG-Sketch); a new SVG editing dataset (VG-Edit) featuring complex, multi-step edits with higher-order primitives; Text2SVG generation (VG-Text); and SVG captioning (VG-Cap). Unlike prior benchmarks that rely on synthetic edits, VectorGym provides gold-standard human annotations that require semantic understanding and design intent. We also propose a multi-task reinforcement learning approach that jointly optimizes across all four tasks using rendering-based rewards. Our method, built on GRPO with curriculum learning, trains a Qwen3-VL 8B model that achieves state-of-the-art performance among open-source models, surpassing much larger models including Qwen3-VL 235B and matching GPT-4o. We also introduce a VLM-as-a-Judge metric for SVG generation, validated through human correlation studies. Our evaluation of frontier VLMs reveals significant performance gaps, positioning VectorGym as a rigorous framework for advancing visual code generation. VectorGym is publicly available on huggingface.co/datasets/ServiceNow/VectorGym.

Zizheng Zhan, Ken Deng, Jinghui Wang et al.

Recent advances in large language models (LLMs) have enabled progress in agentic coding, where models autonomously reason, plan, and act within interactive software development workflows. However, bridging the gap between static text-based training and dynamic real-world agentic execution remains a core challenge. In this technical report, we present KAT-Coder, a large-scale agentic code model trained through a multi-stage curriculum encompassing Mid-Term Training, Supervised Fine-Tuning (SFT), Reinforcement Fine-Tuning (RFT), and Reinforcement-to-Deployment Adaptation. The Mid-Term stage enhances reasoning, planning, and reflection capabilities through a corpus of real software engineering data and synthetic agentic interactions. The SFT stage constructs a million-sample dataset balancing twenty programming languages, ten development contexts, and ten task archetypes. The RFT stage introduces a novel multi-ground-truth reward formulation for stable and sample-efficient policy optimization. Finally, the Reinforcement-to-Deployment phase adapts the model to production-grade IDE environments using Error-Masked SFT and Tree-Structured Trajectory Training. In summary, these stages enable KAT-Coder to achieve robust tool-use reliability, instruction alignment, and long-context reasoning, forming a deployable foundation for real-world intelligent coding agents. Our KAT series 32B model, KAT-Dev, has been open-sourced on https://huggingface.co/Kwaipilot/KAT-Dev.

Haotian Zhang, Han Guo, Keyan Chen et al.

Despite the remarkable progress achieved in remote sensing semantic change detection (SCD), two major challenges remain. At the data level, existing SCD datasets suffer from limited change categories, insufficient change types, and a lack of fine-grained class definitions, making them inadequate to fully support practical applications. At the methodological level, most current approaches underutilize change information, typically treating it as a post-processing step to enhance spatial consistency, which constrains further improvements in model performance. To address these issues, we construct a new benchmark for remote sensing SCD, LevirSCD. Focused on the Beijing area, the dataset covers 16 change categories and 210 specific change types, with more fine-grained class definitions (e.g., roads are divided into unpaved and paved roads). Furthermore, we propose a foreground-background co-guided SCD (FoBa) method, which leverages foregrounds that focus on regions of interest and backgrounds enriched with contextual information to guide the model collaboratively, thereby alleviating semantic ambiguity while enhancing its ability to detect subtle changes. Considering the requirements of bi-temporal interaction and spatial consistency in SCD, we introduce a Gated Interaction Fusion (GIF) module along with a simple consistency loss to further enhance the model's detection performance. Extensive experiments on three datasets (SECOND, JL1, and the proposed LevirSCD) demonstrate that FoBa achieves competitive results compared to current SOTA methods, with improvements of 1.48%, 3.61%, and 2.81% in the SeK metric, respectively. Our code and dataset are available at https://github.com/zmoka-zht/FoBa.

Haotian Zhang, Junting Zhou, Haowei Lin et al.

Continual Learning (CL) poses a significant challenge in Artificial Intelligence, aiming to mirror the human ability to incrementally acquire knowledge and skills. While extensive research has focused on CL within the context of classification tasks, the advent of increasingly powerful generative models necessitates the exploration of Continual Learning of Generative models (CLoG). This paper advocates for shifting the research focus from classification-based CL to CLoG. We systematically identify the unique challenges presented by CLoG compared to traditional classification-based CL. We adapt three types of existing CL methodologies, replay-based, regularization-based, and parameter-isolation-based methods to generative tasks and introduce comprehensive benchmarks for CLoG that feature great diversity and broad task coverage. Our benchmarks and results yield intriguing insights that can be valuable for developing future CLoG methods. Additionally, we will release a codebase designed to facilitate easy benchmarking and experimentation in CLoG publicly at https://github.com/linhaowei1/CLoG. We believe that shifting the research focus to CLoG will benefit the continual learning community and illuminate the path for next-generation AI-generated content (AIGC) in a lifelong learning paradigm.

Haotian Zhang, Keyan Chen, Chenyang Liu et al.

Recently, the Mamba architecture based on state space models has demonstrated remarkable performance in a series of natural language processing tasks and has been rapidly applied to remote sensing change detection (CD) tasks. However, most methods enhance the global receptive field by directly modifying the scanning mode of Mamba, neglecting the crucial role that local information plays in dense prediction tasks (e.g., binary CD). In this article, we propose a model called CDMamba, which effectively combines global and local features for handling binary CD tasks. Specifically, the Scaled Residual ConvMamba (SRCM) block is proposed to utilize the ability of Mamba to extract global features and convolution to enhance the local details to alleviate the issue that current Mamba-based methods lack detailed clues and are difficult to achieve fine detection in dense prediction tasks. Furthermore, considering the characteristics of bi-temporal feature interaction required for CD, the Adaptive Global Local Guided Fusion (AGLGF) block is proposed to dynamically facilitate the bi-temporal interaction guided by other temporal global/local features. Our intuition is that more discriminative change features can be acquired with the guidance of other temporal features. Extensive experiments on five datasets demonstrate that our proposed CDMamba is comparable to the current methods (such as the F1/IoU scores are improved by 2.10%/3.00% and 2.44%/2.91% on LEVIR+CD and CLCD, respectively). Our code is open-sourced at https://github.com/zmoka-zht/CDMamba.