Zheng Ge

Yifan Sui, Xin Huang, Hongbing Li et al.

The rapid development of GUI foundation models and mobile GUI agents has spurred numerous evaluation benchmarks, yet most rely on simulated environments or open-source applications, leaving real-world closed-source applications largely unevaluated. The core difficulty is that closed-source applications do not expose internal states, making traditional automatic verification inapplicable. To bridge this gap, we introduce AndroidDaily, a large-scale benchmark comprising 350 realistic daily-use tasks across 94 high-frequency Android applications spanning transportation, shopping, local services, entertainment, content creation, social media, and everyday utilities. To enable automatic and verifiable assessment in these opaque environments, we propose Guideline-grounded Reviewer for Automatic Diagnostic Evaluation (GRADE), a process-aware evaluator built on a three-tiered system of observable external guidelines: operational obligations, output quality, and negative constraints. GRADE tracks the agent's visual trajectory against these criteria and produces step-level diagnostic judgments, turning long-horizon, open-ended mobile interactions into verifiable evaluation without relying on hidden internal states. Experiments show that GRADE achieves 87.37\% agreement with human evaluators. The strongest model reaches a 62.0\% success rate on AndroidDaily, highlighting a substantial gap between current reasoning capabilities and practical execution in realistic mobile workflows.

Shichao Dong, Jin Wang, Renhe Ji et al.

In this paper, we analyse the generalization ability of binary classifiers for the task of deepfake detection. We find that the stumbling block to their generalization is caused by the unexpected learned identity representation on images. Termed as the Implicit Identity Leakage, this phenomenon has been qualitatively and quantitatively verified among various DNNs. Furthermore, based on such understanding, we propose a simple yet effective method named the ID-unaware Deepfake Detection Model to reduce the influence of this phenomenon. Extensive experimental results demonstrate that our method outperforms the state-of-the-art in both in-dataset and cross-dataset evaluation. The code is available at https://github.com/megvii-research/CADDM.

Yinhao Li, Han Bao, Zheng Ge et al.

Bounded by the inherent ambiguity of depth perception, contemporary camera-based 3D object detection methods fall into the performance bottleneck. Intuitively, leveraging temporal multi-view stereo (MVS) technology is the natural knowledge for tackling this ambiguity. However, traditional attempts of MVS are flawed in two aspects when applying to 3D object detection scenes: 1) The affinity measurement among all views suffers expensive computation cost; 2) It is difficult to deal with outdoor scenarios where objects are often mobile. To this end, we introduce an effective temporal stereo method to dynamically select the scale of matching candidates, enable to significantly reduce computation overhead. Going one step further, we design an iterative algorithm to update more valuable candidates, making it adaptive to moving candidates. We instantiate our proposed method to multi-view 3D detector, namely BEVStereo. BEVStereo achieves the new state-of-the-art performance (i.e., 52.5% mAP and 61.0% NDS) on the camera-only track of nuScenes dataset. Meanwhile, extensive experiments reflect our method can deal with complex outdoor scenarios better than contemporary MVS approaches. Codes have been released at https://github.com/Megvii-BaseDetection/BEVStereo.

Zekun Qi, Runpei Dong, Guofan Fan et al.

Mainstream 3D representation learning approaches are built upon contrastive or generative modeling pretext tasks, where great improvements in performance on various downstream tasks have been achieved. However, we find these two paradigms have different characteristics: (i) contrastive models are data-hungry that suffer from a representation over-fitting issue; (ii) generative models have a data filling issue that shows inferior data scaling capacity compared to contrastive models. This motivates us to learn 3D representations by sharing the merits of both paradigms, which is non-trivial due to the pattern difference between the two paradigms. In this paper, we propose Contrast with Reconstruct (ReCon) that unifies these two paradigms. ReCon is trained to learn from both generative modeling teachers and single/cross-modal contrastive teachers through ensemble distillation, where the generative student guides the contrastive student. An encoder-decoder style ReCon-block is proposed that transfers knowledge through cross attention with stop-gradient, which avoids pretraining over-fitting and pattern difference issues. ReCon achieves a new state-of-the-art in 3D representation learning, e.g., 91.26% accuracy on ScanObjectNN. Codes have been released at https://github.com/qizekun/ReCon.

Runpei Dong, Zekun Qi, Linfeng Zhang et al.

The success of deep learning heavily relies on large-scale data with comprehensive labels, which is more expensive and time-consuming to fetch in 3D compared to 2D images or natural languages. This promotes the potential of utilizing models pretrained with data more than 3D as teachers for cross-modal knowledge transferring. In this paper, we revisit masked modeling in a unified fashion of knowledge distillation, and we show that foundational Transformers pretrained with 2D images or natural languages can help self-supervised 3D representation learning through training Autoencoders as Cross-Modal Teachers (ACT). The pretrained Transformers are transferred as cross-modal 3D teachers using discrete variational autoencoding self-supervision, during which the Transformers are frozen with prompt tuning for better knowledge inheritance. The latent features encoded by the 3D teachers are used as the target of masked point modeling, wherein the dark knowledge is distilled to the 3D Transformer students as foundational geometry understanding. Our ACT pretrained 3D learner achieves state-of-the-art generalization capacity across various downstream benchmarks, e.g., 88.21% overall accuracy on ScanObjectNN. Codes have been released at https://github.com/RunpeiDong/ACT.

Runpei Dong, Chunrui Han, Yuang Peng et al. · tsinghua

This paper presents DreamLLM, a learning framework that first achieves versatile Multimodal Large Language Models (MLLMs) empowered with frequently overlooked synergy between multimodal comprehension and creation. DreamLLM operates on two fundamental principles. The first focuses on the generative modeling of both language and image posteriors by direct sampling in the raw multimodal space. This approach circumvents the limitations and information loss inherent to external feature extractors like CLIP, and a more thorough multimodal understanding is obtained. Second, DreamLLM fosters the generation of raw, interleaved documents, modeling both text and image contents, along with unstructured layouts. This allows DreamLLM to learn all conditional, marginal, and joint multimodal distributions effectively. As a result, DreamLLM is the first MLLM capable of generating free-form interleaved content. Comprehensive experiments highlight DreamLLM's superior performance as a zero-shot multimodal generalist, reaping from the enhanced learning synergy. Project page: https://dreamllm.github.io.

Haoran Wei, Chenglong Liu, Jinyue Chen et al. · tsinghua

Traditional OCR systems (OCR-1.0) are increasingly unable to meet people's usage due to the growing demand for intelligent processing of man-made optical characters. In this paper, we collectively refer to all artificial optical signals (e.g., plain texts, math/molecular formulas, tables, charts, sheet music, and even geometric shapes) as "characters" and propose the General OCR Theory along with an excellent model, namely GOT, to promote the arrival of OCR-2.0. The GOT, with 580M parameters, is a unified, elegant, and end-to-end model, consisting of a high-compression encoder and a long-contexts decoder. As an OCR-2.0 model, GOT can handle all the above "characters" under various OCR tasks. On the input side, the model supports commonly used scene- and document-style images in slice and whole-page styles. On the output side, GOT can generate plain or formatted results (markdown/tikz/smiles/kern) via an easy prompt. Besides, the model enjoys interactive OCR features, i.e., region-level recognition guided by coordinates or colors. Furthermore, we also adapt dynamic resolution and multi-page OCR technologies to GOT for better practicality. In experiments, we provide sufficient results to prove the superiority of our model.

Yinhao Li, Zheng Ge, Guanyi Yu et al.

In this research, we propose a new 3D object detector with a trustworthy depth estimation, dubbed BEVDepth, for camera-based Bird's-Eye-View (BEV) 3D object detection. Our work is based on a key observation -- depth estimation in recent approaches is surprisingly inadequate given the fact that depth is essential to camera 3D detection. Our BEVDepth resolves this by leveraging explicit depth supervision. A camera-awareness depth estimation module is also introduced to facilitate the depth predicting capability. Besides, we design a novel Depth Refinement Module to counter the side effects carried by imprecise feature unprojection. Aided by customized Efficient Voxel Pooling and multi-frame mechanism, BEVDepth achieves the new state-of-the-art 60.9% NDS on the challenging nuScenes test set while maintaining high efficiency. For the first time, the NDS score of a camera model reaches 60%.

Chunrui Han, Jinrong Yang, Jianjian Sun et al. · tsinghua

Long-term temporal fusion is a crucial but often overlooked technique in camera-based Bird's-Eye-View (BEV) 3D perception. Existing methods are mostly in a parallel manner. While parallel fusion can benefit from long-term information, it suffers from increasing computational and memory overheads as the fusion window size grows. Alternatively, BEVFormer adopts a recurrent fusion pipeline so that history information can be efficiently integrated, yet it fails to benefit from longer temporal frames. In this paper, we explore an embarrassingly simple long-term recurrent fusion strategy built upon the LSS-based methods and find it already able to enjoy the merits from both sides, i.e., rich long-term information and efficient fusion pipeline. A temporal embedding module is further proposed to improve the model's robustness against occasionally missed frames in practical scenarios. We name this simple but effective fusing pipeline VideoBEV. Experimental results on the nuScenes benchmark show that VideoBEV obtains strong performance on various camera-based 3D perception tasks, including object detection (55.4\% mAP and 62.9\% NDS), segmentation (48.6\% vehicle mIoU), tracking (54.8\% AMOTA), and motion prediction (0.80m minADE and 0.463 EPA).

Zhi Cai, Songtao Liu, Guodong Wang et al.

DETR has set up a simple end-to-end pipeline for object detection by formulating this task as a set prediction problem, showing promising potential. Despite its notable advancements, this paper identifies two key forms of misalignment within the model: classification-regression misalignment and cross-layer target misalignment. Both issues impede DETR's convergence and degrade its overall performance. To tackle both issues simultaneously, we introduce a novel loss function, termed as Align Loss, designed to resolve the discrepancy between the two tasks. Align Loss guides the optimization of DETR through a joint quality metric, strengthening the connection between classification and regression. Furthermore, it incorporates an exponential down-weighting term to facilitate a smooth transition from positive to negative samples. Align-DETR also employs many-to-one matching for supervision of intermediate layers, akin to the design of H-DETR, which enhances robustness against instability. We conducted extensive experiments, yielding highly competitive results. Notably, our method achieves a 49.3% (+0.6) AP on the H-DETR baseline with the ResNet-50 backbone. It also sets a new state-of-the-art performance, reaching 50.5% AP in the 1x setting and 51.7% AP in the 2x setting, surpassing several strong competitors. Our code is available at https://github.com/FelixCaae/AlignDETR.

Liang Zhao, En Yu, Zheng Ge et al. · tsinghua

Human-AI interactivity is a critical aspect that reflects the usability of multimodal large language models (MLLMs). However, existing end-to-end MLLMs only allow users to interact with them through language instructions, leading to the limitation of the interactive accuracy and efficiency. In this study, we present precise referring instructions that utilize diverse reference representations such as points and boxes as referring prompts to refer to the special region. This enables MLLMs to focus on the region of interest and achieve finer-grained interaction. Based on precise referring instruction, we propose ChatSpot, a unified end-to-end multimodal large language model that supports diverse forms of interactivity including mouse clicks, drag-and-drop, and drawing boxes, which provides a more flexible and seamless interactive experience. We also construct a multi-grained vision-language instruction-following dataset based on existing datasets and GPT-4 generating. Furthermore, we design a series of evaluation tasks to assess the effectiveness of region recognition and interaction. Experimental results showcase ChatSpot's promising performance.

Ailin Huang, Chengyuan Yao, Chunrui Han et al.

We present STEP3-VL-10B, a lightweight open-source foundation model designed to redefine the trade-off between compact efficiency and frontier-level multimodal intelligence. STEP3-VL-10B is realized through two strategic shifts: first, a unified, fully unfrozen pre-training strategy on 1.2T multimodal tokens that integrates a language-aligned Perception Encoder with a Qwen3-8B decoder to establish intrinsic vision-language synergy; and second, a scaled post-training pipeline featuring over 1k iterations of reinforcement learning. Crucially, we implement Parallel Coordinated Reasoning (PaCoRe) to scale test-time compute, allocating resources to scalable perceptual reasoning that explores and synthesizes diverse visual hypotheses. Consequently, despite its compact 10B footprint, STEP3-VL-10B rivals or surpasses models 10$\times$-20$\times$ larger (e.g., GLM-4.6V-106B, Qwen3-VL-235B) and top-tier proprietary flagships like Gemini 2.5 Pro and Seed-1.5-VL. Delivering best-in-class performance, it records 92.2% on MMBench and 80.11% on MMMU, while excelling in complex reasoning with 94.43% on AIME2025 and 75.95% on MathVision. We release the full model suite to provide the community with a powerful, efficient, and reproducible baseline.

Jingcheng Hu, Yinmin Zhang, Shijie Shang et al.

We introduce Parallel Coordinated Reasoning (PaCoRe), a training-and-inference framework designed to overcome a central limitation of contemporary language models: their inability to scale test-time compute (TTC) far beyond sequential reasoning under a fixed context window. PaCoRe departs from the traditional sequential paradigm by driving TTC through massive parallel exploration coordinated via a message-passing architecture in multiple rounds. Each round launches many parallel reasoning trajectories, compacts their findings into context-bounded messages, and synthesizes these messages to guide the next round and ultimately produce the final answer. Trained end-to-end with large-scale, outcome-based reinforcement learning, the model masters the synthesis abilities required by PaCoRe and scales to multi-million-token effective TTC without exceeding context limits. The approach yields strong improvements across diverse domains, and notably pushes reasoning beyond frontier systems in mathematics: an 8B model reaches 94.5% on HMMT 2025, surpassing GPT-5's 93.2% by scaling effective TTC to roughly two million tokens. We open-source model checkpoints, training data, and the full inference pipeline to accelerate follow-up work.

Hongyu Zhou, Zheng Ge, Songtao Liu et al.

To date, the most powerful semi-supervised object detectors (SS-OD) are based on pseudo-boxes, which need a sequence of post-processing with fine-tuned hyper-parameters. In this work, we propose replacing the sparse pseudo-boxes with the dense prediction as a united and straightforward form of pseudo-label. Compared to the pseudo-boxes, our Dense Pseudo-Label (DPL) does not involve any post-processing method, thus retaining richer information. We also introduce a region selection technique to highlight the key information while suppressing the noise carried by dense labels. We name our proposed SS-OD algorithm that leverages the DPL as Dense Teacher. On COCO and VOC, Dense Teacher shows superior performance under various settings compared with the pseudo-box-based methods.

Zengran Wang, Chen Min, Zheng Ge et al.

Learning accurate depth is essential to multi-view 3D object detection. Recent approaches mainly learn depth from monocular images, which confront inherent difficulties due to the ill-posed nature of monocular depth learning. Instead of using a sole monocular depth method, in this work, we propose a novel Surround-view Temporal Stereo (STS) technique that leverages the geometry correspondence between frames across time to facilitate accurate depth learning. Specifically, we regard the field of views from all cameras around the ego vehicle as a unified view, namely surroundview, and conduct temporal stereo matching on it. The resulting geometrical correspondence between different frames from STS is utilized and combined with the monocular depth to yield final depth prediction. Comprehensive experiments on nuScenes show that STS greatly boosts 3D detection ability, notably for medium and long distance objects. On BEVDepth with ResNet-50 backbone, STS improves mAP and NDS by 2.6% and 1.4%, respectively. Consistent improvements are observed when using a larger backbone and a larger image resolution, demonstrating its effectiveness

Hongyu Zhou, Zheng Ge, Zeming Li et al.

This paper proposes an efficient multi-camera to Bird's-Eye-View (BEV) view transformation method for 3D perception, dubbed MatrixVT. Existing view transformers either suffer from poor transformation efficiency or rely on device-specific operators, hindering the broad application of BEV models. In contrast, our method generates BEV features efficiently with only convolutions and matrix multiplications (MatMul). Specifically, we propose describing the BEV feature as the MatMul of image feature and a sparse Feature Transporting Matrix (FTM). A Prime Extraction module is then introduced to compress the dimension of image features and reduce FTM's sparsity. Moreover, we propose the Ring \& Ray Decomposition to replace the FTM with two matrices and reformulate our pipeline to reduce calculation further. Compared to existing methods, MatrixVT enjoys a faster speed and less memory footprint while remaining deploy-friendly. Extensive experiments on the nuScenes benchmark demonstrate that our method is highly efficient but obtains results on par with the SOTA method in object detection and map segmentation tasks

Zhuoling Li, Chunrui Han, Zheng Ge et al.

Efficiency is quite important for 3D lane detection due to practical deployment demand. In this work, we propose a simple, fast, and end-to-end detector that still maintains high detection precision. Specifically, we devise a set of fully convolutional heads based on row-wise classification. In contrast to previous counterparts, ours supports recognizing both vertical and horizontal lanes. Besides, our method is the first one to perform row-wise classification in bird-eye-view. In the heads, we split feature into multiple groups and every group of feature corresponds to a lane instance. During training, the predictions are associated with lane labels using the proposed single-win one-to-one matching to compute loss, and no post-processing operation is demanded for inference. In this way, our proposed fully convolutional detector, GroupLane, realizes end-to-end detection like DETR. Evaluated on 3 real world 3D lane benchmarks, OpenLane, Once-3DLanes, and OpenLane-Huawei, GroupLane adopting ConvNext-Base as the backbone outperforms the published state-of-the-art PersFormer by 13.6% F1 score in the OpenLane validation set. Besides, GroupLane with ResNet18 still surpasses PersFormer by 4.9% F1 score, while the inference speed is nearly 7x faster and the FLOPs is only 13.3% of it.

En Yu, Liang Zhao, Yana Wei et al.

Humans possess the remarkable ability to foresee the future to a certain extent based on present observations, a skill we term as foresight minds. However, this capability remains largely under explored within existing Multimodal Large Language Models (MLLMs), hindering their capacity to learn the fundamental principles of how things operate and the intentions behind the observed subjects. To address this issue, we introduce the integration of future modeling into the existing learning frameworks of MLLMs. By utilizing the subject trajectory, a highly structured representation of a consecutive frame sequence, as a learning objective, we aim to bridge the gap between the past and the future. We propose two innovative methods to empower MLLMs with foresight minds, Foresight Pre-Training (FPT) and Foresight Instruction-Tuning (FIT), which are inspired by the modern learning paradigm of LLMs. Specifically, FPT jointly training various tasks centered on trajectories, enabling MLLMs to learn how to attend and predict entire trajectories from a given initial observation. Then, FIT requires MLLMs to first predict trajectories of related objects and then reason about potential future events based on them. Aided by FPT and FIT, we build a novel and unified MLLM named Merlin that supports multi-images input and analysis about potential actions of multiple objects for the future reasoning. Experimental results show Merlin powerful foresight minds with impressive performance on both future reasoning and visual comprehension tasks.

Haodong Li, Jingwei Wu, Quan Sun et al.

Recent advancements in image generation models have enabled the prediction of future Graphical User Interface (GUI) states based on user instructions. However, existing benchmarks primarily focus on general domain visual fidelity, leaving the evaluation of state transitions and temporal coherence in GUI-specific contexts underexplored. To address this gap, we introduce GEBench, a comprehensive benchmark for evaluating dynamic interaction and temporal coherence in GUI generation. GEBench comprises 700 carefully curated samples spanning five task categories, covering both single-step interactions and multi-step trajectories across real-world and fictional scenarios, as well as grounding point localization. To support systematic evaluation, we propose GE-Score, a novel five-dimensional metric that assesses Goal Achievement, Interaction Logic, Content Consistency, UI Plausibility, and Visual Quality. Extensive evaluations on current models indicate that while they perform well on single-step transitions, they struggle significantly with maintaining temporal coherence and spatial grounding over longer interaction sequences. Our findings identify icon interpretation, text rendering, and localization precision as critical bottlenecks. This work provides a foundation for systematic assessment and suggests promising directions for future research toward building high-fidelity generative GUI environments. The code is available at: https://github.com/stepfun-ai/GEBench.

Dongming Wu, Fan Jia, Jiahao Chang et al.

We present the 1st-place solution of OpenLane Topology in Autonomous Driving Challenge. Considering that topology reasoning is based on centerline detection and traffic element detection, we develop a multi-stage framework for high performance. Specifically, the centerline is detected by the powerful PETRv2 detector and the popular YOLOv8 is employed to detect the traffic elements. Further, we design a simple yet effective MLP-based head for topology prediction. Our method achieves 55\% OLS on the OpenLaneV2 test set, surpassing the 2nd solution by 8 points.

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.

Hongyu Zhou, Zheng Ge, Weixin Mao et al.

Currently, detecting 3D objects in Bird's-Eye-View (BEV) is superior to other 3D detectors for autonomous driving and robotics. However, transforming image features into BEV necessitates special operators to conduct feature sampling. These operators are not supported on many edge devices, bringing extra obstacles when deploying detectors. To address this problem, we revisit the generation of BEV representation and propose detecting objects in perspective BEV -- a new BEV representation that does not require feature sampling. We demonstrate that perspective BEV features can likewise enjoy the benefits of the BEV paradigm. Moreover, the perspective BEV improves detection performance by addressing issues caused by feature sampling. We propose PersDet for high-performance object detection in perspective BEV space based on this discovery. While implementing a simple and memory-efficient structure, PersDet outperforms existing state-of-the-art monocular methods on the nuScenes benchmark, reaching 34.6% mAP and 40.8% NDS when using ResNet-50 as the backbone.

Jinrong Yang, Tiancai Wang, Zheng Ge et al.

The great progress of 3D object detectors relies on large-scale data and 3D annotations. The annotation cost for 3D bounding boxes is extremely expensive while the 2D ones are easier and cheaper to collect. In this paper, we introduce a hybrid training framework, enabling us to learn a visual 3D object detector with massive 2D (pseudo) labels, even without 3D annotations. To break through the information bottleneck of 2D clues, we explore a new perspective: Temporal 2D Supervision. We propose a temporal 2D transformation to bridge the 3D predictions with temporal 2D labels. Two steps, including homography wraping and 2D box deduction, are taken to transform the 3D predictions into 2D ones for supervision. Experiments conducted on the nuScenes dataset show strong results (nearly 90% of its fully-supervised performance) with only 25% 3D annotations. We hope our findings can provide new insights for using a large number of 2D annotations for 3D perception.

Weixin Mao, Jinrong Yang, Zheng Ge et al.

Depth perception is a crucial component of monoc-ular 3D detection tasks that typically involve ill-posed problems. In light of the success of sample mining techniques in 2D object detection, we propose a simple yet effective mining strategy for improving depth perception in 3D object detection. Concretely, we introduce a plain metric to evaluate the quality of depth predictions, which chooses the mined sample for the model. Moreover, we propose a Gradient-aware and Model-perceive Mining strategy (GMM) for depth learning, which exploits the predicted depth quality for better depth learning through easy mining. GMM is a general strategy that can be readily applied to several state-of-the-art monocular 3D detectors, improving the accuracy of depth prediction. Extensive experiments on the nuScenes dataset demonstrate that the proposed methods significantly improve the performance of 3D object detection while outperforming other state-of-the-art sample mining techniques by a considerable margin. On the nuScenes benchmark, GMM achieved the state-of-the-art (42.1% mAP and 47.3% NDS) performance in monocular object detection.

Yinhao Li, Jinrong Yang, Jianjian Sun et al.

Bounded by the inherent ambiguity of depth perception, contemporary multi-view 3D object detection methods fall into the performance bottleneck. Intuitively, leveraging temporal multi-view stereo (MVS) technology is the natural knowledge for tackling this ambiguity. However, traditional attempts of MVS has two limitations when applying to 3D object detection scenes: 1) The affinity measurement among all views suffers expensive computational cost; 2) It is difficult to deal with outdoor scenarios where objects are often mobile. To this end, we propose BEVStereo++: by introducing a dynamic temporal stereo strategy, BEVStereo++ is able to cut down the harm that is brought by introducing temporal stereo when dealing with those two scenarios. Going one step further, we apply Motion Compensation Module and long sequence Frame Fusion to BEVStereo++, which shows further performance boosting and error reduction. Without bells and whistles, BEVStereo++ achieves state-of-the-art(SOTA) on both Waymo and nuScenes dataset.

Shiyu Liu, Yucheng Han, Peng Xing et al. · tsinghua

In recent years, image editing models have witnessed remarkable and rapid development. The recent unveiling of cutting-edge multimodal models such as GPT-4o and Gemini2 Flash has introduced highly promising image editing capabilities. These models demonstrate an impressive aptitude for fulfilling a vast majority of user-driven editing requirements, marking a significant advancement in the field of image manipulation. However, there is still a large gap between the open-source algorithm with these closed-source models. Thus, in this paper, we aim to release a state-of-the-art image editing model, called Step1X-Edit, which can provide comparable performance against the closed-source models like GPT-4o and Gemini2 Flash. More specifically, we adopt the Multimodal LLM to process the reference image and the user's editing instruction. A latent embedding has been extracted and integrated with a diffusion image decoder to obtain the target image. To train the model, we build a data generation pipeline to produce a high-quality dataset. For evaluation, we develop the GEdit-Bench, a novel benchmark rooted in real-world user instructions. Experimental results on GEdit-Bench demonstrate that Step1X-Edit outperforms existing open-source baselines by a substantial margin and approaches the performance of leading proprietary models, thereby making significant contributions to the field of image editing.

Haoran Wei, Lingyu Kong, Jinyue Chen et al.

Modern Large Vision-Language Models (LVLMs) enjoy the same vision vocabulary -- CLIP, which can cover most common vision tasks. However, for some special vision task that needs dense and fine-grained vision perception, e.g., document-level OCR or chart understanding, especially in non-English scenarios, the CLIP-style vocabulary may encounter low efficiency in tokenizing the vision knowledge and even suffer out-of-vocabulary problem. Accordingly, we propose Vary, an efficient and effective method to scale up the vision vocabulary of LVLMs. The procedures of Vary are naturally divided into two folds: the generation and integration of a new vision vocabulary. In the first phase, we devise a vocabulary network along with a tiny decoder-only transformer to produce the desired vocabulary via autoregression. In the next, we scale up the vanilla vision vocabulary by merging the new one with the original one (CLIP), enabling the LVLMs can quickly garner new features. Compared to the popular BLIP-2, MiniGPT4, and LLaVA, Vary can maintain its vanilla capabilities while enjoying more excellent fine-grained perception and understanding ability. Specifically, Vary is competent in new document parsing features (OCR or markdown conversion) while achieving 78.2% ANLS in DocVQA and 36.2% in MMVet. Our code will be publicly available on the homepage.

Guoqing Ma, Haoyang Huang, Kun Yan et al.

We present Step-Video-T2V, a state-of-the-art text-to-video pre-trained model with 30B parameters and the ability to generate videos up to 204 frames in length. A deep compression Variational Autoencoder, Video-VAE, is designed for video generation tasks, achieving 16x16 spatial and 8x temporal compression ratios, while maintaining exceptional video reconstruction quality. User prompts are encoded using two bilingual text encoders to handle both English and Chinese. A DiT with 3D full attention is trained using Flow Matching and is employed to denoise input noise into latent frames. A video-based DPO approach, Video-DPO, is applied to reduce artifacts and improve the visual quality of the generated videos. We also detail our training strategies and share key observations and insights. Step-Video-T2V's performance is evaluated on a novel video generation benchmark, Step-Video-T2V-Eval, demonstrating its state-of-the-art text-to-video quality when compared with both open-source and commercial engines. Additionally, we discuss the limitations of current diffusion-based model paradigm and outline future directions for video foundation models. We make both Step-Video-T2V and Step-Video-T2V-Eval available at https://github.com/stepfun-ai/Step-Video-T2V. The online version can be accessed from https://yuewen.cn/videos as well. Our goal is to accelerate the innovation of video foundation models and empower video content creators.

Dinging Li, Yingxiu Zhao, Xinrui Cheng et al.

Spatial reasoning over three-dimensional scenes is a core capability for embodied intelligence, yet continuous model improvement remains bottlenecked by the cost of geometric annotation. The self-evolving paradigm offers a promising path, but its reliance on model consensus to construct pseudo-labels causes training to reinforce rather than correct the model's own geometric errors. We identify a property unique to 3D spatial reasoning that circumvents this limitation: ground truth is a deterministic consequence of the underlying geometry, computable exactly from point clouds and camera poses without any model involvement. Building on this insight, we present SpatialEvo, a self-evolving framework for 3D spatial reasoning, centered on the Deterministic Geometric Environment (DGE). The DGE formalizes 16 spatial reasoning task categories under explicit geometric validation rules and converts unannotated 3D scenes into zero-noise interactive oracles, replacing model consensus with objective physical feedback. A single shared-parameter policy co-evolves across questioner and solver roles under DGE constraints: the questioner generates physically valid spatial questions grounded in scene observations, while the solver derives precise answers against DGE-verified ground truth. A task-adaptive scheduler endogenously concentrates training on the model's weakest categories, producing a dynamic curriculum without manual design. Experiments across nine benchmarks demonstrate that SpatialEvo achieves the highest average score at both 3B and 7B scales, with consistent gains on spatial reasoning benchmarks and no degradation on general visual understanding.

Ailin Huang, Boyong Wu, Bruce Wang et al.

Real-time speech interaction, serving as a fundamental interface for human-machine collaboration, holds immense potential. However, current open-source models face limitations such as high costs in voice data collection, weakness in dynamic control, and limited intelligence. To address these challenges, this paper introduces Step-Audio, the first production-ready open-source solution. Key contributions include: 1) a 130B-parameter unified speech-text multi-modal model that achieves unified understanding and generation, with the Step-Audio-Chat version open-sourced; 2) a generative speech data engine that establishes an affordable voice cloning framework and produces the open-sourced lightweight Step-Audio-TTS-3B model through distillation; 3) an instruction-driven fine control system enabling dynamic adjustments across dialects, emotions, singing, and RAP; 4) an enhanced cognitive architecture augmented with tool calling and role-playing abilities to manage complex tasks effectively. Based on our new StepEval-Audio-360 evaluation benchmark, Step-Audio achieves state-of-the-art performance in human evaluations, especially in terms of instruction following. On open-source benchmarks like LLaMA Question, shows 9.3% average performance improvement, demonstrating our commitment to advancing the development of open-source multi-modal language technologies. Our code and models are available at https://github.com/stepfun-ai/Step-Audio.

Yuhong Dai, Yanlin Lai, Mitt Huang et al.

Existing web-generation benchmarks rely on text prompts or static screenshots as input. However, videos naturally convey richer signals such as interaction flow, transition timing, and motion continuity, which are essential for faithful webpage recreation. Despite this potential, video-conditioned webpage generation remains largely unexplored, with no dedicated benchmark for this task. To fill this gap, we introduce WebVR, a benchmark that evaluates whether MLLMs can faithfully recreate webpages from demonstration videos. WebVR contains 175 webpages across diverse categories, all constructed through a controlled synthesis pipeline rather than web crawling, ensuring varied and realistic demonstrations without overlap with existing online pages. We also design a fine-grained, human-aligned visual rubric that evaluates the generated webpages across multiple dimensions. Experiments on 19 models reveal substantial gaps in recreating fine-grained style and motion quality, while the rubric-based automatic evaluation achieves 96% agreement with human preferences. We release the dataset, evaluation toolkit, and baseline results to support future research on video-to-webpage generation.

Ke Zhu, Zheng Ge, Liang Zhao et al.

This paper makes the first attempt towards unsupervised preference alignment in Vision-Language Models (VLMs). We generate chosen and rejected responses with regard to the original and augmented image pairs, and conduct preference alignment with direct preference optimization. It is based on a core idea: properly designed augmentation to the image input will induce VLM to generate false but hard negative responses, which helps the model to learn from and produce more robust and powerful answers. The whole pipeline no longer hinges on supervision from GPT-4 or human involvement during alignment, and is highly efficient with few lines of code. With only 8k randomly sampled unsupervised data, it achieves 90\% relative score to GPT-4 on complex reasoning in LLaVA-Bench, and improves LLaVA-7B/13B by 6.7\%/5.6\% score on complex multi-modal benchmark MM-Vet. Visualizations shows its improved ability to align with user-intentions. A series of ablations are firmly conducted to reveal the latent mechanism of the approach, which also indicates its potential towards further scaling. Code are available in https://github.com/Kevinz-code/SeVa.

Haolong Yan, Yeqing Shen, Xin Huang et al.

With the rapid development of Large Vision Language Models, the focus of Graphical User Interface (GUI) agent tasks shifts from single-screen tasks to complex screen navigation challenges. However, real-world GUI environments, such as PC software and mobile Apps, are often complex and proprietary, making it difficult to obtain the comprehensive environment information needed for agent training and evaluation. This limitation hinders systematic investigation and benchmarking of agent navigation capabilities. To address this limitation, we introduce GUI Exploration Lab, a simulation environment engine for GUI agent navigation research that enables flexible definition and composition of screens, icons, and navigation graphs, while providing full access to environment information for comprehensive agent training and evaluation. Through extensive experiments, we find that supervised fine-tuning enables effective memorization of fundamental knowledge, serving as a crucial foundation for subsequent training. Building on this, single-turn reinforcement learning further enhances generalization to unseen scenarios. Finally, multi-turn reinforcement learning encourages the development of exploration strategies through interactive trial and error, leading to further improvements in screen navigation performance. We validate our methods on both static and interactive benchmarks, demonstrating that our findings generalize effectively to real-world scenarios. These findings demonstrate the advantages of reinforcement learning approaches in GUI navigation and offer practical guidance for building more capable and generalizable GUI agents.

Yanlin Lai, Mitt Huang, Hangyu Guo et al.

Reinforcement Learning from Human Feedback (RLHF) remains indispensable for aligning large language models (LLMs) in subjective domains. To enhance robustness, recent work shifts toward Generative Reward Models (GenRMs) that generate rationales before predicting preferences. Yet in GenRM training and evaluation, practice remains outcome-label-only, leaving reasoning quality unchecked. We show that reasoning fidelity-the consistency between a GenRM's preference decision and reference decision rationales-is highly predictive of downstream RLHF outcomes, beyond standard label accuracy. Specifically, we repurpose existing reward-model benchmarks to compute Spurious Correctness (S-Corr)-the fraction of label-correct decisions with rationales misaligned with golden judgments. Our empirical evaluation reveals substantial S-Corr even for competitive GenRMs, and higher S-Corr is associated with policy degeneration under optimization. To improve fidelity, we propose Rationale-Centric Alignment, R-Align, which augments training with gold judgments and explicitly supervises rationale alignment. R-Align reduces S-Corr on RM benchmarks and yields consistent gains in actor performance across STEM, coding, instruction following, and general tasks.

NextStep Team, Chunrui Han, Guopeng Li et al. · tsinghua

Prevailing autoregressive (AR) models for text-to-image generation either rely on heavy, computationally-intensive diffusion models to process continuous image tokens, or employ vector quantization (VQ) to obtain discrete tokens with quantization loss. In this paper, we push the autoregressive paradigm forward with NextStep-1, a 14B autoregressive model paired with a 157M flow matching head, training on discrete text tokens and continuous image tokens with next-token prediction objectives. NextStep-1 achieves state-of-the-art performance for autoregressive models in text-to-image generation tasks, exhibiting strong capabilities in high-fidelity image synthesis. Furthermore, our method shows strong performance in image editing, highlighting the power and versatility of our unified approach. To facilitate open research, we will release our code and models to the community.

Haoyang Huang, Guoqing Ma, Nan Duan et al.

We present Step-Video-TI2V, a state-of-the-art text-driven image-to-video generation model with 30B parameters, capable of generating videos up to 102 frames based on both text and image inputs. We build Step-Video-TI2V-Eval as a new benchmark for the text-driven image-to-video task and compare Step-Video-TI2V with open-source and commercial TI2V engines using this dataset. Experimental results demonstrate the state-of-the-art performance of Step-Video-TI2V in the image-to-video generation task. Both Step-Video-TI2V and Step-Video-TI2V-Eval are available at https://github.com/stepfun-ai/Step-Video-TI2V.

Yana Wei, Liang Zhao, Jianjian Sun et al. · tsinghua

The remarkable reasoning capability of large language models (LLMs) stems from cognitive behaviors that emerge through reinforcement with verifiable rewards. This work investigates how to transfer this principle to Multimodal LLMs (MLLMs) to unlock advanced visual reasoning. We introduce a two-stage paradigm built on Qwen2.5-VL-7B: a massive linguistic cold-start fine-tuning, followed by multimodal reinforcement learning (RL) spanning nearly 1,000 steps, surpassing all previous open-source efforts in scale. This pioneering work reveals three fundamental insights: 1) Behavior transfer emerges surprisingly early in cold start due to linguistic mental imagery. 2) Cold start broadly memorizes visual behaviors, while RL critically discerns and scales up effective patterns. 3) Transfer strategically favors high-utility behaviors such as visual reflection. Our resulting model, Open-Vision-Reasoner (OVR), achieves state-of-the-art performance on a suite of reasoning benchmarks, including 95.3% on MATH500, 51.8% on MathVision and 54.6% on MathVerse. We release our model, data, and training dynamics to catalyze the development of more capable, behavior-aligned multimodal reasoners.

Xiangfeng Wang, Hangyu Guo, Yanlin Lai et al.

While model-based verifiers are essential for scaling Reinforcement Learning with Verifiable Rewards (RLVR), current outcome-centric verification paradigms primarily focus on the consistency between the final result and the ground truth, often neglecting potential errors in the derivation process. This leads to assigning positive rewards to correct answers produced from incorrect derivations. To bridge this gap, we introduce PRIME, a benchmark for evaluating verifiers on Process-Outcome Alignment verification in Mathematics and Engineering. Curated from a comprehensive collection of college-level STEM problems, PRIME comprises 2,530 high-difficulty samples through a consistency-based filtering pipeline. Through extensive evaluation, we find that current verifiers frequently fail to detect derivation flaws. Furthermore, we propose a process-aware RLVR training paradigm utilizing verifiers selected via PRIME. This approach substantially outperforms the outcome-only verification baseline, achieving absolute performance gains of 8.29%, 9.12%, and 7.31% on AIME24, AIME25, and Beyond-AIME, respectively, for the Qwen3-14B-Base model. Finally, we demonstrate a strong linear correlation ($R^2 > 0.92$) between verifier accuracy on PRIME and RLVR training effectiveness, validating PRIME as a reliable predictor for verifier selection.

Zekun Qi, Runpei Dong, Shaochen Zhang et al. · berkeley

This paper presents ShapeLLM, the first 3D Multimodal Large Language Model (LLM) designed for embodied interaction, exploring a universal 3D object understanding with 3D point clouds and languages. ShapeLLM is built upon an improved 3D encoder by extending ReCon to ReCon++ that benefits from multi-view image distillation for enhanced geometry understanding. By utilizing ReCon++ as the 3D point cloud input encoder for LLMs, ShapeLLM is trained on constructed instruction-following data and tested on our newly human-curated benchmark, 3D MM-Vet. ReCon++ and ShapeLLM achieve state-of-the-art performance in 3D geometry understanding and language-unified 3D interaction tasks, such as embodied visual grounding. Project page: https://qizekun.github.io/shapellm/

Haolong Yan, Kaijun Tan, Yeqing Shen et al.

We investigate a critical yet under-explored question in Large Vision-Language Models (LVLMs): Do LVLMs genuinely comprehend interleaved image-text in the document? Existing document understanding benchmarks often assess LVLMs using question-answer formats, which are information-sparse and difficult to guarantee the coverage of long-range dependencies. To address this issue, we introduce a novel and challenging Multimodal Document Summarization Benchmark (M-DocSum-Bench), which comprises 500 high-quality arXiv papers, along with interleaved multimodal summaries aligned with human preferences. M-DocSum-Bench is a reference-based generation task and necessitates the generation of interleaved image-text summaries using provided reference images, thereby simultaneously evaluating capabilities in understanding, reasoning, localization, and summarization within complex multimodal document scenarios. To facilitate this benchmark, we develop an automated framework to construct summaries and propose a fine-grained evaluation method called M-DocEval. Moreover, we further develop a robust summarization baseline, i.e., M-DocSum-7B, by progressive two-stage training with diverse instruction and preference data. The extensive results on our M-DocSum-Bench reveal that the leading LVLMs struggle to maintain coherence and accurately integrate information within long and interleaved contexts, often exhibiting confusion between similar images and a lack of robustness. Notably, M-DocSum-7B achieves state-of-the-art performance compared to larger and closed-source models (including GPT-4o, Gemini Pro, Claude-3.5-Sonnet and Qwen2.5-VL-72B, etc.), demonstrating the potential of LVLMs for improved interleaved image-text understanding. The code, data, and models are available at https://github.com/stepfun-ai/M-DocSum-Bench.

Songyang Zhang, Lin Song, Songtao Liu et al.

In this report, we introduce our real-time 2D object detection system for the realistic autonomous driving scenario. Our detector is built on a newly designed YOLO model, called YOLOX. On the Argoverse-HD dataset, our system achieves 41.0 streaming AP, which surpassed second place by 7.8/6.1 on detection-only track/fully track, respectively. Moreover, equipped with TensorRT, our model achieves the 30FPS inference speed with a high-resolution input size (e.g., 1440-2304). Code and models will be available at https://github.com/Megvii-BaseDetection/YOLOX

Zheng Ge, Songtao Liu, Feng Wang et al.

In this report, we present some experienced improvements to YOLO series, forming a new high-performance detector -- YOLOX. We switch the YOLO detector to an anchor-free manner and conduct other advanced detection techniques, i.e., a decoupled head and the leading label assignment strategy SimOTA to achieve state-of-the-art results across a large scale range of models: For YOLO-Nano with only 0.91M parameters and 1.08G FLOPs, we get 25.3% AP on COCO, surpassing NanoDet by 1.8% AP; for YOLOv3, one of the most widely used detectors in industry, we boost it to 47.3% AP on COCO, outperforming the current best practice by 3.0% AP; for YOLOX-L with roughly the same amount of parameters as YOLOv4-CSP, YOLOv5-L, we achieve 50.0% AP on COCO at a speed of 68.9 FPS on Tesla V100, exceeding YOLOv5-L by 1.8% AP. Further, we won the 1st Place on Streaming Perception Challenge (Workshop on Autonomous Driving at CVPR 2021) using a single YOLOX-L model. We hope this report can provide useful experience for developers and researchers in practical scenes, and we also provide deploy versions with ONNX, TensorRT, NCNN, and Openvino supported. Source code is at https://github.com/Megvii-BaseDetection/YOLOX.

Zheng Ge, Songtao Liu, Zeming Li et al.

Recent advances in label assignment in object detection mainly seek to independently define positive/negative training samples for each ground-truth (gt) object. In this paper, we innovatively revisit the label assignment from a global perspective and propose to formulate the assigning procedure as an Optimal Transport (OT) problem -- a well-studied topic in Optimization Theory. Concretely, we define the unit transportation cost between each demander (anchor) and supplier (gt) pair as the weighted summation of their classification and regression losses. After formulation, finding the best assignment solution is converted to solve the optimal transport plan at minimal transportation costs, which can be solved via Sinkhorn-Knopp Iteration. On COCO, a single FCOS-ResNet-50 detector equipped with Optimal Transport Assignment (OTA) can reach 40.7% mAP under 1X scheduler, outperforming all other existing assigning methods. Extensive experiments conducted on COCO and CrowdHuman further validate the effectiveness of our proposed OTA, especially its superiority in crowd scenarios. The code is available at https://github.com/Megvii-BaseDetection/OTA.

Haoran Wei, Lingyu Kong, Jinyue Chen et al.

Playing Large Vision Language Models (LVLMs) in 2023 is trendy among the AI community. However, the relatively large number of parameters (more than 7B) of popular LVLMs makes it difficult to train and deploy on consumer GPUs, discouraging many researchers with limited resources. Imagine how cool it would be to experience all the features of current LVLMs on an old GTX1080ti (our only game card). Accordingly, we present Vary-toy in this report, a small-size Vary along with Qwen-1.8B as the base ``large'' language model. In Vary-toy, we introduce an improved vision vocabulary, allowing the model to not only possess all features of Vary but also gather more generality. Specifically, we replace negative samples of natural images with positive sample data driven by object detection in the procedure of generating vision vocabulary, more sufficiently utilizing the capacity of the vocabulary network and enabling it to efficiently encode visual information corresponding to natural objects. For experiments, Vary-toy can achieve 65.6% ANLS on DocVQA, 59.1% accuracy on ChartQA, 88.1% accuracy on RefCOCO, and 29% on MMVet. The code will be publicly available on the homepage.

Chenglong Liu, Haoran Wei, Jinyue Chen et al.

Modern LVLMs still struggle to achieve fine-grained document understanding, such as OCR/translation/caption for regions of interest to the user, tasks that require the context of the entire page, or even multiple pages. Accordingly, this paper proposes Fox, an effective pipeline, hybrid data, and tuning strategy, that catalyzes LVLMs to focus anywhere on single/multi-page documents. We introduce a novel task to boost the document understanding by making LVLMs focus attention on the document-level region, such as redefining full-page OCR as foreground focus. We employ multiple vision vocabularies to extract visual hybrid knowledge for interleaved document pages (e.g., a page containing a photo). Meanwhile, we render cross-vocabulary vision data as the catalyzer to achieve a full reaction of multiple visual vocabularies and in-document figure understanding. Further, without modifying the weights of multiple vision vocabularies, the above catalyzed fine-grained understanding capabilities can be efficiently tuned to multi-page documents, enabling the model to focus anywhere in both format-free and page-free manners. Besides, we build a benchmark including 9 fine-grained sub-tasks (e.g., region-level OCR/summary, color-guided OCR) to promote document analysis in the community. The experimental results verify the superiority of our model.

En Yu, Kangheng Lin, Liang Zhao et al. · tsinghua

Inspired by the success of DeepSeek-R1, we explore the potential of rule-based reinforcement learning (RL) in MLLM post-training for perception policy learning. While promising, our initial experiments reveal that incorporating a thinking process through RL does not consistently lead to performance gains across all visual perception tasks. This leads us to delve into the essential role of RL in the context of visual perception. In this work, we return to the fundamentals and explore the effects of RL on different perception tasks. We observe that the perceptual complexity is a major factor in determining the effectiveness of RL. We also observe that reward design plays a crucial role in further approching the upper limit of model perception. To leverage these findings, we propose Perception-R1, a scalable RL framework using GRPO during MLLM post-training. With a standard Qwen2.5-VL-3B-Instruct, Perception-R1 achieves +4.2% on RefCOCO+, +17.9% on PixMo-Count, +4.2% on PageOCR, and notably, 31.9% AP on COCO2017 val for the first time, establishing a strong baseline for perception policy learning.

Jinyue Chen, Lingyu Kong, Haoran Wei et al.

Chart parsing poses a significant challenge due to the diversity of styles, values, texts, and so forth. Even advanced large vision-language models (LVLMs) with billions of parameters struggle to handle such tasks satisfactorily. To address this, we propose OneChart: a reliable agent specifically devised for the structural extraction of chart information. Similar to popular LVLMs, OneChart incorporates an autoregressive main body. Uniquely, to enhance the reliability of the numerical parts of the output, we introduce an auxiliary token placed at the beginning of the total tokens along with an additional decoder. The numerically optimized (auxiliary) token allows subsequent tokens for chart parsing to capture enhanced numerical features through causal attention. Furthermore, with the aid of the auxiliary token, we have devised a self-evaluation mechanism that enables the model to gauge the reliability of its chart parsing results by providing confidence scores for the generated content. Compared to current state-of-the-art (SOTA) chart parsing models, e.g., DePlot, ChartVLM, ChartAst, OneChart significantly outperforms in Average Precision (AP) for chart structural extraction across multiple public benchmarks, despite enjoying only 0.2 billion parameters. Moreover, as a chart parsing agent, it also brings 10%+ accuracy gains for the popular LVLM (LLaVA-1.6) in the downstream ChartQA benchmark.

Haochen Wang, Anlin Zheng, Yucheng Zhao et al.

This paper introduces reconstructive visual instruction tuning (ROSS), a family of Large Multimodal Models (LMMs) that exploit vision-centric supervision signals. In contrast to conventional visual instruction tuning approaches that exclusively supervise text outputs, ROSS prompts LMMs to supervise visual outputs via reconstructing input images. By doing so, it capitalizes on the inherent richness and detail present within input images themselves, which are often lost in pure text supervision. However, producing meaningful feedback from natural images is challenging due to the heavy spatial redundancy of visual signals. To address this issue, ROSS employs a denoising objective to reconstruct latent representations of input images, avoiding directly regressing exact raw RGB values. This intrinsic activation design inherently encourages LMMs to maintain image detail, thereby enhancing their fine-grained comprehension capabilities and reducing hallucinations. Empirically, ROSS consistently brings significant improvements across different visual encoders and language models. In comparison with extrinsic assistance state-of-the-art alternatives that aggregate multiple visual experts, ROSS delivers competitive performance with a single SigLIP visual encoder, demonstrating the efficacy of our vision-centric supervision tailored for visual outputs.

En Yu, Kangheng Lin, Liang Zhao et al.

In the pursuit of superior video-processing MLLMs, we have encountered a perplexing paradox: the "anti-scaling law", where more data and larger models lead to worse performance. This study unmasks the culprit: "temporal hacking", a phenomenon where models shortcut by fixating on select frames, missing the full video narrative. In this work, we systematically establish a comprehensive theory of temporal hacking, defining it from a reinforcement learning perspective, introducing the Temporal Perplexity (TPL) score to assess this misalignment, and proposing the Unhackable Temporal Rewarding (UTR) framework to mitigate the temporal hacking. Both theoretically and empirically, TPL proves to be a reliable indicator of temporal modeling quality, correlating strongly with frame activation patterns. Extensive experiments reveal that UTR not only counters temporal hacking but significantly elevates video comprehension capabilities. This work not only advances video-AI systems but also illuminates the critical importance of aligning proxy rewards with true objectives in MLLM development.

Deyu Zhou, Quan Sun, Yuang Peng et al. · tsinghua

We introduce MAGI, a hybrid video generation framework that combines masked modeling for intra-frame generation with causal modeling for next-frame generation. Our key innovation, Complete Teacher Forcing (CTF), conditions masked frames on complete observation frames rather than masked ones (namely Masked Teacher Forcing, MTF), enabling a smooth transition from token-level (patch-level) to frame-level autoregressive generation. CTF significantly outperforms MTF, achieving a +23% improvement in FVD scores on first-frame conditioned video prediction. To address issues like exposure bias, we employ targeted training strategies, setting a new benchmark in autoregressive video generation. Experiments show that MAGI can generate long, coherent video sequences exceeding 100 frames, even when trained on as few as 16 frames, highlighting its potential for scalable, high-quality video generation.