Zhiyuan Zhang

Omar Khattab, Arnav Singhvi, Paridhi Maheshwari et al.

The ML community is rapidly exploring techniques for prompting language models (LMs) and for stacking them into pipelines that solve complex tasks. Unfortunately, existing LM pipelines are typically implemented using hard-coded "prompt templates", i.e. lengthy strings discovered via trial and error. Toward a more systematic approach for developing and optimizing LM pipelines, we introduce DSPy, a programming model that abstracts LM pipelines as text transformation graphs, i.e. imperative computational graphs where LMs are invoked through declarative modules. DSPy modules are parameterized, meaning they can learn (by creating and collecting demonstrations) how to apply compositions of prompting, finetuning, augmentation, and reasoning techniques. We design a compiler that will optimize any DSPy pipeline to maximize a given metric. We conduct two case studies, showing that succinct DSPy programs can express and optimize sophisticated LM pipelines that reason about math word problems, tackle multi-hop retrieval, answer complex questions, and control agent loops. Within minutes of compiling, a few lines of DSPy allow GPT-3.5 and llama2-13b-chat to self-bootstrap pipelines that outperform standard few-shot prompting (generally by over 25% and 65%, respectively) and pipelines with expert-created demonstrations (by up to 5-46% and 16-40%, respectively). On top of that, DSPy programs compiled to open and relatively small LMs like 770M-parameter T5 and llama2-13b-chat are competitive with approaches that rely on expert-written prompt chains for proprietary GPT-3.5. DSPy is available at https://github.com/stanfordnlp/dspy

Sishuo Chen, Wenkai Yang, Zhiyuan Zhang et al. · pku

Natural language processing (NLP) models are known to be vulnerable to backdoor attacks, which poses a newly arisen threat to NLP models. Prior online backdoor defense methods for NLP models only focus on the anomalies at either the input or output level, still suffering from fragility to adaptive attacks and high computational cost. In this work, we take the first step to investigate the unconcealment of textual poisoned samples at the intermediate-feature level and propose a feature-based efficient online defense method. Through extensive experiments on existing attacking methods, we find that the poisoned samples are far away from clean samples in the intermediate feature space of a poisoned NLP model. Motivated by this observation, we devise a distance-based anomaly score (DAN) to distinguish poisoned samples from clean samples at the feature level. Experiments on sentiment analysis and offense detection tasks demonstrate the superiority of DAN, as it substantially surpasses existing online defense methods in terms of defending performance and enjoys lower inference costs. Moreover, we show that DAN is also resistant to adaptive attacks based on feature-level regularization. Our code is available at https://github.com/lancopku/DAN.

Bin Fan, Yuchao Dai, Zhiyuan Zhang et al.

With the ubiquity of rolling shutter (RS) cameras, it is becoming increasingly attractive to recover the latent global shutter (GS) video from two consecutive RS frames, which also places a higher demand on realism. Existing solutions, using deep neural networks or optimization, achieve promising performance. However, these methods generate intermediate GS frames through image warping based on the RS model, which inevitably result in black holes and noticeable motion artifacts. In this paper, we alleviate these issues by proposing a context-aware GS video reconstruction architecture. It facilitates the advantages such as occlusion reasoning, motion compensation, and temporal abstraction. Specifically, we first estimate the bilateral motion field so that the pixels of the two RS frames are warped to a common GS frame accordingly. Then, a refinement scheme is proposed to guide the GS frame synthesis along with bilateral occlusion masks to produce high-fidelity GS video frames at arbitrary times. Furthermore, we derive an approximated bilateral motion field model, which can serve as an alternative to provide a simple but effective GS frame initialization for related tasks. Experiments on synthetic and real data show that our approach achieves superior performance over state-of-the-art methods in terms of objective metrics and subjective visual quality. Code is available at \url{https://github.com/GitCVfb/CVR}.

Cursor Research, Aaron Chan, Ahmed Shalaby et al. · berkeley, microsoft-research

Composer 2 is a specialized model designed for agentic software engineering. The model demonstrates strong long-term planning and coding intelligence while maintaining the ability to efficiently solve problems for interactive use. The model is trained in two phases: first, continued pretraining to improve the model's knowledge and latent coding ability, followed by large-scale reinforcement learning to improve end-to-end coding performance through stronger reasoning, accurate multi-step execution, and coherence on long-horizon realistic coding problems. We develop infrastructure to support training in the same Cursor harness that is used by the deployed model, with equivalent tools and structure, and use environments that match real problems closely. To measure the ability of the model on increasingly difficult tasks, we introduce a benchmark derived from real software engineering problems in large codebases including our own. Composer 2 is a frontier-level coding model and demonstrates a process for training strong domain-specialized models. On our CursorBench evaluations the model achieves a major improvement in accuracy compared to previous Composer models (61.3). On public benchmarks the model scores 61.7 on Terminal-Bench and 73.7 on SWE-bench Multilingual in our harness, comparable to state-of-the-art systems.

Yifan Liu, Yanling Sang, Xishun Liao et al. · stanford

Tourist mobility poses a distinct challenge for urban transportation planning. Unlike resident commuting, tourist travel is largely non-routine, attraction driven, and highly sensitive to trip purpose, travel season, and trip member composition. Existing approaches either measure aggregate tourist spatial patterns without generating individual schedules, or synthesize mobility without tourist specific structure such as trip duration conditioning, month varying attraction demand, and household co-travel rules. To address these challenges, we propose a four stage simulation framework combining month conditioned spatial priors derived from GPS and survey data, trip extent prediction from tourist demographics, distance feasible ward sequence assignment, and LLM-based activity chain generation under household and spatial constraints. GPS data are used only in privacy preserving aggregated form as month conditioned spatial priors, with no individual traces retained or exposed. Experiments on tourism in Tokyo demonstrate that the GPS based tourist cohort extraction recovers spatial visitation signatures consistent with survey references, and our framework produces demographically aligned synthetic schedules whose ward-level visitation shares align closely with both survey distributions and staypoint derived monthly visitation patterns. The results demonstrate the framework's effectiveness as a geographically grounded, demographically aware approach to tourist mobility modeling.

Nan Jiang, Tengyu Liu, Zhexuan Cao et al. · pku

Fine-grained capturing of 3D HOI boosts human activity understanding and facilitates downstream visual tasks, including action recognition, holistic scene reconstruction, and human motion synthesis. Despite its significance, existing works mostly assume that humans interact with rigid objects using only a few body parts, limiting their scope. In this paper, we address the challenging problem of f-AHOI, wherein the whole human bodies interact with articulated objects, whose parts are connected by movable joints. We present CHAIRS, a large-scale motion-captured f-AHOI dataset, consisting of 16.2 hours of versatile interactions between 46 participants and 81 articulated and rigid sittable objects. CHAIRS provides 3D meshes of both humans and articulated objects during the entire interactive process, as well as realistic and physically plausible full-body interactions. We show the value of CHAIRS with object pose estimation. By learning the geometrical relationships in HOI, we devise the very first model that leverage human pose estimation to tackle the estimation of articulated object poses and shapes during whole-body interactions. Given an image and an estimated human pose, our model first reconstructs the pose and shape of the object, then optimizes the reconstruction according to a learned interaction prior. Under both evaluation settings (e.g., with or without the knowledge of objects' geometries/structures), our model significantly outperforms baselines. We hope CHAIRS will promote the community towards finer-grained interaction understanding. We will make the data/code publicly available.

Zhiyuan Zhang, Lingjuan Lyu, Xingjun Ma et al. · pku

Deep Neural Networks (DNNs) are known to be vulnerable to backdoor attacks. In Natural Language Processing (NLP), DNNs are often backdoored during the fine-tuning process of a large-scale Pre-trained Language Model (PLM) with poisoned samples. Although the clean weights of PLMs are readily available, existing methods have ignored this information in defending NLP models against backdoor attacks. In this work, we take the first step to exploit the pre-trained (unfine-tuned) weights to mitigate backdoors in fine-tuned language models. Specifically, we leverage the clean pre-trained weights via two complementary techniques: (1) a two-step Fine-mixing technique, which first mixes the backdoored weights (fine-tuned on poisoned data) with the pre-trained weights, then fine-tunes the mixed weights on a small subset of clean data; (2) an Embedding Purification (E-PUR) technique, which mitigates potential backdoors existing in the word embeddings. We compare Fine-mixing with typical backdoor mitigation methods on three single-sentence sentiment classification tasks and two sentence-pair classification tasks and show that it outperforms the baselines by a considerable margin in all scenarios. We also show that our E-PUR method can benefit existing mitigation methods. Our work establishes a simple but strong baseline defense for secure fine-tuned NLP models against backdoor attacks.

Lei Li, Zhiyuan Zhang, Ruihan Bao et al. · pku

Traditional knowledge distillation in classification problems transfers the knowledge via class correlations in the soft label produced by teacher models, which are not available in regression problems like stock trading volume prediction. To remedy this, we present a novel distillation framework for training a light-weight student model to perform trading volume prediction given historical transaction data. Specifically, we turn the regression model into a probabilistic forecasting model, by training models to predict a Gaussian distribution to which the trading volume belongs. The student model can thus learn from the teacher at a more informative distributional level, by matching its predicted distributions to that of the teacher. Two correlational distillation objectives are further introduced to encourage the student to produce consistent pair-wise relationships with the teacher model. We evaluate the framework on a real-world stock volume dataset with two different time window settings. Experiments demonstrate that our framework is superior to strong baseline models, compressing the model size by $5\times$ while maintaining $99.6\%$ prediction accuracy. The extensive analysis further reveals that our framework is more effective than vanilla distillation methods under low-resource scenarios.

Xiaosong Jia, Yuqian Shao, Zhenjie Yang et al.

With the rise of vision-language models (VLM), their application for autonomous driving (VLM4AD) has gained significant attention. Meanwhile, in autonomous driving, closed-loop evaluation has become widely recognized as a more reliable validation method than open-loop evaluation, as it can evaluate the performance of the model under cumulative errors and out-of-distribution inputs. However, existing VLM4AD benchmarks evaluate the model`s scene understanding ability under open-loop, i.e., via static question-answer (QA) dataset. This kind of evaluation fails to assess the VLMs performance under out-of-distribution states rarely appeared in the human collected datasets.To this end, we present Bench2Drive-VL, an extension of Bench2Drive that brings closed-loop evaluation to VLM-based driving, which introduces: (1) DriveCommenter, a closed-loop generator that automatically generates diverse, behavior-grounded question-answer pairs for all driving situations in CARLA,including severe off-route and off-road deviations previously unassessable in simulation. (2) A unified protocol and interface that allows modern VLMs to be directly plugged into the Bench2Drive closed-loop environment to compare with traditional agents. (3) A flexible reasoning and control framework, supporting multi-format visual inputs and configurable graph-based chain-of-thought execution. (4) A complete development ecosystem. Together, these components form a comprehensive closed-loop benchmark for VLM4AD. All codes and annotated datasets are open sourced.

Zhiyuan Zhang, Hua Yang, Zhouping Yin

High-resolution multi-modality information acquired by vision-based tactile sensors can support more dexterous manipulations for robot fingers. Optical flow is low-level information directly obtained by vision-based tactile sensors, which can be transformed into other modalities like force, geometry and depth. Current vision-tactile sensors employ optical flow methods from OpenCV to estimate the deformation of markers in gels. However, these methods need to be more precise for accurately measuring the displacement of markers during large elastic deformation of the gel, as this can significantly impact the accuracy of downstream tasks. This study proposes a self-supervised optical flow method based on deep learning to achieve high accuracy in displacement measurement for vision-based tactile sensors. The proposed method employs a coarse-to-fine strategy to handle large deformations by constructing a multi-scale feature pyramid from the input image. To better deal with the elastic deformation caused by the gel, the Helmholtz velocity decomposition constraint combined with the elastic deformation constraint are adopted to address the distortion rate and area change rate, respectively. A local flow fusion module is designed to smooth the optical flow, taking into account the prior knowledge of the blurred effect of gel deformation. We trained the proposed self-supervised network using an open-source dataset and compared it with traditional and deep learning-based optical flow methods. The results show that the proposed method achieved the highest displacement measurement accuracy, thereby demonstrating its potential for enabling more precise measurement of downstream tasks using vision-based tactile sensors.

Zhiyuan Zhang, Qi Su, Xu Sun · pku

Despite the potential of federated learning, it is known to be vulnerable to backdoor attacks. Many robust federated aggregation methods are proposed to reduce the potential backdoor risk. However, they are mainly validated in the CV field. In this paper, we find that NLP backdoors are hard to defend against than CV, and we provide a theoretical analysis that the malicious update detection error probabilities are determined by the relative backdoor strengths. NLP attacks tend to have small relative backdoor strengths, which may result in the failure of robust federated aggregation methods for NLP attacks. Inspired by the theoretical results, we can choose some dimensions with higher backdoor strengths to settle this issue. We propose a novel federated aggregation algorithm, Dim-Krum, for NLP tasks, and experimental results validate its effectiveness.

Zhiyuan Zhang, Ruixuan Luo, Qi Su et al. · pku

Recently, Sharpness-Aware Minimization (SAM) algorithm has shown state-of-the-art generalization abilities in vision tasks. It demonstrates that flat minima tend to imply better generalization abilities. However, it has some difficulty implying SAM to some natural language tasks, especially to models with drastic gradient changes, such as RNNs. In this work, we analyze the relation between the flatness of the local minimum and its generalization ability from a novel and straightforward theoretical perspective. We propose that the shift of the training and test distributions can be equivalently seen as a virtual parameter corruption or perturbation, which can explain why flat minima that are robust against parameter corruptions or perturbations have better generalization performances. On its basis, we propose a Gradient-Strength based Adaptive Sharpness-Aware Minimization (GA-SAM) algorithm to help to learn algorithms find flat minima that generalize better. Results in various language benchmarks validate the effectiveness of the proposed GA-SAM algorithm on natural language tasks.

Xiaosong Jia, Chenhe Zhang, Yule Jiang et al.

Existing autonomous driving systems rely on onboard sensors (cameras, LiDAR, IMU, etc) for environmental perception. However, this paradigm is limited by the drive-time perception horizon and often fails under limited view scope, occlusion or extreme conditions such as darkness and rain. In contrast, human drivers are able to recall road structure even under poor visibility. To endow models with this ``recall" ability, we propose the spatial retrieval paradigm, introducing offline retrieved geographic images as an additional input. These images are easy to obtain from offline caches (e.g, Google Maps or stored autonomous driving datasets) without requiring additional sensors, making it a plug-and-play extension for existing AD tasks. For experiments, we first extend the nuScenes dataset with geographic images retrieved via Google Maps APIs and align the new data with ego-vehicle trajectories. We establish baselines across five core autonomous driving tasks: object detection, online mapping, occupancy prediction, end-to-end planning, and generative world modeling. Extensive experiments show that the extended modality could enhance the performance of certain tasks. We will open-source dataset curation code, data, and benchmarks for further study of this new autonomous driving paradigm.

Zhiyuan Zhang, Jiadai Sun, Yuchao Dai et al.

3D point cloud registration is fragile to outliers, which are labeled as the points without corresponding points. To handle this problem, a widely adopted strategy is to estimate the relative pose based only on some accurate correspondences, which is achieved by building correspondences on the identified inliers or by selecting reliable ones. However, these approaches are usually complicated and time-consuming. By contrast, the virtual point-based methods learn the virtual corresponding points (VCPs) for all source points uniformly without distinguishing the outliers and the inliers. Although this strategy is time-efficient, the learned VCPs usually exhibit serious collapse degeneration due to insufficient supervision and the inherent distribution limitation. In this paper, we propose to exploit the best of both worlds and present a novel robust 3D point cloud registration framework. We follow the idea of the virtual point-based methods but learn a new type of virtual points called rectified virtual corresponding points (RCPs), which are defined as the point set with the same shape as the source and with the same pose as the target. Hence, a pair of consistent point clouds, i.e. source and RCPs, is formed by rectifying VCPs to RCPs (VRNet), through which reliable correspondences between source and RCPs can be accurately obtained. Since the relative pose between source and RCPs is the same as the relative pose between source and target, the input point clouds can be registered naturally. Specifically, we first construct the initial VCPs by using an estimated soft matching matrix to perform a weighted average on the target points. Then, we design a correction-walk module to learn an offset to rectify VCPs to RCPs, which effectively breaks the distribution limitation of VCPs. Finally, we develop a hybrid loss function to enforce the shape and geometry structure consistency ...

Zhiyuan Zhang, Jiadai Sun, Yuchao Dai et al.

3D point cloud registration in remote sensing field has been greatly advanced by deep learning based methods, where the rigid transformation is either directly regressed from the two point clouds (correspondences-free approaches) or computed from the learned correspondences (correspondences-based approaches). Existing correspondences-free methods generally learn the holistic representation of the entire point cloud, which is fragile for partial and noisy point clouds. In this paper, we propose a correspondences-free unsupervised point cloud registration (UPCR) method from the representation separation perspective. First, we model the input point cloud as a combination of pose-invariant representation and pose-related representation. Second, the pose-related representation is used to learn the relative pose wrt a "latent canonical shape" for the source and target point clouds respectively. Third, the rigid transformation is obtained from the above two learned relative poses. Our method not only filters out the disturbance in pose-invariant representation but also is robust to partial-to-partial point clouds or noise. Experiments on benchmark datasets demonstrate that our unsupervised method achieves comparable if not better performance than state-of-the-art supervised registration methods.

Ruibo Chen, Wei Li, Zhiyuan Zhang et al. · pku

Volume prediction is one of the fundamental objectives in the Fintech area, which is helpful for many downstream tasks, e.g., algorithmic trading. Previous methods mostly learn a universal model for different stocks. However, this kind of practice omits the specific characteristics of individual stocks by applying the same set of parameters for different stocks. On the other hand, learning different models for each stock would face data sparsity or cold start problems for many stocks with small capitalization. To take advantage of the data scale and the various characteristics of individual stocks, we propose a dual-process meta-learning method that treats the prediction of each stock as one task under the meta-learning framework. Our method can model the common pattern behind different stocks with a meta-learner, while modeling the specific pattern for each stock across time spans with stock-dependent parameters. Furthermore, we propose to mine the pattern of each stock in the form of a latent variable which is then used for learning the parameters for the prediction module. This makes the prediction procedure aware of the data pattern. Extensive experiments on volume predictions show that our method can improve the performance of various baseline models. Further analyses testify the effectiveness of our proposed meta-learning framework.

Zhiyuan Zhang, Licheng Yang, Zhiyu Xiang

Despite the progress on 3D point cloud deep learning, most prior works focus on learning features that are invariant to translation and point permutation, and very limited efforts have been devoted for rotation invariant property. Several recent studies achieve rotation invariance at the cost of lower accuracies. In this work, we close this gap by proposing a novel yet effective rotation invariant architecture for 3D point cloud classification and segmentation. Instead of traditional pointwise operations, we construct local triangle surfaces to capture more detailed surface structure, based on which we can extract highly expressive rotation invariant surface properties which are then integrated into an attention-augmented convolution operator named RISurConv to generate refined attention features via self-attention layers. Based on RISurConv we build an effective neural network for 3D point cloud analysis that is invariant to arbitrary rotations while maintaining high accuracy. We verify the performance on various benchmarks with supreme results obtained surpassing the previous state-of-the-art by a large margin. We achieve an overall accuracy of 96.0% (+4.7%) on ModelNet40, 93.1% (+12.8%) on ScanObjectNN, and class accuracies of 91.5% (+3.6%), 82.7% (+5.1%), and 78.5% (+9.2%) on the three categories of the FG3D dataset for the fine-grained classification task. Additionally, we achieve 81.5% (+1.0%) mIoU on ShapeNet for the segmentation task. Code is available here: https://github.com/cszyzhang/RISurConv

Bo Yang, Haoxuan Ma, Yifan Liu et al. · stanford

Large-scale GPS trajectory data offer rich observations of human mobility, yet assigning trip purposes to detected stops remains challenging due to the absence of individual-level ground truth, spatial uncertainty from GPS noise and incomplete points of interest (POIs) coverage, and fundamental behavioral differences across trip purposes. We propose a weakly supervised framework integrating neighborhood-level POI semantic zones with distance-weighted spatial likelihoods, differentiated inference strategies for mandatory and non-mandatory activities, and a multi-phase Pareto optimization that jointly minimizes distributional divergence from household travel survey statistics and maximizes inference reliability without requiring annotated labels. Evaluated on over 81 million staypoints in Los Angeles, the framework reduces activity type frequency Jensen-Shannon distance (JSD) by 23%, start time JSD by 48%, and duration JSD by 12% respectively relative to a comparable baseline. The proposed approach provides a scalable and uncertainty-aware path from raw GPS trajectories to semantically annotated mobility data for travel demand modeling and transportation policy analysis.

Zhiyuan Zhang, Yuchao Dai, Bin Fan et al.

Existing learning-based point feature descriptors are usually task-agnostic, which pursue describing the individual 3D point clouds as accurate as possible. However, the matching task aims at describing the corresponding points consistently across different 3D point clouds. Therefore these too accurate features may play a counterproductive role due to the inconsistent point feature representations of correspondences caused by the unpredictable noise, partiality, deformation, \etc, in the local geometry. In this paper, we propose to learn a robust task-specific feature descriptor to consistently describe the correct point correspondence under interference. Born with an Encoder and a Dynamic Fusion module, our method EDFNet develops from two aspects. First, we augment the matchability of correspondences by utilizing their repetitive local structure. To this end, a special encoder is designed to exploit two input point clouds jointly for each point descriptor. It not only captures the local geometry of each point in the current point cloud by convolution, but also exploits the repetitive structure from paired point cloud by Transformer. Second, we propose a dynamical fusion module to jointly use different scale features. There is an inevitable struggle between robustness and discriminativeness of the single scale feature. Specifically, the small scale feature is robust since little interference exists in this small receptive field. But it is not sufficiently discriminative as there are many repetitive local structures within a point cloud. Thus the resultant descriptors will lead to many incorrect matches. In contrast, the large scale feature is more discriminative by integrating more neighborhood information. ...

Tuan-Anh Vu, Srinjay Sarkar, Zhiyuan Zhang et al.

Data augmentation is a powerful technique to enhance the performance of a deep learning task but has received less attention in 3D deep learning. It is well known that when 3D shapes are sparsely represented with low point density, the performance of the downstream tasks drops significantly. This work explores test-time augmentation (TTA) for 3D point clouds. We are inspired by the recent revolution of learning implicit representation and point cloud upsampling, which can produce high-quality 3D surface reconstruction and proximity-to-surface, respectively. Our idea is to leverage the implicit field reconstruction or point cloud upsampling techniques as a systematic way to augment point cloud data. Mainly, we test both strategies by sampling points from the reconstructed results and using the sampled point cloud as test-time augmented data. We show that both strategies are effective in improving accuracy. We observed that point cloud upsampling for test-time augmentation can lead to more significant performance improvement on downstream tasks such as object classification and segmentation on the ModelNet40, ShapeNet, ScanObjectNN, and SemanticKITTI datasets, especially for sparse point clouds.

Jiaqi Gu, Zhiyu Xiang, Pan Zhao et al.

In recent years 3D object detection from LiDAR point clouds has made great progress thanks to the development of deep learning technologies. Although voxel or point based methods are popular in 3D object detection, they usually involve time-consuming operations such as 3D convolutions on voxels or ball query among points, making the resulting network inappropriate for time critical applications. On the other hand, 2D view-based methods feature high computing efficiency while usually obtaining inferior performance than the voxel or point based methods. In this work, we present a real-time view-based single stage 3D object detector, namely CVFNet to fulfill this task. To strengthen the cross-view feature learning under the condition of demanding efficiency, our framework extracts the features of different views and fuses them in an efficient progressive way. We first propose a novel Point-Range feature fusion module that deeply integrates point and range view features in multiple stages. Then, a special Slice Pillar is designed to well maintain the 3D geometry when transforming the obtained deep point-view features into bird's eye view. To better balance the ratio of samples, a sparse pillar detection head is presented to focus the detection on the nonempty grids. We conduct experiments on the popular KITTI and NuScenes benchmark, and state-of-the-art performances are achieved in terms of both accuracy and speed.

Ruibo Chen, Zhiyuan Zhang, Yi Liu et al. · pku

Multimodal stock trading volume movement prediction with stock-related news is one of the fundamental problems in the financial area. Existing multimodal works that train models from scratch face the problem of lacking universal knowledge when modeling financial news. In addition, the models ability may be limited by the lack of domain-related knowledge due to insufficient data in the datasets. To handle this issue, we propose the Prompt-based MUltimodal Stock volumE prediction model (ProMUSE) to process text and time series modalities. We use pre-trained language models for better comprehension of financial news and adopt prompt learning methods to leverage their capability in universal knowledge to model textual information. Besides, simply fusing two modalities can cause harm to the unimodal representations. Thus, we propose a novel cross-modality contrastive alignment while reserving the unimodal heads beside the fusion head to mitigate this problem. Extensive experiments demonstrate that our proposed ProMUSE outperforms existing baselines. Comprehensive analyses further validate the effectiveness of our architecture compared to potential variants and learning mechanisms.

Zhiyuan Zhang, Zijian Zhou, Linjun Li et al.

3D texture generation is receiving increasing attention, as it enables the creation of realistic and aesthetic texture materials for untextured 3D meshes. However, existing 3D texture generation methods are limited to producing only a few types of non-emissive PBR materials (e.g., albedo, metallic maps and roughness maps), making them difficult to replicate highly popular styles, such as cyberpunk, failing to achieve effects like realistic LED emissions. To address this limitation, we propose a novel task, emission texture generation, which enables the synthesized 3D objects to faithfully reproduce the emission materials from input reference images. Our key contributions include: first, We construct the Objaverse-Emission dataset, the first dataset that contains 40k 3D assets with high-quality emission materials. Second, we propose EmissionGen, a novel baseline for the emission texture generation task. Third, we define detailed evaluation metrics for the emission texture generation task. Our results demonstrate significant potential for future industrial applications. Dataset will be available at https://github.com/yx345kw/EmissionGen.

Junyu Shi, Haoting Wu, Zhiyuan Zhang et al.

3D human motion prediction aims to generate coherent future motions from observed sequences, yet existing end-to-end regression frameworks often fail to capture complex dynamics and tend to produce temporally inconsistent or static predictions-a limitation rooted in representation shortcutting, where models rely on superficial cues rather than learning meaningful motion structure. We propose a two-stage self-supervised framework that decouples representation learning from prediction. In the pretraining stage, the model performs unified past-future self-reconstruction, reconstructing the past sequence while recovering masked joints in the future sequence under full historical guidance. A velocity-based masking strategy selects highly dynamic joints, forcing the model to focus on informative motion components and internalize the statistical dependencies between past and future states without regression interference. In the fine-tuning stage, the pretrained model predicts the entire future sequence, now treated as fully masked, and is further equipped with a lightweight future-text prediction head for joint optimization of low-level motion prediction and high-level motion understanding. Experiments on Human3.6M, 3DPW, and AMASS show that our method reduces average prediction errors by 8.8% over state-of-the-art methods while achieving competitive future-motion understanding performance compared to LLM-based models. Code is available at: https://github.com/JunyuShi02/MoReFun

Zhiyuan Zhang, Jiadai Sun, Yuchao Dai et al.

Although 3D point cloud data has received widespread attentions as a general form of 3D signal expression, applying point clouds to the task of dense correspondence estimation between 3D shapes has not been investigated widely. Furthermore, even in the few existing 3D point cloud-based methods, an important and widely acknowledged principle, i.e . one-to-one matching, is usually ignored. In response, this paper presents a novel end-to-end learning-based method to estimate the dense correspondence of 3D point clouds, in which the problem of point matching is formulated as a zero-one assignment problem to achieve a permutation matching matrix to implement the one-to-one principle fundamentally. Note that the classical solutions of this assignment problem are always non-differentiable, which is fatal for deep learning frameworks. Thus we design a special matching module, which solves a doubly stochastic matrix at first and then projects this obtained approximate solution to the desired permutation matrix. Moreover, to guarantee end-to-end learning and the accuracy of the calculated loss, we calculate the loss from the learned permutation matrix but propagate the gradient to the doubly stochastic matrix directly which bypasses the permutation matrix during the backward propagation. Our method can be applied to both non-rigid and rigid 3D point cloud data and extensive experiments show that our method achieves state-of-the-art performance for dense correspondence learning.

Zhiyuan Zhang, Zhenghao Jin, Yanlun Peng et al.

Robustness is a critical requirement for deploying autonomous driving systems in the real world. Existing robustness benchmarks for autonomous driving have made important progress in studying the effects of image-level corruptions, such as adverse weather or camera degradation, on perception modules and open-loop planning outputs. However, deployment can also involve system-level imperfections, such as inference latency and ego-state estimation errors, which remain less studied in closed-loop E2E-AD evaluation. These imperfections can accumulate through the feedback loop and destabilize control. In this work, we present Bench2Drive-Robust, to our knowledge the first device-centric robustness benchmark for closed-loop end-to-end autonomous driving under realistic deployment perturbations. We systematically evaluate deployment-oriented perturbations arising from three major sources: camera-stream failures (frame drop, partial observation), ego-state estimation errors (GPS noise, and speed or odometry errors), and compute-induced control delay (model inference delay). We evaluate representative end-to-end driving methods and analyze their robustness under different perturbation severities. Our results show that these deployment-related perturbations can substantially degrade closed-loop driving performance, revealing robustness challenges that are not fully captured by conventional image-level corruption evaluations. By establishing a closed-loop evaluation protocol and demonstrating the substantial impact of these deployment-oriented perturbations, Bench2Drive-Robust defines practical robustness problems for end-to-end autonomous driving and encourages further research on deployment-aware robust driving systems.

Zhiyuan Zhang, Zhitong Huang, Jing Liao

Recent advancements in large-scale text-to-image diffusion models have enabled many applications in image editing. However, none of these methods have been able to edit the layout of single existing images. To address this gap, we propose the first framework for layout editing of a single image while preserving its visual properties, thus allowing for continuous editing on a single image. Our approach is achieved through two key modules. First, to preserve the characteristics of multiple objects within an image, we disentangle the concepts of different objects and embed them into separate textual tokens using a novel method called masked textual inversion. Next, we propose a training-free optimization method to perform layout control for a pre-trained diffusion model, which allows us to regenerate images with learned concepts and align them with user-specified layouts. As the first framework to edit the layout of existing images, we demonstrate that our method is effective and outperforms other baselines that were modified to support this task. Our code will be freely available for public use upon acceptance.

Junqi You, Xiaosong Jia, Zhiyuan Zhang et al.

For end-to-end autonomous driving (E2E-AD), the evaluation system remains an open problem. Existing closed-loop evaluation protocols usually rely on simulators like CARLA being less realistic; while NAVSIM using real-world vision data, yet is limited to fixed planning trajectories in short horizon and assumes other agents are not reactive. We introduce Bench2Drive-R, a generative framework that enables reactive closed-loop evaluation. Unlike existing video generative models for AD, the proposed designs are tailored for interactive simulation, where sensor rendering and behavior rollout are decoupled by applying a separate behavioral controller to simulate the reactions of surrounding agents. As a result, the renderer could focus on image fidelity, control adherence, and spatial-temporal coherence. For temporal consistency, due to the step-wise interaction nature of simulation, we design a noise modulating temporal encoder with Gaussian blurring to encourage long-horizon autoregressive rollout of image sequences without deteriorating distribution shifts. For spatial consistency, a retrieval mechanism, which takes the spatially nearest images as references, is introduced to to ensure scene-level rendering fidelity during the generation process. The spatial relations between target and reference are explicitly modeled with 3D relative position encodings and the potential over-reliance of reference images is mitigated with hierarchical sampling and classifier-free guidance. We compare the generation quality of Bench2Drive-R with existing generative models and achieve state-of-the-art performance. We further integrate Bench2Drive-R into nuPlan and evaluate the generative qualities with closed-loop simulation results. We will open source our code.

Gheorghe Comanici, Eric Bieber, Mike Schaekermann et al. · amazon-science, baidu

In this report, we introduce the Gemini 2.X model family: Gemini 2.5 Pro and Gemini 2.5 Flash, as well as our earlier Gemini 2.0 Flash and Flash-Lite models. Gemini 2.5 Pro is our most capable model yet, achieving SoTA performance on frontier coding and reasoning benchmarks. In addition to its incredible coding and reasoning skills, Gemini 2.5 Pro is a thinking model that excels at multimodal understanding and it is now able to process up to 3 hours of video content. Its unique combination of long context, multimodal and reasoning capabilities can be combined to unlock new agentic workflows. Gemini 2.5 Flash provides excellent reasoning abilities at a fraction of the compute and latency requirements and Gemini 2.0 Flash and Flash-Lite provide high performance at low latency and cost. Taken together, the Gemini 2.X model generation spans the full Pareto frontier of model capability vs cost, allowing users to explore the boundaries of what is possible with complex agentic problem solving.

Zhiyuan Zhang, Xiaosong Jia, Guanyu Chen et al.

In this technical report, we introduce TrajTok, a trajectory tokenizer for discrete next-token-prediction based behavior generation models, which combines data-driven and rule-based methods with better coverage, symmetry and robustness, along with a spatial-aware label smoothing method for cross-entropy loss. We adopt the tokenizer and loss for the SMART model and reach a superior performance with realism score of 0.7852 on the Waymo Open Sim Agents Challenge 2025. We will open-source the code in the future.

Zhiyuan Zhang, Yanzhao Li, Zhiqiang Zou et al.

While 4-bit quantization is essential for high-throughput deployment of Large Language Models, activation outliers often lead to significant accuracy degradation due to the restricted dynamic range of low-bit formats. In this paper, we systematically investigate the spatial distribution of outliers and demonstrate a token-persistent structural clustering effect, where high-magnitude outliers consistently occupy fixed channels across tokens. Building on this insight, we propose OSC, a hardware-efficient framework for outlier suppression. During inference, OSC executes a dual-path computation consisting of a low-precision 4-bit General Matrix Multiplication (GEMM) path and a high-precision 16-bit branch GEMM path. Specifically, OSC uses an offline group-wise strategy to identify the channels where outliers are located and then performs structured sub-tensor extraction to coalesce these scattered activation channels into a compact dense tensor online. This mechanism implements outlier protection through regularized and high-throughput GEMM operations, achieving a seamless fit with modern 4-bit micro-scaling hardware. Furthermore, for the inputs of W2 where outlier clustering is less pronounced, we integrate a fallback strategy to FP8. Evaluation on Qwen3-8B and Qwen3-30B restricts the average accuracy drop to 2.19 and 1.12 points, respectively. Notably, OSC is highly hardware-friendly, achieving a peak speedup of 1.78x over the W8A8 GEMM baseline on a modern AI accelerator.

Yupeng Sun, Yanzhao Li, Zhiqiang Zou et al.

FlashAttention-style online softmax enables exact attention computation with linear memory by streaming score tiles through on-chip memory and maintaining a running maximum and normalizer. However, as attention kernels approach peak tensor-core/cube-core throughput on modern accelerators, non-matmul components of online softmax -- especially per-tile rowmax and rowsum reductions and rescale chains -- can become vector or SIMD limited and dominate latency. This paper revisits FlashAttention and proposes Vector Relieved Flash Attention (VFA), a hardware-friendly method that reduces rowmax-driven updates of the running maximum while retaining the online-softmax structure. VFA initializes the running maximum via a cheap approximation from key-block representations, reorders key-block traversal to prioritize high-impact sink and local blocks, and freezes the maximum for remaining blocks to avoid repeated reductions and rescaling. We further integrate VFA with block-sparse skipping methods such as BLASST to form Vector Relieved Sparse Attention (VSA), which reduces both block count and per-block overhead. Notably, VFA and VSA completely avoid the conditional rescale operation in the update stage used in FA4.0. Extensive evaluations on benchmarks including MMLU and MATH500, together with attention statistics, verify our design: (i) sink and local reordering stabilizes the running maximum early; (ii) simple Q and K block summaries fail due to intra-block heterogeneity; (iii) m-initialization is required when maxima appear in middle blocks. Overall, VFA and VSA efficiently alleviate online-softmax reduction bottlenecks without performance loss. Compared to the C16V32 baseline, C8V32, C4V32 and C4V16 achieve nearly two times speedup on modern hardware while hitting the vector bottleneck. With upcoming architecture improvements, C4V16 will deliver six times speedup by enhancing exponent capacity.

Tianhui Liu, Jie Feng, Hetian Pang et al.

Understanding urban socioeconomic conditions through visual data is a challenging yet essential task for sustainable urban development and policy planning. In this work, we introduce $\textbf{CityLens}$, a comprehensive benchmark designed to evaluate the capabilities of large language-vision models (LLVMs) in predicting socioeconomic indicators from satellite and street view imagery. We construct a multi-modal dataset covering a total of 17 globally distributed cities, spanning 6 key domains: economy, education, crime, transport, health, and environment, reflecting the multifaceted nature of urban life. Based on this dataset, we define 11 prediction tasks and utilize three evaluation paradigms: Direct Metric Prediction, Normalized Metric Estimation, and Feature-Based Regression. We benchmark 17 state-of-the-art LLVMs across these tasks. Our results reveal that while LLVMs demonstrate promising perceptual and reasoning capabilities, they still exhibit limitations in predicting urban socioeconomic indicators. CityLens provides a unified framework for diagnosing these limitations and guiding future efforts in using LLVMs to understand and predict urban socioeconomic patterns. Our codes and datasets are open-sourced via https://github.com/tsinghua-fib-lab/CityLens.

Zhiyuan Zhang, Aditya Mohan, Seungho Han et al.

Robotic imitation learning has achieved impressive success in learning complex manipulation behaviors from demonstrations. However, many existing robot learning methods do not explicitly account for the physical symmetries of robotic systems, often resulting in asymmetric or inconsistent behaviors under symmetric observations. This limitation is particularly pronounced in dual-arm manipulation, where bilateral symmetry is inherent to both the robot morphology and the structure of many tasks. In this paper, we introduce EquiBim, a symmetry-equivariant policy learning framework for bimanual manipulation that enforces bilateral equivariance between observations and actions during training. Our approach formulates physical symmetry as a group action on both observation and action spaces, and imposes an equivariance constraint on policy predictions under symmetric transformations. The framework is model-agnostic and can be seamlessly integrated into a wide range of imitation learning pipelines with diverse observation modalities and action representations, including point cloud-based and image-based policies, as well as both end-effector-space and joint-space parameterizations. We evaluate EquiBim on RoboTwin, a dual-arm robotic platform with symmetric kinematics, and evaluate it across diverse observation and action configurations in simulation. We further validate the approach on a real-world dual-arm system. Across both simulation and physical experiments, our method consistently improves performance and robustness under distribution shifts. These results suggest that explicitly enforcing physical symmetry provides a simple yet effective inductive bias for bimanual robot learning.

Viktor Rogowski, Maarten L. Terpstra, Niklas Wahl et al.

Radiation therapy (RT) requires precise dose delivery over multiple fractions, with CT fundamental for treatment planning due to its electron density information. Repeated CT acquisitions impose radiation exposure and logistical burdens, MRI lacks electron density, and cone-beam CT (CBCT) requires correction for dose calculation. Synthetic CT (sCT) generation addresses these by converting MRI or CBCT into CT-equivalent images with accurate Hounsfield Unit (HU) values, enabling MRI-only RT and CBCT-based adaptive workflows. Building on SynthRAD2023, SynthRAD2025 benchmarked sCT methods on 2,362 patients from five European centers across head and neck, thorax, and abdomen. Two tasks: MRI-to-CT (890 cases) and CBCT-to-CT (1,472 cases), evaluated via image similarity (MAE, PSNR, MS-SSIM), segmentation (Dice, HD95), and dosimetric metrics from photon and proton plans. With 803 participants and 12/13 valid submissions, Task 1 top performance reached MAE $64.8\pm21.3$ HU, PSNR $\sim$30 dB, MS-SSIM $\sim$0.936, Dice 0.79, photon $γ_{2\%/2\text{mm}}>98\%$, proton $γ\approx85\%$. Task 2 improved: MAE $48.3\pm13.4$ HU, PSNR 32.6 dB, MS-SSIM 0.968, Dice 0.86, photon $γ>99\%$, proton $γ\approx89\%$. Strong image--segmentation correlations ($ρ=0.78$--$0.79$) but moderate dose correlations confirmed image quality is insufficient as a dosimetric surrogate. Head-and-neck cases were most consistent; thoracic and abdominal cases showed greater variability. Residual errors at tissue interfaces propagate along beam paths, affecting proton dose more than photon. SynthRAD2025 demonstrates that deep learning yields clinically relevant sCTs, especially for CBCT-to-CT, while identifying persistent MRI-to-CT challenges and underscoring dose-based evaluation as essential for clinical validation.

Yi Li, Zhiyuan Zhang, Jiangnan Xia et al.

This paper presents a novel Two-Stage Diffusion Model (TS-Diff) for enhancing extremely low-light RAW images. In the pre-training stage, TS-Diff synthesizes noisy images by constructing multiple virtual cameras based on a noise space. Camera Feature Integration (CFI) modules are then designed to enable the model to learn generalizable features across diverse virtual cameras. During the aligning stage, CFIs are averaged to create a target-specific CFI$^T$, which is fine-tuned using a small amount of real RAW data to adapt to the noise characteristics of specific cameras. A structural reparameterization technique further simplifies CFI$^T$ for efficient deployment. To address color shifts during the diffusion process, a color corrector is introduced to ensure color consistency by dynamically adjusting global color distributions. Additionally, a novel dataset, QID, is constructed, featuring quantifiable illumination levels and a wide dynamic range, providing a comprehensive benchmark for training and evaluation under extreme low-light conditions. Experimental results demonstrate that TS-Diff achieves state-of-the-art performance on multiple datasets, including QID, SID, and ELD, excelling in denoising, generalization, and color consistency across various cameras and illumination levels. These findings highlight the robustness and versatility of TS-Diff, making it a practical solution for low-light imaging applications. Source codes and models are available at https://github.com/CircccleK/TS-Diff

Junyu Shi, Yong Sun, Zhiyuan Zhang et al.

Existing text-driven motion generation methods often treat synthesis as a bidirectional mapping between language and motion, but remain limited in capturing the causal logic of action execution and the human intentions that drive behavior. The absence of visual grounding further restricts precision and personalization, as language alone cannot specify fine-grained spatiotemporal details. We propose MoGIC, a unified framework that integrates intention modeling and visual priors into multimodal motion synthesis. By jointly optimizing multimodal-conditioned motion generation and intention prediction, MoGIC uncovers latent human goals, leverages visual priors to enhance generation, and exhibits versatile multimodal generative capability. We further introduce a mixture-of-attention mechanism with adaptive scope to enable effective local alignment between conditional tokens and motion subsequences. To support this paradigm, we curate Mo440H, a 440-hour benchmark from 21 high-quality motion datasets. Experiments show that after finetuning, MoGIC reduces FID by 38.6\% on HumanML3D and 34.6\% on Mo440H, surpasses LLM-based methods in motion captioning with a lightweight text head, and further enables intention prediction and vision-conditioned generation, advancing controllable motion synthesis and intention understanding. The code is available at https://github.com/JunyuShi02/MoGIC

Yanyin Guo, Runxuan An, Junwei Li et al.

Traditional ship detection methods primarily rely on single-modal approaches, such as visible or infrared images, which limit their application in complex scenarios involving varying lighting conditions and heavy fog. To address this issue, we explore the advantages of short-wave infrared (SWIR) and long-wave infrared (LWIR) in ship detection and propose a novel single-stage image fusion detection algorithm called LSFDNet. This algorithm leverages feature interaction between the image fusion and object detection subtask networks, achieving remarkable detection performance and generating visually impressive fused images. To further improve the saliency of objects in the fused images and improve the performance of the downstream detection task, we introduce the Multi-Level Cross-Fusion (MLCF) module. This module combines object-sensitive fused features from the detection task and aggregates features across multiple modalities, scales, and tasks to obtain more semantically rich fused features. Moreover, we utilize the position prior from the detection task in the Object Enhancement (OE) loss function, further increasing the retention of object semantics in the fused images. The detection task also utilizes preliminary fused features from the fusion task to complement SWIR and LWIR features, thereby enhancing detection performance. Additionally, we have established a Nearshore Ship Long-Short Wave Registration (NSLSR) dataset to train effective SWIR and LWIR image fusion and detection networks, bridging a gap in this field. We validated the superiority of our proposed single-stage fusion detection algorithm on two datasets. The source code and dataset are available at https://github.com/Yanyin-Guo/LSFDNet

Yajie Fu, Chaorui Huang, Junwei Li et al.

We propose HDiffTG, a novel 3D Human Pose Estimation (3DHPE) method that integrates Transformer, Graph Convolutional Network (GCN), and diffusion model into a unified framework. HDiffTG leverages the strengths of these techniques to significantly improve pose estimation accuracy and robustness while maintaining a lightweight design. The Transformer captures global spatiotemporal dependencies, the GCN models local skeletal structures, and the diffusion model provides step-by-step optimization for fine-tuning, achieving a complementary balance between global and local features. This integration enhances the model's ability to handle pose estimation under occlusions and in complex scenarios. Furthermore, we introduce lightweight optimizations to the integrated model and refine the objective function design to reduce computational overhead without compromising performance. Evaluation results on the Human3.6M and MPI-INF-3DHP datasets demonstrate that HDiffTG achieves state-of-the-art (SOTA) performance on the MPI-INF-3DHP dataset while excelling in both accuracy and computational efficiency. Additionally, the model exhibits exceptional robustness in noisy and occluded environments. Source codes and models are available at https://github.com/CirceJie/HDiffTG

Zhiyuan Zhang, Yuxin He, Yong Sun et al.

Visual Language Models (VLMs) have emerged as pivotal tools for robotic systems, enabling cross-task generalization, dynamic environmental interaction, and long-horizon planning through multimodal perception and semantic reasoning. However, existing open-source VLMs predominantly trained for generic vision-language alignment tasks fail to model temporally correlated action semantics that are crucial for robotic manipulation effectively. While current image-based fine-tuning methods partially adapt VLMs to robotic applications, they fundamentally disregard temporal evolution patterns in video sequences and suffer from visual feature entanglement between robotic agents, manipulated objects, and environmental contexts, thereby limiting semantic decoupling capability for atomic actions and compromising model generalizability.To overcome these challenges, this work presents RoboAct-CLIP with dual technical contributions: 1) A dataset reconstruction framework that performs semantic-constrained action unit segmentation and re-annotation on open-source robotic videos, constructing purified training sets containing singular atomic actions (e.g., "grasp"); 2) A temporal-decoupling fine-tuning strategy based on Contrastive Language-Image Pretraining (CLIP) architecture, which disentangles temporal action features across video frames from object-centric characteristics to achieve hierarchical representation learning of robotic atomic actions.Experimental results in simulated environments demonstrate that the RoboAct-CLIP pretrained model achieves a 12% higher success rate than baseline VLMs, along with superior generalization in multi-object manipulation tasks.

Wenkai Yang, Lei Li, Zhiyuan Zhang et al.

Recent studies have revealed a security threat to natural language processing (NLP) models, called the Backdoor Attack. Victim models can maintain competitive performance on clean samples while behaving abnormally on samples with a specific trigger word inserted. Previous backdoor attacking methods usually assume that attackers have a certain degree of data knowledge, either the dataset which users would use or proxy datasets for a similar task, for implementing the data poisoning procedure. However, in this paper, we find that it is possible to hack the model in a data-free way by modifying one single word embedding vector, with almost no accuracy sacrificed on clean samples. Experimental results on sentiment analysis and sentence-pair classification tasks show that our method is more efficient and stealthier. We hope this work can raise the awareness of such a critical security risk hidden in the embedding layers of NLP models. Our code is available at https://github.com/lancopku/Embedding-Poisoning.

Guangxiang Zhao, Junyang Lin, Zhiyuan Zhang et al.

Self-attention based Transformer has demonstrated the state-of-the-art performances in a number of natural language processing tasks. Self-attention is able to model long-term dependencies, but it may suffer from the extraction of irrelevant information in the context. To tackle the problem, we propose a novel model called \textbf{Explicit Sparse Transformer}. Explicit Sparse Transformer is able to improve the concentration of attention on the global context through an explicit selection of the most relevant segments. Extensive experimental results on a series of natural language processing and computer vision tasks, including neural machine translation, image captioning, and language modeling, all demonstrate the advantages of Explicit Sparse Transformer in model performance. We also show that our proposed sparse attention method achieves comparable or better results than the previous sparse attention method, but significantly reduces training and testing time. For example, the inference speed is twice that of sparsemax in Transformer model. Code will be available at \url{https://github.com/lancopku/Explicit-Sparse-Transformer}

Guangxiang Zhao, Xu Sun, Jingjing Xu et al.

In sequence to sequence learning, the self-attention mechanism proves to be highly effective, and achieves significant improvements in many tasks. However, the self-attention mechanism is not without its own flaws. Although self-attention can model extremely long dependencies, the attention in deep layers tends to overconcentrate on a single token, leading to insufficient use of local information and difficultly in representing long sequences. In this work, we explore parallel multi-scale representation learning on sequence data, striving to capture both long-range and short-range language structures. To this end, we propose the Parallel MUlti-Scale attEntion (MUSE) and MUSE-simple. MUSE-simple contains the basic idea of parallel multi-scale sequence representation learning, and it encodes the sequence in parallel, in terms of different scales with the help from self-attention, and pointwise transformation. MUSE builds on MUSE-simple and explores combining convolution and self-attention for learning sequence representations from more different scales. We focus on machine translation and the proposed approach achieves substantial performance improvements over Transformer, especially on long sequences. More importantly, we find that although conceptually simple, its success in practice requires intricate considerations, and the multi-scale attention must build on unified semantic space. Under common setting, the proposed model achieves substantial performance and outperforms all previous models on three main machine translation tasks. In addition, MUSE has potential for accelerating inference due to its parallelism. Code will be available at https://github.com/lancopku/MUSE

Nan Jiang, Zhiyuan Zhang, Hongjie Li et al. · pku

Confronting the challenges of data scarcity and advanced motion synthesis in human-scene interaction modeling, we introduce the TRUMANS dataset alongside a novel HSI motion synthesis method. TRUMANS stands as the most comprehensive motion-captured HSI dataset currently available, encompassing over 15 hours of human interactions across 100 indoor scenes. It intricately captures whole-body human motions and part-level object dynamics, focusing on the realism of contact. This dataset is further scaled up by transforming physical environments into exact virtual models and applying extensive augmentations to appearance and motion for both humans and objects while maintaining interaction fidelity. Utilizing TRUMANS, we devise a diffusion-based autoregressive model that efficiently generates HSI sequences of any length, taking into account both scene context and intended actions. In experiments, our approach shows remarkable zero-shot generalizability on a range of 3D scene datasets (e.g., PROX, Replica, ScanNet, ScanNet++), producing motions that closely mimic original motion-captured sequences, as confirmed by quantitative experiments and human studies.

Peiyuan Zhi, Zhiyuan Zhang, Yu Zhao et al.

Autonomous robot navigation and manipulation in open environments require reasoning and replanning with closed-loop feedback. In this work, we present COME-robot, the first closed-loop robotic system utilizing the GPT-4V vision-language foundation model for open-ended reasoning and adaptive planning in real-world scenarios.COME-robot incorporates two key innovative modules: (i) a multi-level open-vocabulary perception and situated reasoning module that enables effective exploration of the 3D environment and target object identification using commonsense knowledge and situated information, and (ii) an iterative closed-loop feedback and restoration mechanism that verifies task feasibility, monitors execution success, and traces failure causes across different modules for robust failure recovery. Through comprehensive experiments involving 8 challenging real-world mobile and tabletop manipulation tasks, COME-robot demonstrates a significant improvement in task success rate (~35%) compared to state-of-the-art methods. We further conduct comprehensive analyses to elucidate how COME-robot's design facilitates failure recovery, free-form instruction following, and long-horizon task planning.

Xiaosong Jia, Junqi You, Zhiyuan Zhang et al.

End-to-end autonomous driving (E2E-AD) has emerged as a trend in the field of autonomous driving, promising a data-driven, scalable approach to system design. However, existing E2E-AD methods usually adopt the sequential paradigm of perception-prediction-planning, which leads to cumulative errors and training instability. The manual ordering of tasks also limits the system`s ability to leverage synergies between tasks (for example, planning-aware perception and game-theoretic interactive prediction and planning). Moreover, the dense BEV representation adopted by existing methods brings computational challenges for long-range perception and long-term temporal fusion. To address these challenges, we present DriveTransformer, a simplified E2E-AD framework for the ease of scaling up, characterized by three key features: Task Parallelism (All agent, map, and planning queries direct interact with each other at each block), Sparse Representation (Task queries direct interact with raw sensor features), and Streaming Processing (Task queries are stored and passed as history information). As a result, the new framework is composed of three unified operations: task self-attention, sensor cross-attention, temporal cross-attention, which significantly reduces the complexity of system and leads to better training stability. DriveTransformer achieves state-of-the-art performance in both simulated closed-loop benchmark Bench2Drive and real world open-loop benchmark nuScenes with high FPS.

Yong Sun, Zhengjie Zhang, Junyu Shi et al.

Identification of fine-grained embryo developmental stages during In Vitro Fertilization (IVF) is crucial for assessing embryo viability. Although recent deep learning methods have achieved promising accuracy, existing discriminative models fail to utilize the distributional prior of embryonic development to improve accuracy. Moreover, their reliance on single-focal information leads to incomplete embryonic representations, making them susceptible to feature ambiguity under cell occlusions. To address these limitations, we propose EmbryoDiff, a two-stage diffusion-based framework that formulates the task as a conditional sequence denoising process. Specifically, we first train and freeze a frame-level encoder to extract robust multi-focal features. In the second stage, we introduce a Multi-Focal Feature Fusion Strategy that aggregates information across focal planes to construct a 3D-aware morphological representation, effectively alleviating ambiguities arising from cell occlusions. Building on this fused representation, we derive complementary semantic and boundary cues and design a Hybrid Semantic-Boundary Condition Block to inject them into the diffusion-based denoising process, enabling accurate embryonic stage classification. Extensive experiments on two benchmark datasets show that our method achieves state-of-the-art results. Notably, with only a single denoising step, our model obtains the best average test performance, reaching 82.8% and 81.3% accuracy on the two datasets, respectively.

Ruihan Wu, Erchi Wang, Zhiyuan Zhang et al.

Retrieval-augmented generation (RAG) enhances large language models (LLMs) by retrieving documents from an external corpus at inference time. When this corpus contains sensitive information, however, unprotected RAG systems are at risk of leaking private information. Prior work has introduced differential privacy (DP) guarantees for RAG, but only in single-query settings, which fall short of realistic usage. In this paper, we study the more practical multi-query setting and propose two DP-RAG algorithms. The first, MURAG, leverages an individual privacy filter so that the accumulated privacy loss only depends on how frequently each document is retrieved rather than the total number of queries. The second, MURAG-ADA, further improves utility by privately releasing query-specific thresholds, enabling more precise selection of relevant documents. Our experiments across multiple LLMs and datasets demonstrate that the proposed methods scale to hundreds of queries within a practical DP budget ($\varepsilon\approx10$), while preserving meaningful utility.

Tatsuki Koga, Ruihan Wu, Zhiyuan Zhang et al.

With the recent remarkable advancement of large language models (LLMs), there has been a growing interest in utilizing them in the domains with highly sensitive data that lies outside their training data. For this purpose, retrieval-augmented generation (RAG) is particularly effective -- it assists LLMs by directly providing relevant information from the external knowledge sources. However, without extra privacy safeguards, RAG outputs risk leaking sensitive information from the external data source. In this work, we explore RAG under differential privacy (DP), a formal guarantee of data privacy. The main challenge with differentially private RAG is how to generate long accurate answers within a moderate privacy budget. We address this by proposing an algorithm that smartly spends privacy budget only for the tokens that require the sensitive information and uses the non-private LLM for other tokens. Our extensive empirical evaluations reveal that our algorithm outperforms the non-RAG baseline under a reasonable privacy budget of $ε\approx 10$ across different models and datasets.

Shu Pu, Yaochen Wang, Dongping Chen et al.

Evaluating generative foundation models on open-ended multimodal understanding (MMU) and generation (MMG) tasks across diverse modalities (e.g., images, audio, video) poses significant challenges due to the complexity of cross-modal interactions. To this end, the idea of utilizing Multimodal LLMs (MLLMs) as automated judges has emerged, with encouraging results in assessing vision-language understanding tasks. Moving further, this paper extends MLLM-as-a-Judge across modalities to a unified manner by introducing two benchmarks, TaskAnything and JudgeAnything, to respectively evaluate the overall performance and judging capabilities of MLLMs across any-to-any modality tasks. Specifically, TaskAnything evaluates the MMU and MMG capabilities across 15 any-to-any modality categories, employing 1,500 queries curated from well-established benchmarks. Furthermore, JudgeAnything evaluates the judging capabilities of 5 advanced (e.g., GPT-4o and Gemini-2.0-Flash) from the perspectives of Pair Comparison and Score Evaluation, providing a standardized testbed that incorporates human judgments and detailed rubrics. Our extensive experiments reveal that while these MLLMs show promise in assessing MMU (i.e., achieving an average of 66.55% in Pair Comparison setting and 42.79% in Score Evaluation setting), they encounter significant challenges with MMG tasks (i.e., averaging only 53.37% in Pair Comparison setting and 30.05% in Score Evaluation setting), exposing cross-modality biases and hallucination issues. To address this, we present OmniArena, an automated platform for evaluating omni-models and multimodal reward models. Our work highlights the need for fairer evaluation protocols and stronger alignment with human preferences. The source code and dataset are publicly available at: https://urrealhero.github.io/judgeanythingweb/.