Li Tang

Haoyi Duan, Yan Xia, Mingze Zhou et al. · stanford

In recent years, the deployment of large-scale pre-trained models in audio-visual downstream tasks has yielded remarkable outcomes. However, these models, primarily trained on single-modality unconstrained datasets, still encounter challenges in feature extraction for multi-modal tasks, leading to suboptimal performance. This limitation arises due to the introduction of irrelevant modality-specific information during encoding, which adversely affects the performance of downstream tasks. To address this challenge, this paper proposes a novel Dual-Guided Spatial-Channel-Temporal (DG-SCT) attention mechanism. This mechanism leverages audio and visual modalities as soft prompts to dynamically adjust the parameters of pre-trained models based on the current multi-modal input features. Specifically, the DG-SCT module incorporates trainable cross-modal interaction layers into pre-trained audio-visual encoders, allowing adaptive extraction of crucial information from the current modality across spatial, channel, and temporal dimensions, while preserving the frozen parameters of large-scale pre-trained models. Experimental evaluations demonstrate that our proposed model achieves state-of-the-art results across multiple downstream tasks, including AVE, AVVP, AVS, and AVQA. Furthermore, our model exhibits promising performance in challenging few-shot and zero-shot scenarios. The source code and pre-trained models are available at https://github.com/haoyi-duan/DG-SCT.

Aoxiong Yin, Tianyun Zhong, Li Tang et al.

Most sign language translation (SLT) methods to date require the use of gloss annotations to provide additional supervision information, however, the acquisition of gloss is not easy. To solve this problem, we first perform an analysis of existing models to confirm how gloss annotations make SLT easier. We find that it can provide two aspects of information for the model, 1) it can help the model implicitly learn the location of semantic boundaries in continuous sign language videos, 2) it can help the model understand the sign language video globally. We then propose \emph{gloss attention}, which enables the model to keep its attention within video segments that have the same semantics locally, just as gloss helps existing models do. Furthermore, we transfer the knowledge of sentence-to-sentence similarity from the natural language model to our gloss attention SLT network (GASLT) to help it understand sign language videos at the sentence level. Experimental results on multiple large-scale sign language datasets show that our proposed GASLT model significantly outperforms existing methods. Our code is provided in \url{https://github.com/YinAoXiong/GASLT}.

Lei Guo, Jinyu Zhang, Li Tang et al.

Shared-account Cross-domain Sequential Recommendation (SCSR) task aims to recommend the next item via leveraging the mixed user behaviors in multiple domains. It is gaining immense research attention as more and more users tend to sign up on different platforms and share accounts with others to access domain-specific services. Existing works on SCSR mainly rely on mining sequential patterns via Recurrent Neural Network (RNN)-based models, which suffer from the following limitations: 1) RNN-based methods overwhelmingly target discovering sequential dependencies in single-user behaviors. They are not expressive enough to capture the relationships among multiple entities in SCSR. 2) All existing methods bridge two domains via knowledge transfer in the latent space, and ignore the explicit cross-domain graph structure. 3) None existing studies consider the time interval information among items, which is essential in the sequential recommendation for characterizing different items and learning discriminative representations for them. In this work, we propose a new graph-based solution, namely TiDA-GCN, to address the above challenges. Specifically, we first link users and items in each domain as a graph. Then, we devise a domain-aware graph convolution network to learn userspecific node representations. To fully account for users' domainspecific preferences on items, two effective attention mechanisms are further developed to selectively guide the message passing process. Moreover, to further enhance item- and account-level representation learning, we incorporate the time interval into the message passing, and design an account-aware self-attention module for learning items' interactive characteristics. Experiments demonstrate the superiority of our proposed method from various aspects.

Ruofan Hu, Menghui Zhu, Jieming Zhu et al.

Multimodal documents contain diverse elements, such as tables, figures, and layouts, which can complicate retrieval tasks. While current approaches typically combine dense visual embedding models with supervised rerankers to achieve high-precision retrieval, they face inherent limitations. First, the coarse-grained nature of dense embeddings tends to obfuscate explicit semantics, failing to leverage structurally salient information. Second, supervised reranking models suffer from generalization bottlenecks, as their performance heavily relies on domain-specific training data. Furthermore, existing benchmarks often lack diverse assessment dimensions and comprehensive relevance annotations, limiting reliable evaluation. To address these challenges, we propose DocRetriever, a plug-and-play framework. It enhances visual retrieval via a layout-aware sparse embedding technique, enabling effective hybrid encoding without the overhead of optical character recognition (OCR). We also introduce a generalizable reranker that leverages reasoning-augmented demonstrations and optimized sampling to improve accuracy in few-shot settings. Finally, we construct a new benchmark, MultiDocR, to enable more rigorous evaluation. Experiments across diverse benchmarks validate DocRetriever's superiority over state-of-the-art methods.

Sha Lu, Xuecheng Xu, Li Tang et al.

LiDAR based place recognition is popular for loop closure detection and re-localization. In recent years, deep learning brings improvements to place recognition by learnable feature extraction. However, these methods degenerate when the robot re-visits previous places with large perspective difference. To address the challenge, we propose DeepRING to learn the roto-translation invariant representation from LiDAR scan, so that robot visits the same place with different perspective can have similar representations. There are two keys in DeepRING: the feature is extracted from sinogram, and the feature is aggregated by magnitude spectrum. The two steps keeps the final representation with both discrimination and roto-translation invariance. Moreover, we state the place recognition as a one-shot learning problem with each place being a class, leveraging relation learning to build representation similarity. Substantial experiments are carried out on public datasets, validating the effectiveness of each proposed component, and showing that DeepRING outperforms the comparative methods, especially in dataset level generalization.

Yong Zou, Haoran Li, Fanxiao Li et al.

Recent progress in image generation models (IGMs) enables high-fidelity content creation but also amplifies risks, including the reproduction of copyrighted content and the generation of offensive content. Image Generation Model Unlearning (IGMU) mitigates these risks by removing harmful concepts without full retraining. Despite growing attention, the robustness under adversarial inputs, particularly image-side threats in black-box settings, remains underexplored. To bridge this gap, we present REFORGE, a black-box red-teaming framework that evaluates IGMU robustness via adversarial image prompts. REFORGE initializes stroke-based images and optimizes perturbations with a cross-attention-guided masking strategy that allocates noise to concept-relevant regions, balancing attack efficacy and visual fidelity. Extensive experiments across representative unlearning tasks and defenses demonstrate that REFORGE significantly improves attack success rate while achieving stronger semantic alignment and higher efficiency than involved baselines. These results expose persistent vulnerabilities in current IGMU methods and highlight the need for robustness-aware unlearning against multi-modal adversarial attacks. Our code is at: https://github.com/Imfatnoily/REFORGE.

Ling Team, Binwei Zeng, Chao Huang et al.

In this technical report, we tackle the challenges of training large-scale Mixture of Experts (MoE) models, focusing on overcoming cost inefficiency and resource limitations prevalent in such systems. To address these issues, we present two differently sized MoE large language models (LLMs), namely Ling-Lite and Ling-Plus (referred to as "Bailing" in Chinese, spelled Bǎilíng in Pinyin). Ling-Lite contains 16.8 billion parameters with 2.75 billion activated parameters, while Ling-Plus boasts 290 billion parameters with 28.8 billion activated parameters. Both models exhibit comparable performance to leading industry benchmarks. This report offers actionable insights to improve the efficiency and accessibility of AI development in resource-constrained settings, promoting more scalable and sustainable technologies. Specifically, to reduce training costs for large-scale MoE models, we propose innovative methods for (1) optimization of model architecture and training processes, (2) refinement of training anomaly handling, and (3) enhancement of model evaluation efficiency. Additionally, leveraging high-quality data generated from knowledge graphs, our models demonstrate superior capabilities in tool use compared to other models. Ultimately, our experimental findings demonstrate that a 300B MoE LLM can be effectively trained on lower-performance devices while achieving comparable performance to models of a similar scale, including dense and MoE models. Compared to high-performance devices, utilizing a lower-specification hardware system during the pre-training phase demonstrates significant cost savings, reducing computing costs by approximately 20%. The models can be accessed at https://huggingface.co/inclusionAI.

Ling Team, Anqi Shen, Baihui Li et al.

We present Ring-1T, the first open-source, state-of-the-art thinking model with a trillion-scale parameter. It features 1 trillion total parameters and activates approximately 50 billion per token. Training such models at a trillion-parameter scale introduces unprecedented challenges, including train-inference misalignment, inefficiencies in rollout processing, and bottlenecks in the RL system. To address these, we pioneer three interconnected innovations: (1) IcePop stabilizes RL training via token-level discrepancy masking and clipping, resolving instability from training-inference mismatches; (2) C3PO++ improves resource utilization for long rollouts under a token budget by dynamically partitioning them, thereby obtaining high time efficiency; and (3) ASystem, a high-performance RL framework designed to overcome the systemic bottlenecks that impede trillion-parameter model training. Ring-1T delivers breakthrough results across critical benchmarks: 93.4 on AIME-2025, 86.72 on HMMT-2025, 2088 on CodeForces, and 55.94 on ARC-AGI-1. Notably, it attains a silver medal-level result on the IMO-2025, underscoring its exceptional reasoning capabilities. By releasing the complete 1T parameter MoE model to the community, we provide the research community with direct access to cutting-edge reasoning capabilities. This contribution marks a significant milestone in democratizing large-scale reasoning intelligence and establishes a new baseline for open-source model performance.

Runze Su, Jiayin Jin, Jiacheng Li et al.

Large embedding tables are indispensable in modern recommendation systems, thanks to their ability to effectively capture and memorize intricate details of interactions among diverse entities. As we explore integrating large embedding tables into Pinterest's ads ranking models, we encountered not only common challenges such as sparsity and scalability, but also several obstacles unique to our context. Notably, our initial attempts to train large embedding tables from scratch resulted in neutral metrics. To tackle this, we introduced a novel multi-faceted pretraining scheme that incorporates multiple pretraining algorithms. This approach greatly enriched the embedding tables and resulted in significant performance improvements. As a result, the multi-faceted large embedding tables bring great performance gain on both the Click-Through Rate (CTR) and Conversion Rate (CVR) domains. Moreover, we designed a CPU-GPU hybrid serving infrastructure to overcome GPU memory limits and elevate the scalability. This framework has been deployed in the Pinterest Ads system and achieved 1.34% online CPC reduction and 2.60% CTR increase with neutral end-to-end latency change.

Li Tang, Peter A. Torrione, Cihat Eldeniz et al.

Ground penetrating radar (GPR) provides a promising technology for accurate subsurface object detection. In particular, it has shown promise for detecting landmines with low metal content. However, the ground bounce (GB) that is present in GPR data, which is caused by the dielectric discontinuity between soil and air, is a major source of interference and degrades landmine detection performance. To mitigate this interference, GB tracking algorithms formulated using both a Kalman filter (KF) and a particle filter (PF) framework are proposed. In particular, the location of the GB in the radar signal is modeled as the hidden state in a stochastic system for the PF approach. The observations are the 2D radar images, which arrive scan by scan along the down-track direction. An initial training stage sets parameters automatically to accommodate different ground and weather conditions. The features associated with the GB description are updated adaptively with the arrival of new data. The prior distribution for a given location is predicted by propagating information from two adjacent channels/scans, which ensures that the overall GB surface remains smooth. The proposed algorithms are verified in experiments utilizing real data, and their performances are compared with other GB tracking approaches. We demonstrate that improved GB tracking contributes to improved performance for the landmine detection problem.

Zehan Wang, Yang Zhao, Xize Cheng et al.

Multi-modal Contrastive Representation learning aims to encode different modalities into a semantically aligned shared space. This paradigm shows remarkable generalization ability on numerous downstream tasks across various modalities. However, the reliance on massive high-quality data pairs limits its further development on more modalities. This paper proposes a novel training-efficient method for learning MCR without paired data called Connecting Multi-modal Contrastive Representations (C-MCR). Specifically, given two existing MCRs pre-trained on (A, B) and (B, C) modality pairs, we project them to a new space and use the data from the overlapping modality B to aligning the two MCRs in the new space. Meanwhile, since the modality pairs (A, B) and (B, C) are already aligned within each MCR, the connection learned by overlapping modality can also be transferred to non-overlapping modality pair (A, C). To unleash the potential of C-MCR, we further introduce a semantic-enhanced inter- and intra-MCR connection method. We first enhance the semantic consistency and completion of embeddings across different modalities for more robust alignment. Then we utilize the inter-MCR alignment to establish the connection, and employ the intra-MCR alignment to better maintain the connection for inputs from non-overlapping modalities. To demonstrate the effectiveness of C-MCR, we connect CLIP and CLAP via texts to derive audio-visual representations, and integrate CLIP and ULIP via images for 3D-language representations. Remarkably, without using any paired data, C-MCR for audio-visual achieves state-of-the-art performance on audio-image retrieval, audio-visual source localization, and counterfactual audio-image recognition tasks. Furthermore, C-MCR for 3D-language also attains advanced zero-shot 3D point cloud classification accuracy on ModelNet40.

Lei Guo, Li Tang, Tong Chen et al.

Shared-account Cross-domain Sequential recommendation (SCSR) is the task of recommending the next item based on a sequence of recorded user behaviors, where multiple users share a single account, and their behaviours are available in multiple domains. Existing work on solving SCSR mainly relies on mining sequential patterns via RNN-based models, which are not expressive enough to capture the relationships among multiple entities. Moreover, all existing algorithms try to bridge two domains via knowledge transfer in the latent space, and the explicit cross-domain graph structure is unexploited. In this work, we propose a novel graph-based solution, namely DA-GCN, to address the above challenges. Specifically, we first link users and items in each domain as a graph. Then, we devise a domain-aware graph convolution network to learn user-specific node representations. To fully account for users' domain-specific preferences on items, two novel attention mechanisms are further developed to selectively guide the message passing process. Extensive experiments on two real-world datasets are conducted to demonstrate the superiority of our DA-GCN method.

Xiaqing Ding, Yue Wang, Li Tang et al.

Visual localization occupies an important position in many areas such as Augmented Reality, robotics and 3D reconstruction. The state-of-the-art visual localization methods perform pose estimation using geometry based solver within the RANSAC framework. However, these methods require accurate pixel-level matching at high image resolution, which is hard to satisfy under significant changes from appearance, dynamics or perspective of view. End-to-end learning based regression networks provide a solution to circumvent the requirement for precise pixel-level correspondences, but demonstrate poor performance towards cross-scene generalization. In this paper, we explicitly add a learnable matching layer within the network to isolate the pose regression solver from the absolute image feature values, and apply dimension regularization on both the correlation feature channel and the image scale to further improve performance towards generalization and large viewpoint change. We implement this dimension regularization strategy within a two-layer pyramid based framework to regress the localization results from coarse to fine. In addition, the depth information is fused for absolute translational scale recovery. Through experiments on real world RGBD datasets we validate the effectiveness of our design in terms of improving both generalization performance and robustness towards viewpoint change, and also show the potential of regression based visual localization networks towards challenging occasions that are difficult for geometry based visual localization methods.

Huan Yin, Yue Wang, Li Tang et al.

Radar and lidar, provided by two different range sensors, each has pros and cons of various perception tasks on mobile robots or autonomous driving. In this paper, a Monte Carlo system is used to localize the robot with a rotating radar sensor on 2D lidar maps. We first train a conditional generative adversarial network to transfer raw radar data to lidar data, and achieve reliable radar points from generator. Then an efficient radar odometry is included in the Monte Carlo system. Combining the initial guess from odometry, a measurement model is proposed to match the radar data and prior lidar maps for final 2D positioning. We demonstrate the effectiveness of the proposed localization framework on the public multi-session dataset. The experimental results show that our system can achieve high accuracy for long-term localization in outdoor scenes.

Huifang Ma, Yue Wang, Rong Xiong et al.

Recent research on automotive driving developed an efficient end-to-end learning mode that directly maps visual input to control commands. However, it models distinct driving variations in a single network, which increases learning complexity and is less adaptive for modular integration. In this paper, we re-investigate human's driving style and propose to learn an intermediate driving intention region to relax difficulties in end-to-end approach. The intention region follows both road structure in image and direction towards goal in public route planner, which addresses visual variations only and figures out where to go without conventional precise localization. Then the learned visual intention is projected on vehicle local coordinate and fused with reliable obstacle perception to render a navigation score map widely used for motion planning. The core of the proposed system is a weakly-supervised cGAN-LSTM model trained to learn driving intention from human demonstration. The adversarial loss learns from limited demonstration data with one local planned route and enables reasoning of multi-modal behavior with diverse routes while testing. Comprehensive experiments are conducted with real-world datasets. Results show the proposed paradigm can produce more consistent motion commands with human demonstration, and indicates better reliability and robustness to environment change.

Huifang Ma, Yue Wang, Li Tang et al.

Autonomous navigation based on precise localization has been widely developed in both academic research and practical applications. The high demand for localization accuracy has been essential for safe robot planing and navigation while it makes the current geometric solutions less robust to environmental changes. Recent research on end-to-end methods handle raw sensory data with forms of navigation instructions and directly output the command for robot control. However, the lack of intermediate semantics makes the system more rigid and unstable for practical use. To explore these issues, this paper proposes an innovate navigation framework based on the GPS-level localization, which takes the raw perception data with publicly accessible navigation maps to produce an intermediate navigation cost map that allows subsequent flexible motion planning. A deterministic conditional adversarial network is adopted in our method to generate visual goal-directed paths under diverse navigation conditions. The adversarial loss avoids the pixel-level annotation and enables a weakly supervised training strategy to implicitly learn both of the traffic semantics in image perceptions and the planning intentions in navigation instructions. The navigation cost map is then rendered from the goal-directed path and the concurrently collected laser data, indicating the way towards the destination. Comprehensive experiments have been conducted with a real vehicle running in our campus and the results have verified the robustness to localization error of the proposed navigation system.

Bo Fu, Yue Wang, Xiaqing Ding et al.

LiDAR-camera calibration is a precondition for many heterogeneous systems that fuse data from LiDAR and camera. However, the constraint from common field of view and the requirement for strict time synchronization make the calibration a challenging problem. In this paper, we propose a novel LiDAR-camera calibration method aiming to eliminate these two constraints. Specifically, we capture a scan of 3D LiDAR when both the environment and the sensors are stationary, then move the camera to reconstruct the 3D environment using the sequentially obtained images. Finally, we align 3D visual points to the laser scan based on tightly couple graph optimization method to calculate the extrinsic parameters between LiDAR and camera. Under this design, the configuration of these two sensors are free from the common field of view constraint owing to the extended view from the moving camera. And we also eliminate the requirement for strict time synchronization as we only use the single scan of laser data when the sensors are stationary. We theoretically derive the conditions of minimal observability for our method and prove that the accuracy of calibration is improved by collecting more observations from multiple scattered calibration targets. We validate our method on both simulation platform and real-world datasets. Experiments show that our method achieves higher accuracy than other comparable methods, which is in accordance with our theoretical analysis. In addition, the proposed method is beneficial to not only plane measurement error based chessboard, but also other point measurement error based calibration targets, such as boxes and polygonal boards.

Xiaqing Ding, Yue Wang, Li Tang et al.

Visual localization is one of the primary capabilities for mobile robots. Long-term visual localization in real time is particularly challenging, in which the robot is required to efficiently localize itself using visual data where appearance may change significantly over time. In this paper, we propose a cloud-based visual localization system targeting at long-term localization in real time. On the robot, we employ two estimators to achieve accurate and real-time performance. One is a sliding-window based visual inertial odometry, which integrates constraints from consecutive observations and self-motion measurements, as well as the constraints induced by localization on the cloud. This estimator builds a local visual submap as the virtual observation which is then sent to the cloud as new localization constraints. The other one is a delayed state Extended Kalman Filter to fuse the pose of the robot localized from the cloud, the local odometry and the high-frequency inertial measurements. On the cloud, we propose a longer sliding-window based localization method to aggregate the virtual observations for larger field of view, leading to more robust alignment between virtual observations and the map. Under this architecture, the robot can achieve drift-free and real-time localization using onboard resources even in a network with limited bandwidth, high latency and existence of package loss, which enables the autonomous navigation in real-world environment. We evaluate the effectiveness of our system on a dataset with challenging seasonal and illuminative variations. We further validate the robustness of the system under challenging network conditions.

Li Tang, Konstantinos Konstantinidis, Aditya Ramamoorthy

Distributed matrix multiplication is widely used in several scientific domains. It is well recognized that computation times on distributed clusters are often dominated by the slowest workers (called stragglers). Recent work has demonstrated that straggler mitigation can be viewed as a problem of designing erasure codes. For matrices $\mathbf A$ and $\mathbf B$, the technique essentially maps the computation of $\mathbf A^T \mathbf B$ into the multiplication of smaller (coded) submatrices. The stragglers are treated as erasures in this process. The computation can be completed as long as a certain number of workers (called the recovery threshold) complete their assigned tasks. We present a novel coding strategy for this problem when the absolute values of the matrix entries are sufficiently small. We demonstrate a tradeoff between the assumed absolute value bounds on the matrix entries and the recovery threshold. At one extreme, we are optimal with respect to the recovery threshold and on the other extreme, we match the threshold of prior work. Experimental results on cloud-based clusters validate the benefits of our method.

Xiaqing Ding, Yue Wang, Huan Yin et al.

Map construction in large scale outdoor environment is of importance for robots to robustly fulfill their tasks. Massive sessions of data should be merged to distinguish low dynamics in the map, which otherwise might debase the performance of localization and navigation algorithms. In this paper we propose a method for multi-session map construction in large scale outdoor environment using 3D LiDAR. To efficiently align the maps from different sessions, a laser-based loop closure detection method is integrated and the sequential information within the submaps is utilized for higher robustness. Furthermore, a dynamic detection method is proposed to detect dynamics in the overlapping areas among sessions of maps. We test the method in the real-world environment with a VLP-16 Velodyne LiDAR and the experimental results prove the validity and robustness of the proposed method.

Xiaqing Ding, Yue Wang, Dongxuan Li et al.

Long-term visual localization in outdoor environment is a challenging problem, especially faced with the cross-seasonal, bi-directional tasks and changing environment. In this paper we propose a novel visual inertial localization framework that localizes against the LiDAR-built map. Based on the geometry information of the laser map, a hybrid bundle adjustment framework is proposed, which estimates the poses of the cameras with respect to the prior laser map as well as optimizes the state variables of the online visual inertial odometry system simultaneously. For more accurate cross-modal data association, the laser map is optimized using multi-session laser and visual data to extract the salient and stable subset for localization. To validate the efficiency of the proposed method, we collect data in south part of our campus in different seasons, along the same and opposite-direction route. In all sessions of localization data, our proposed method gives satisfactory results, and shows the superiority of the hybrid bundle adjustment and map optimization.

Huan Yin, Li Tang, Xiaqing Ding et al.

Global localization in 3D point clouds is a challenging problem of estimating the pose of vehicles without any prior knowledge. In this paper, a solution to this problem is presented by achieving place recognition and metric pose estimation in the global prior map. Specifically, we present a semi-handcrafted representation learning method for LiDAR point clouds using siamese LocNets, which states the place recognition problem to a similarity modeling problem. With the final learned representations by LocNet, a global localization framework with range-only observations is proposed. To demonstrate the performance and effectiveness of our global localization system, KITTI dataset is employed for comparison with other algorithms, and also on our long-time multi-session datasets for evaluation. The result shows that our system can achieve high accuracy.

Qianhui Luo, Huifang Ma, Yue Wang et al.

This paper aims at developing a faster and a more accurate solution to the amodal 3D object detection problem for indoor scenes. It is achieved through a novel neural network that takes a pair of RGB-D images as the input and delivers oriented 3D bounding boxes as the output. The network, named 3D-SSD, composed of two parts: hierarchical feature fusion and multi-layer prediction. The hierarchical feature fusion combines appearance and geometric features from RGB-D images while the multi-layer prediction utilizes multi-scale features for object detection. As a result, the network can exploit 2.5D representations in a synergetic way to improve the accuracy and efficiency. The issue of object sizes is addressed by attaching a set of 3D anchor boxes with varying sizes to every location of the prediction layers. At the end stage, the category scores for 3D anchor boxes are generated with adjusted positions, sizes and orientations respectively, leading to the final detections using non-maximum suppression. In the training phase, the positive samples are identified with the aid of 2D ground truth to avoid the noisy estimation of depth from raw data, which guide to a better converged model. Experiments performed on the challenging SUN RGB-D dataset show that our algorithm outperforms the state-of-the-art Deep Sliding Shape by 10.2% mAP and 88x faster. Further, experiments also suggest our approach achieves comparable accuracy and is 386x faster than the state-of-art method on the NYUv2 dataset even with a smaller input image size.