Haodong Wang

Kuntai Du, Qizheng Zhang, Anton Arapin et al. · stanford

With more videos being recorded by edge sensors (cameras) and analyzed by computer-vision deep neural nets (DNNs), a new breed of video streaming systems has emerged, with the goal to compress and stream videos to remote servers in real time while preserving enough information to allow highly accurate inference by the server-side DNNs. An ideal design of the video streaming system should simultaneously meet three key requirements: (1) low latency of encoding and streaming, (2) high accuracy of server-side DNNs, and (3) low compute overheads on the camera. Unfortunately, despite many recent efforts, such video streaming system has hitherto been elusive, especially when serving advanced vision tasks such as object detection or semantic segmentation. This paper presents AccMPEG, a new video encoding and streaming system that meets all the three requirements. The key is to learn how much the encoding quality at each (16x16) macroblock can influence the server-side DNN accuracy, which we call accuracy gradient. Our insight is that these macroblock-level accuracy gradient can be inferred with sufficient precision by feeding the video frames through a cheap model. AccMPEG provides a suite of techniques that, given a new server-side DNN, can quickly create a cheap model to infer the accuracy gradient on any new frame in near realtime. Our extensive evaluation of AccMPEG on two types of edge devices (one Intel Xeon Silver 4100 CPU or NVIDIA Jetson Nano) and three vision tasks (six recent pre-trained DNNs) shows that AccMPEG (with the same camera-side compute resources) can reduce the end-to-end inference delay by 10-43% without hurting accuracy compared to the state-of-the-art baselines

Rajrup Ghosh, Haodong Wang, Haoran Hong et al.

Dynamic 3D Gaussian Splatting (3DGS) holds great promise as a 3D video streaming technology since it can represent complex 3D scenes with high fidelity. In this approach, every frame in a 3D video represents the environment as a collection of Gaussians with position and other attributes such as scale, rotation, opacity, and color. Frames capture fine details, permit views from any arbitrary perspective, but are an order of magnitude, or more, larger than 2D video frames. A line of recent work has explored how to compress dynamic 3DGS frames, but these approaches are often slow, in part because their compression techniques are not amenable to efficient acceleration. GS-NFS accelerates dynamic 3DGS compression and decompression on a GPU, to the point where it can encode and decode at full frame rate. It achieves this by developing novel GPU-based parallelizations of existing algorithms for encoding both positions and attributes of Gaussians. As a result, it is 1-2 orders of magnitude faster than the state-of-the-art in encoding and decoding a frame, while offering competitive compression performance and rendering quality.

Kuntai Du, Yuhan Liu, Yitian Hao et al. · stanford

Deep learning inference on streaming media data, such as object detection in video or LiDAR feeds and text extraction from audio waves, is now ubiquitous. To achieve high inference accuracy, these applications typically require significant network bandwidth to gather high-fidelity data and extensive GPU resources to run deep neural networks (DNNs). While the high demand for network bandwidth and GPU resources could be substantially reduced by optimally adapting the configuration knobs, such as video resolution and frame rate, current adaptation techniques fail to meet three requirements simultaneously: adapt configurations (i) with minimum extra GPU or bandwidth overhead; (ii) to reach near-optimal decisions based on how the data affects the final DNN's accuracy, and (iii) do so for a range of configuration knobs. This paper presents OneAdapt, which meets these requirements by leveraging a gradient-ascent strategy to adapt configuration knobs. The key idea is to embrace DNNs' differentiability to quickly estimate the accuracy's gradient to each configuration knob, called AccGrad. Specifically, OneAdapt estimates AccGrad by multiplying two gradients: InputGrad (i.e. how each configuration knob affects the input to the DNN) and DNNGrad (i.e. how the DNN input affects the DNN inference output). We evaluate OneAdapt across five types of configurations, four analytic tasks, and five types of input data. Compared to state-of-the-art adaptation schemes, OneAdapt cuts bandwidth usage and GPU usage by 15-59% while maintaining comparable accuracy or improves accuracy by 1-5% while using equal or fewer resources.

Haodong Wang, Junjie Liu, Zicong Hong et al.

4-bit quantization reduces the memory footprint and latency of large language model inference, but its aggressive precision reduction can severely degrade accuracy. Prior methods address this by decomposing each weight matrix into two components (e.g., via singular value decomposition) and quantizing them separately, assigning the bulk of values to a low-precision residual component while handling outliers with a high-precision low-rank component. However, such decompositions are designed to minimize the real-valued energy of the residual, rather than the post-quantization error of the residual and low-rank components. We propose TwinQuant, a 4-bit quantization framework that learns quantization-friendly decomposed subspaces and jointly reshapes both the low-rank and residual components. TwinQuant learns component-specific transformations via a joint optimization over the Stiefel and general linear manifolds, flattening their distributions and reducing dynamic-range imbalance. To enable efficient end-to-end execution, we further design a fused dual-component kernel that pipelines the two-stage low-rank computation on-chip and merges both components with a single epilogue, avoiding intermediate global-memory traffic. Across LLaMA3 and Qwen3 models, TwinQuant preserves near-FP16 accuracy and delivers up to $1.8\times$ end-to-end speedup over an FP16 baseline.

Changli Wu, Haodong Wang, Jiayi Ji et al.

Most existing 3D referring expression segmentation (3DRES) methods rely on dense, high-quality point clouds, while real-world agents such as robots and mobile phones operate with only a few sparse RGB views and strict latency constraints. We introduce Multi-view 3D Referring Expression Segmentation (MV-3DRES), where the model must recover scene structure and segment the referred object directly from sparse multi-view images. Traditional two-stage pipelines, which first reconstruct a point cloud and then perform segmentation, often yield low-quality geometry, produce coarse or degraded target regions, and run slowly. We propose the Multimodal Visual Geometry Grounded Transformer (MVGGT), an efficient end-to-end framework that integrates language information into sparse-view geometric reasoning through a dual-branch design. Training in this setting exposes a critical optimization barrier, termed Foreground Gradient Dilution (FGD), where sparse 3D signals lead to weak supervision. To resolve this, we introduce Per-view No-target Suppression Optimization (PVSO), which provides stronger and more balanced gradients across views, enabling stable and efficient learning. To support consistent evaluation, we build MVRefer, a benchmark that defines standardized settings and metrics for MV-3DRES. Experiments show that MVGGT establishes the first strong baseline and achieves both high accuracy and fast inference, outperforming existing alternatives. The code is available at https://mvggt.github.io/.

Yilun Jin, Zheng Li, Chenwei Zhang et al.

Online shopping is a complex multi-task, few-shot learning problem with a wide and evolving range of entities, relations, and tasks. However, existing models and benchmarks are commonly tailored to specific tasks, falling short of capturing the full complexity of online shopping. Large Language Models (LLMs), with their multi-task and few-shot learning abilities, have the potential to profoundly transform online shopping by alleviating task-specific engineering efforts and by providing users with interactive conversations. Despite the potential, LLMs face unique challenges in online shopping, such as domain-specific concepts, implicit knowledge, and heterogeneous user behaviors. Motivated by the potential and challenges, we propose Shopping MMLU, a diverse multi-task online shopping benchmark derived from real-world Amazon data. Shopping MMLU consists of 57 tasks covering 4 major shopping skills: concept understanding, knowledge reasoning, user behavior alignment, and multi-linguality, and can thus comprehensively evaluate the abilities of LLMs as general shop assistants. With Shopping MMLU, we benchmark over 20 existing LLMs and uncover valuable insights about practices and prospects of building versatile LLM-based shop assistants. Shopping MMLU can be publicly accessed at https://github.com/KL4805/ShoppingMMLU. In addition, with Shopping MMLU, we host a competition in KDD Cup 2024 with over 500 participating teams. The winning solutions and the associated workshop can be accessed at our website https://amazon-kddcup24.github.io/.

Jian Lin, Jiazhi Mi, Zicong Hong et al.

Supporting long-context LLMs is challenging due to the substantial memory demands of the key-value (KV) cache. Existing offloading systems store the full cache in host memory and selectively fetch critical entries during decoding, but this strategy quickly hits a ceiling: sparsity cannot be pushed further without degrading accuracy. As a result, when context length and batch size grow, the volume of KV transfers rises sharply and becomes the dominant source of decoding latency. We present KVDrive, a holistic multi-tier KV cache management system spanning GPU memory, host DRAM, and SSD. Unlike prior work that pursues greater sparsity through algorithmic refinements, KVDrive tackles the problem from a systems perspective - jointly orchestrating cache placement, pipeline scheduling, and cross-tier coordination to sustain high-throughput inference under tight GPU budgets. KVDrive advances three fundamental capabilities: it adapts cache management to attention behavior to maximize reuse and minimize redundant data movement; it restructures the decoding pipeline to overlap I/O- and CPU/GPU compute-bound stages, eliminating stalls across heterogeneous resources; and it harmonizes data movement across memory tiers to unlock scalable long-context inference far beyond GPU and DRAM limits. We have implemented a fully functional prototype of KVDrive and evaluated it on long-context benchmarks with popular LLMs. The system achieves up to 1.74x higher throughput compared to state-of-the-art works while preserving accuracy.

Zile Wang, Qianli Liu, Kaibin Guo et al.

Deploying large language model (LLM) on edge device enables personalized LLM agents for various users. The growing availability of diverse personalized agents presents a unique opportunity for peer-to-peer (P2P) collaboration, wherein each user can delegate tasks beyond the local agent's expertise to remote agents more suited for the specific query. This paper introduces PPAI, the first personalized LLM agent interoperability system, which enables users to collaborate with each other based on agent specialization. However, the ever-changing pool of agents and their interchangeable capacity introduce new challenges when it comes to matching queries to agents and balancing loads, compared with existing P2P systems. Therefore, we propose a scalable query-agent pair scoring mechanism based on prototypes to identify suitable agents within a P2P network with churn. Moreover, we propose a multi-agent interoperability Bayesian game to balance local demand and global efficiency, when changes in remote agent load occur too quickly to be observed. Finally, we implement a prototype of PPAI and demonstrate that it substantially broadens the range of tasks that could be carried out while maintaining load balance. On average, it achieves an accuracy improvement of up to 7.96% across multiple tasks, while reducing latency by 16.34% compared to the baseline.

Zhihan Zhang, Shiyang Li, Zixuan Zhang et al.

The instruction hierarchy, which establishes a priority order from system messages to user messages, conversation history, and tool outputs, is essential for ensuring consistent and safe behavior in language models (LMs). Despite its importance, this topic receives limited attention, and there is a lack of comprehensive benchmarks for evaluating models' ability to follow the instruction hierarchy. We bridge this gap by introducing IHEval, a novel benchmark comprising 3,538 examples across nine tasks, covering cases where instructions in different priorities either align or conflict. Our evaluation of popular LMs highlights their struggle to recognize instruction priorities. All evaluated models experience a sharp performance decline when facing conflicting instructions, compared to their original instruction-following performance. Moreover, the most competitive open-source model only achieves 48% accuracy in resolving such conflicts. Our results underscore the need for targeted optimization in the future development of LMs.

Zhaoxuan Tan, Zheng Li, Tianyi Liu et al.

Learning from preference feedback is essential for aligning large language models (LLMs) with human values and improving the quality of generated responses. However, existing preference learning methods rely heavily on curated data from humans or advanced LLMs, which is costly and difficult to scale. In this work, we present PUGC, a novel framework that leverages implicit human Preferences in unlabeled User-Generated Content (UGC) to generate preference data. Although UGC is not explicitly created to guide LLMs in generating human-preferred responses, it often reflects valuable insights and implicit preferences from its creators that has the potential to address readers' questions. PUGC transforms UGC into user queries and generates responses from the policy model. The UGC is then leveraged as a reference text for response scoring, aligning the model with these implicit preferences. This approach improves the quality of preference data while enabling scalable, domain-specific alignment. Experimental results on Alpaca Eval 2 show that models trained with DPO and PUGC achieve a 9.37% performance improvement over traditional methods, setting a 35.93% state-of-the-art length-controlled win rate using Mistral-7B-Instruct. Further studies highlight gains in reward quality, domain-specific alignment effectiveness, robustness against UGC quality, and theory of mind capabilities. Our code and dataset are available at https://zhaoxuan.info/PUGC.github.io/

Chi Han, Xin Liu, Haodong Wang et al.

Despite significant achievements in improving the instruction-following capabilities of large language models (LLMs), the ability to process multiple potentially entangled or conflicting instructions remains a considerable challenge. Real-world scenarios often require consistency across multiple instructions over time, such as secret privacy, personal preferences, and prioritization, which demand sophisticated abilities to integrate multiple turns and carefully balance competing objectives when instructions intersect or conflict. This work presents a systematic investigation of LLMs' capabilities in handling multiple turns of instructions, covering three levels of difficulty: (1) retrieving information from instructions, (2) tracking and reasoning across turns, and (3) resolving conflicts among instructions. We construct MultiTurnInstruct~with $\sim$1.1K high-quality multi-turn conversations through the human-in-the-loop approach and result in nine capability categories, including statics and dynamics, reasoning, and multitasking. Our finding reveals an intriguing trade-off between different capabilities. While GPT models demonstrate superior memorization, they show reduced effectiveness in privacy-protection tasks requiring selective information withholding. Larger models exhibit stronger reasoning capabilities but still struggle with resolving conflicting instructions. Importantly, these performance gaps cannot be attributed solely to information loss, as models demonstrate strong BLEU scores on memorization tasks. Still, their attention mechanisms fail to integrate multiple related instructions effectively. These findings highlight critical areas for improvement in complex real-world tasks involving multi-turn instructions. Data and codes are released at https://github.com/Glaciohound/Multi-Turn-Instruct.

Haoran Gong, Haodong Wang, Di Wang

Semantic segmentation of large-scale outdoor point clouds is of significant importance in environment perception and scene understanding. However, this task continues to present a significant research challenge, due to the inherent complexity of outdoor objects and their diverse distributions in real-world environments. In this study, we propose the Multilateral Cascading Network (MCNet) designed to address this challenge. The model comprises two key components: a Multilateral Cascading Attention Enhancement (MCAE) module, which facilitates the learning of complex local features through multilateral cascading operations; and a Point Cross Stage Partial (P-CSP) module, which fuses global and local features, thereby optimizing the integration of valuable feature information across multiple scales. Our proposed method demonstrates superior performance relative to state-of-the-art approaches across two widely recognized benchmark datasets: Toronto3D and SensatUrban. Especially on the city-scale SensatUrban dataset, our results surpassed the current best result by 2.1\% in overall mIoU and yielded an improvement of 15.9\% on average for small-sample object categories comprising less than 2\% of the total samples, in comparison to the baseline method.

Jing Yang Lee, Hamed Bonab, Nasser Zalmout et al.

Large Language Models (LLMs) are increasingly employed in multi-turn conversational tasks, yet their pre-training data predominantly consists of continuous prose, creating a potential mismatch between required capabilities and training paradigms. We introduce a novel approach to address this discrepancy by synthesizing conversational data from existing text corpora. We present a pipeline that transforms a cluster of multiple related documents into an extended multi-turn, multi-topic information-seeking dialogue. Applying our pipeline to Wikipedia articles, we curate DocTalk, a multi-turn pre-training dialogue corpus consisting of over 730k long conversations. We hypothesize that exposure to such synthesized conversational structures during pre-training can enhance the fundamental multi-turn capabilities of LLMs, such as context memory and understanding. Empirically, we show that incorporating DocTalk during pre-training results in up to 40% gain in context memory and understanding, without compromising base performance. DocTalk is available at https://huggingface.co/datasets/AmazonScience/DocTalk.

Haodong Wang, Chongyu Wang, Yinghui Quan et al.

Expanding the receptive field in a deep learning model for large-scale 3D point cloud segmentation is an effective technique for capturing rich contextual information, which consequently enhances the network's ability to learn meaningful features. However, this often leads to increased computational complexity and risk of overfitting, challenging the efficiency and effectiveness of the learning paradigm. To address these limitations, we propose the Local Split Attention Pooling (LSAP) mechanism to effectively expand the receptive field through a series of local split operations, thus facilitating the acquisition of broader contextual knowledge. Concurrently, it optimizes the computational workload associated with attention-pooling layers to ensure a more streamlined processing workflow. Based on LSAP, a Parallel Aggregation Enhancement (PAE) module is introduced to enable parallel processing of data using both 2D and 3D neighboring information to further enhance contextual representations within the network. In light of the aforementioned designs, we put forth a novel framework, designated as LSNet, for large-scale point cloud semantic segmentation. Extensive evaluations demonstrated the efficacy of seamlessly integrating the proposed PAE module into existing frameworks, yielding significant improvements in mean intersection over union (mIoU) metrics, with a notable increase of up to 11%. Furthermore, LSNet demonstrated superior performance compared to state-of-the-art semantic segmentation networks on three benchmark datasets, including S3DIS, Toronto3D, and SensatUrban. It is noteworthy that our method achieved a substantial speedup of approximately 38.8% compared to those employing similar-sized receptive fields, which serves to highlight both its computational efficiency and practical utility in real-world large-scale scenes.

Harry Walsh, Ed Fish, Ozge Mercanoglu Sincan et al.

Sign Language Production (SLP) is the task of generating sign language video from spoken language inputs. The field has seen a range of innovations over the last few years, with the introduction of deep learning-based approaches providing significant improvements in the realism and naturalness of generated outputs. However, the lack of standardized evaluation metrics for SLP approaches hampers meaningful comparisons across different systems. To address this, we introduce the first Sign Language Production Challenge, held as part of the third SLRTP Workshop at CVPR 2025. The competition's aims are to evaluate architectures that translate from spoken language sentences to a sequence of skeleton poses, known as Text-to-Pose (T2P) translation, over a range of metrics. For our evaluation data, we use the RWTH-PHOENIX-Weather-2014T dataset, a German Sign Language - Deutsche Gebardensprache (DGS) weather broadcast dataset. In addition, we curate a custom hidden test set from a similar domain of discourse. This paper presents the challenge design and the winning methodologies. The challenge attracted 33 participants who submitted 231 solutions, with the top-performing team achieving BLEU-1 scores of 31.40 and DTW-MJE of 0.0574. The winning approach utilized a retrieval-based framework and a pre-trained language model. As part of the workshop, we release a standardized evaluation network, including high-quality skeleton extraction-based keypoints establishing a consistent baseline for the SLP field, which will enable future researchers to compare their work against a broader range of methods.

Haodong Wang, Qihua Zhou, Zicong Hong et al.

The mixture of experts (MoE) model is a sparse variant of large language models (LLMs), designed to hold a better balance between intelligent capability and computational overhead. Despite its benefits, MoE is still too expensive to deploy on resource-constrained edge devices, especially with the demands of on-device inference services. Recent research efforts often apply model compression techniques, such as quantization, pruning and merging, to restrict MoE complexity. Unfortunately, due to their predefined static model optimization strategies, they cannot always achieve the desired quality-overhead trade-off when handling multiple requests, finally degrading the on-device quality of service. These limitations motivate us to propose the D$^2$MoE, an algorithm-system co-design framework that matches diverse task requirements by dynamically allocating the most proper bit-width to each expert. Specifically, inspired by the nested structure of matryoshka dolls, we propose the matryoshka weight quantization (MWQ) to progressively compress expert weights in a bit-nested manner and reduce the required runtime memory. On top of it, we further optimize the I/O-computation pipeline and design a heuristic scheduling algorithm following our hottest-expert-bit-first (HEBF) principle, which maximizes the expert parallelism between I/O and computation queue under constrained memory budgets, thus significantly reducing the idle temporal bubbles waiting for the experts to load. Evaluations on real edge devices show that D$^2$MoE improves the overall inference throughput by up to 1.39$\times$ and reduces the peak memory footprint by up to 53% over the latest on-device inference frameworks, while still preserving comparable serving accuracy as its INT8 counterparts.

Haodong Wang, Tao Zhuo, Xiuwei Zhang et al.

Achieving pixel-level registration between SAR and optical images remains a challenging task due to their fundamentally different imaging mechanisms and visual characteristics. Although deep learning has achieved great success in many cross-modal tasks, its performance on SAR-Optical registration tasks is still unsatisfactory. Gradient-based information has traditionally played a crucial role in handcrafted descriptors by highlighting structural differences. However, such gradient cues have not been effectively leveraged in deep learning frameworks for SAR-Optical image matching. To address this gap, we propose SOMA, a dense registration framework that integrates structural gradient priors into deep features and refines alignment through a hybrid matching strategy. Specifically, we introduce the Feature Gradient Enhancer (FGE), which embeds multi-scale, multi-directional gradient filters into the feature space using attention and reconstruction mechanisms to boost feature distinctiveness. Furthermore, we propose the Global-Local Affine-Flow Matcher (GLAM), which combines affine transformation and flow-based refinement within a coarse-to-fine architecture to ensure both structural consistency and local accuracy. Experimental results demonstrate that SOMA significantly improves registration precision, increasing the CMR@1px by 12.29% on the SEN1-2 dataset and 18.50% on the GFGE_SO dataset. In addition, SOMA exhibits strong robustness and generalizes well across diverse scenes and resolutions.

Haoxuan Li, Chenxu Wei, Haodong Wang et al.

Change detection typically involves identifying regions with changes between bitemporal images taken at the same location. Besides significant changes, slow changes in bitemporal images are also important in real-life scenarios. For instance, weak changes often serve as precursors to major hazards in scenarios like slopes, dams, and tailings ponds. Therefore, designing a change detection network that simultaneously detects slow and fast changes presents a novel challenge. In this paper, to address this challenge, we propose a change detection network named Flow-CDNet, consisting of two branches: optical flow branch and binary change detection branch. The first branch utilizes a pyramid structure to extract displacement changes at multiple scales. The second one combines a ResNet-based network with the optical flow branch's output to generate fast change outputs. Subsequently, to supervise and evaluate this new change detection framework, a self-built change detection dataset Flow-Change, a loss function combining binary tversky loss and L2 norm loss, along with a new evaluation metric called FEPE are designed. Quantitative experiments conducted on Flow-Change dataset demonstrated that our approach outperforms the existing methods. Furthermore, ablation experiments verified that the two branches can promote each other to enhance the detection performance.