Ziyi Xu

Yi Pan, Yile Gu, Jinbin Luo et al.

Intra-device parallelism addresses resource under-utilization in ML inference and training by overlapping the execution of operators with different resource usage. However, its wide adoption is hindered by a fundamental conflict with the static, sequential programming model of existing frameworks. Integrating these strategies requires invasive, model-specific code overhauls, representing an intractable engineering cost. This is further amplified by the high sensitivity of strategies to execution contexts (e.g., workload, model architecture, hardware), forcing developers to implement and maintain multiple specialized solutions. To address this, we propose DynaFlow, a framework that enables the transparent and flexible integration of intra-device parallelism by decoupling the logical model definition from the physical execution schedule. DynaFlow introduces a flexible frontend with annotations for graph partitioning and a programmable interface for defining custom intra-device parallelism strategies. Its efficient backend manages complex control/data-flow asynchronously, uses custom memory management to eliminate copy overheads, and preserves compatibility with optimizations like CUDA Graphs and TorchInductor. We demonstrate that DynaFlow can integrate representative parallelism strategies into 6 state-of-the-art ML systems with minimal code changes, achieving up to a 1.29x throughput improvement. DynaFlow is publicly available at https://github.com/uw-syfi/DynaFlow.

Lijie Wang, Lianjie Guo, Ziyi Xu et al.

Reconstructing large-scale colored point clouds is an important task in robotics, supporting perception, navigation, and scene understanding. Despite advances in LiDAR inertial visual odometry (LIVO), its performance remains highly sensitive to extrinsic calibration. Meanwhile, 3D vision foundation models, such as VGGT, suffer from limited scalability in large environments and inherently lack metric scale. To overcome these limitations, we propose LiDAR-VGGT, a novel framework that tightly couples LiDAR inertial odometry with the state-of-the-art VGGT model through a two-stage coarse- to-fine fusion pipeline: First, a pre-fusion module with robust initialization refinement efficiently estimates VGGT poses and point clouds with coarse metric scale within each session. Then, a post-fusion module enhances cross-modal 3D similarity transformation, using bounding-box-based regularization to reduce scale distortions caused by inconsistent FOVs between LiDAR and camera sensors. Extensive experiments across multiple datasets demonstrate that LiDAR-VGGT achieves dense, globally consistent colored point clouds and outperforms both VGGT-based methods and LIVO baselines. The implementation of our proposed novel color point cloud evaluation toolkit will be released as open source.

Ziyi Xu, Ziyao Huang, Juan Cao et al.

The generation of anchor-style product promotion videos presents promising opportunities in e-commerce, advertising, and consumer engagement. Despite advancements in pose-guided human video generation, creating product promotion videos remains challenging. In addressing this challenge, we identify the integration of human-object interactions (HOI) into pose-guided human video generation as a core issue. To this end, we introduce AnchorCrafter, a novel diffusion-based system designed to generate 2D videos featuring a target human and a customized object, achieving high visual fidelity and controllable interactions. Specifically, we propose two key innovations: the HOI-appearance perception, which enhances object appearance recognition from arbitrary multi-view perspectives and disentangles object and human appearance, and the HOI-motion injection, which enables complex human-object interactions by overcoming challenges in object trajectory conditioning and inter-occlusion management. Extensive experiments show that our system improves object appearance preservation by 7.5\% and doubles the object localization accuracy compared to existing state-of-the-art approaches. It also outperforms existing approaches in maintaining human motion consistency and high-quality video generation. Project page including data, code, and Huggingface demo: https://github.com/cangcz/AnchorCrafter.

Ziyi Xu, Cem Bilaloglu, Yiming Li et al.

In robotics, a common challenge in imitation learning is the mismatch between training and deployment conditions, caused, for example, by environmental changes or imperfect observation and control. When a robot follows a nominal trajectory under such mismatch, it may become stuck and fail to complete the task. This calls for adaptive online exploration strategies that remain grounded in demonstrations. To this end, we propose an adaptive ergodic imitation approach that constructs a target distribution from the geometry of the retrieved demonstrations and uses it to generate trajectories that adaptively interpolate between tracking and exploration. Our method extends ergodic control beyond its traditional role in area-coverage and search by incorporating demonstrations into a retrieval-based receding-horizon framework for adaptive imitation.

Ziyi Xu, Samy Elshamy, Ziyue Zhao et al.

Estimating time-frequency domain masks for single-channel speech enhancement using deep learning methods has recently become a popular research field with promising results. In this paper, we propose a novel components loss (CL) for the training of neural networks for mask-based speech enhancement. During the training process, the proposed CL offers separate control over preservation of the speech component quality, suppression of the residual noise component, and preservation of a naturally sounding residual noise component. We illustrate the potential of the proposed CL by evaluating a standard convolutional neural network (CNN) for mask-based speech enhancement. The new CL obtains a better and more balanced performance in almost all employed instrumental quality metrics over the baseline losses, the latter comprising the conventional mean squared error (MSE) loss and also auditory-related loss functions, such as the perceptual evaluation of speech quality (PESQ) loss and the recently proposed perceptual weighting filter loss. Particularly, applying the CL offers better speech component quality, better overall enhanced speech perceptual quality, as well as a more naturally sounding residual noise. On average, an at least 0.1 points higher PESQ score on the enhanced speech is obtained while also obtaining a higher SNR improvement by more than 0.5 dB, for seen noise types. This improvement is stronger for unseen noise types, where an about 0.2 points higher PESQ score on the enhanced speech is obtained, while also the output SNR is ahead by more than 0.5 dB. The new proposed CL is easy to implement and code is provided at https://github.com/ifnspaml/Components-Loss.

Yixin Dong, Charlie F. Ruan, Yaxing Cai et al.

The applications of LLM Agents are becoming increasingly complex and diverse, leading to a high demand for structured outputs that can be parsed into code, structured function calls, and embodied agent commands. These developments bring significant demands for structured generation in LLM inference. Context-free grammar is a flexible approach to enable structured generation via constrained decoding. However, executing context-free grammar requires going through several stack states over all tokens in vocabulary during runtime, bringing non-negligible overhead for structured generation. In this paper, we propose XGrammar, a flexible and efficient structure generation engine for large language models. XGrammar accelerates context-free grammar execution by dividing the vocabulary into context-independent tokens that can be prechecked and context-dependent tokens that need to be interpreted during runtime. We further build transformations to expand the grammar context and reduce the number of context-independent tokens. Additionally, we build an efficient persistent stack to accelerate the context-dependent token checks. Finally, we co-design the grammar engine with LLM inference engine to overlap grammar computation with GPU executions. Evaluation results show that XGrammar can achieve up to 100x speedup over existing solutions. Combined with an LLM inference engine, it can generate near-zero overhead structure generation in end-to-end low-LLM serving.

Shan Yu, Junyi Shu, Yuanjiang Ni et al.

As LLM applications grow more complex, developers are increasingly adopting multi-agent architectures to decompose workflows into specialized, collaborative components, introducing structure that constrains agent behavior and exposes useful semantic predictability. Unlike traditional LLM serving, which operates under highly dynamic and uncertain conditions, this structured topology enables opportunities to reduce runtime uncertainty -- yet existing systems fail to exploit it, treating agentic workloads as generic traffic and incurring significant inefficiencies. Our analysis of production traces from an agent-serving platform and an internal coding assistant reveals key bottlenecks, including low prefix cache hit rates, severe resource contention from long-context requests, and substantial queuing delays due to suboptimal scaling. To address these challenges, we propose Pythia, a multi-agent serving system that captures workflow semantics through a simple interface at the serving layer, unlocking new optimization opportunities and substantially improving throughput and job completion time over state-of-the-art baselines.

Nathaniel Demchak, Xin Guan, Zekun Wu et al.

Open-generation bias benchmarks evaluate social biases in Large Language Models (LLMs) by analyzing their outputs. However, the classifiers used in analysis often have inherent biases, leading to unfair conclusions. This study examines such biases in open-generation benchmarks like BOLD and SAGED. Using the MGSD dataset, we conduct two experiments. The first uses counterfactuals to measure prediction variations across demographic groups by altering stereotype-related prefixes. The second applies explainability tools (SHAP) to validate that the observed biases stem from these counterfactuals. Results reveal unequal treatment of demographic descriptors, calling for more robust bias metric models.

Linzhang Li, Yixin Dong, Guanjie Wang et al.

Modern LLM agents are required to handle increasingly complex structured generation tasks, such as tool calling and conditional structured generation. These tasks are significantly more dynamic than predefined structures, posing new challenges to the current structured generation engines. In this paper, we propose XGrammar 2, a highly optimized structured generation engine for agentic LLMs. XGrammar 2 accelerates the mask generation for these dynamic structured generation tasks through a new dynamic dispatching semantics: TagDispatch. We further introduce a just-in-time (JIT) compilation method to reduce compilation time and a cross-grammar caching mechanism to leverage the common sub-structures across different grammars. Additionally, we extend the previous PDA-based mask generation algorithm to the Earley-parser-based one and design a repetition compression algorithm to handle repetition structures in grammars. Evaluation results show that XGrammar 2 can achieve more than 6x speedup over the existing structured generation engines. Integrated with an LLM inference engine, XGrammar 2 can handle dynamic structured generation tasks with near-zero overhead.

Ziyi Xu, Haohong Lin, Shiqi Liu et al.

The robotic assembly task poses a key challenge in building generalist robots due to the intrinsic complexity of part interactions and the sensitivity to noise perturbations in contact-rich settings. The assembly agent is typically designed in a hierarchical manner: high-level multi-part reasoning and low-level precise control. However, implementing such a hierarchical policy is challenging in practice due to the mismatch between high-level skill queries and low-level execution. To address this, we propose the Query-centric Diffusion Policy (QDP), a hierarchical framework that bridges high-level planning and low-level control by utilizing queries comprising objects, contact points, and skill information. QDP introduces a query-centric mechanism that identifies task-relevant components and uses them to guide low-level policies, leveraging point cloud observations to improve the policy's robustness. We conduct comprehensive experiments on the FurnitureBench in both simulation and real-world settings, demonstrating improved performance in skill precision and long-horizon success rate. In the challenging insertion and screwing tasks, QDP improves the skill-wise success rate by over 50% compared to baselines without structured queries.

Ziyi Xu, Maximilian Strake, Tim Fingscheidt

Speech enhancement employing deep neural networks (DNNs) for denoising are called deep noise suppression (DNS). During training, DNS methods are typically trained with mean squared error (MSE) type loss functions, which do not guarantee good perceptual quality. Perceptual evaluation of speech quality (PESQ) is a widely used metric for evaluating speech quality. However, the original PESQ algorithm is non-differentiable, and therefore cannot directly be used as optimization criterion for gradient-based learning. In this work, we propose an end-to-end non-intrusive PESQNet DNN to estimate the PESQ scores of the enhanced speech signal. Thus, by providing a reference-free perceptual loss, it serves as a mediator towards the DNS training, allowing to maximize the PESQ score of the enhanced speech signal. We illustrate the potential of our proposed PESQNet-mediated training on the basis of an already strong baseline DNS. As further novelty, we propose to train the DNS and the PESQNet alternatingly to keep the PESQNet up-to-date and perform well specifically for the DNS under training. Our proposed method is compared to the same DNS trained with MSE-based loss for joint denoising and dereverberation, and the Interspeech 2021 DNS Challenge baseline. Detailed analysis shows that the PESQNet mediation can further increase the DNS performance by about 0.1 PESQ points on synthetic test data and by 0.03 DNSMOS points on real test data, compared to training with the MSE-based loss. Our proposed method also outperforms the Challenge baseline by 0.2 PESQ points on synthetic test data and 0.1 DNSMOS points on real test data.