arXiv subjects

Wei Gao

Publications and source records attributed to Wei Gao.

At least 127 records · Page 7Linked to original sources

arXiv2025

Part II: ROLL Flash -- Accelerating RLVR and Agentic Training with Asynchrony

Synchronous Reinforcement Learning (RL) post-training has emerged as a crucial step for enhancing Large Language Models (LLMs) with diverse capabilities. However, many systems designed to accelerate RL post-training still suffer from low resource utilization and limited scalability. We present ROLL Flash, a system that extends ROLL with native support for asynchronous RL post-training. ROLL Flash is built upon two core design principles: fine-grained parallelism and rollout-train decoupling. Guided by these principles, ROLL Flash provides flexible programming interfaces that enable a fully asynchronous training architecture and support efficient rollout mechanisms, including queue scheduling and environment-level asynchronous execution. Through comprehensive theoretical analysis and extensive experiments, we demonstrate that ROLL Flash significantly improves resource utilization and scalability over synchronous RL post-training. ROLL Flash achieves up to 2.24x speedup on RLVR tasks and 2.72x on agentic tasks, using the same GPU budget as synchronous baselines. Furthermore, we implement several popular off-policy algorithms and verify that asynchronous training can achieve performance on par with synchronous training.

cs.LG↗

arXiv2025

GUIDe: Generative and Uncertainty-Informed Inverse Design for On-Demand Nonlinear Functional Responses

Inverse design is a common yet challenging engineering problem, particularly for nonlinear functional responses such as mechanical behavior or spectral analysis. Deep generative models are motivated by intractability, non-existence or non-uniqueness of solutions, and the need for rapid solution-space exploration. In this study, we show that deep generative model-based and optimization-based approaches can provide incomplete solutions or hallucinate given out-of-distribution targets. To address this, we propose the Generative and Uncertainty-informed Inverse Design (GUIDe) framework, which leverages probabilistic machine learning, statistical inference, and Markov chain Monte Carlo to generate designs with targeted nonlinear behaviors. Instead of inverse mappings, i.e., response $\mapsto$ design, GUIDe adopts design $\mapsto$ response: a forward model predicts each design's nonlinear functional response and evaluates the confidence under a user-specified tolerance. Sampling the solution space by this confidence yields diverse feasible designs. Our validation on nacre-inspired materials finds solutions beyond the training range, even under out-of-distribution targets.

cs.CE↗

arXiv2025

Flow4Agent: Long-form Video Understanding via Motion Prior from Optical Flow

Long-form video understanding has always been a challenging problem due to the significant redundancy in both temporal and spatial contents. This challenge is further exacerbated by the limited context length of Multimodal Large Language Models (MLLMs). To address this issue, many previous works have attempted to extract key video information, where the "key" is typically semantic-aware and heavily dependent on the CLIP model as prior. In this paper, we propose Flow4Agent, a novel framework that pioneeringly incorporates motion priors from optical flow to facilitate LLM-based long video understanding. Flow4Agent mitigates the redundancy in long videos at both temporal and spatial levels through two core modules: Temporal Granularity Optimization (TGO) adaptively refines framelevel hierarchies, which first leverages coarse flow priors to group similar visual contents and then applies semantic priors to filter out highly irrelevant scene information. Motion Token Pruning (MTP) further refines the intra-frame visual representations, pruning high-redundancy video tokens using fine-grained optical flow information. Extensive experiments demonstrate that our Flow4Agent outperforms existing methods across a wide range of video MLLM benchmarks, especially for hour-level video understanding tasks, achieving 64.7% on Video-MME, 71.4% on MLVU and 60.4% on LongVideoBench.

cs.CV↗

arXiv2025

Translation from Wearable PPG to 12-Lead ECG

The 12-lead electrocardiogram (ECG) is the gold standard for cardiovascular monitoring, offering superior diagnostic granularity and specificity compared to photoplethysmography (PPG). However, existing 12-lead ECG systems rely on cumbersome multi-electrode setups, limiting sustained monitoring in ambulatory settings, while current PPG-based methods fail to reconstruct multi-lead ECG due to the absence of inter-lead constraints and insufficient modeling of spatial-temporal dependencies across leads. To bridge this gap, we introduce P2Es, an innovative demographic-aware diffusion framework designed to generate clinically valid 12-lead ECG from PPG signals via three key innovations. Specifically, in the forward process, we introduce frequency-domain blurring followed by temporal noise interference to simulate real-world signal distortions. In the reverse process, we design a temporal multi-scale generation module followed by frequency deblurring. In particular, we leverage KNN-based clustering combined with contrastive learning to assign affinity matrices for the reverse process, enabling demographic-specific ECG translation. Extensive experimental results show that P2Es outperforms baseline models in 12-lead ECG reconstruction.

cs.LG↗

arXiv2025

Experience Paper: Adopting Activity Recognition in On-demand Food Delivery Business

This paper presents the first nationwide deployment of human activity recognition (HAR) technology in the on-demand food delivery industry. We successfully adapted the state-of-the-art LIMU-BERT foundation model to the delivery platform. Spanning three phases over two years, the deployment progresses from a feasibility study in Yangzhou City to nationwide adoption involving 500,000 couriers across 367 cities in China. The adoption enables a series of downstream applications, and large-scale tests demonstrate its significant operational and economic benefits, showcasing the transformative potential of HAR technology in real-world applications. Additionally, we share lessons learned from this deployment and open-source our LIMU-BERT pretrained with millions of hours of sensor data.

cs.AI↗

arXiv2025

ImaginationPolicy: Towards Generalizable, Precise and Reliable End-to-End Policy for Robotic Manipulation

End-to-end robot manipulation policies offer significant potential for enabling embodied agents to understand and interact with the world. Unlike traditional modular pipelines, end-to-end learning mitigates key limitations such as information loss between modules and feature misalignment caused by isolated optimization targets. Despite these advantages, existing end-to-end neural networks for robotic manipulation--including those based on large VLM/VLA models--remain insufficiently performant for large-scale practical deployment. In this paper, we take a step towards an end-to-end manipulation policy that is generalizable, accurate and reliable. To achieve this goal, we propose a novel Chain of Moving Oriented Keypoints (CoMOK) formulation for robotic manipulation. Our formulation is used as the action representation of a neural policy, which can be trained in an end-to-end fashion. Such an action representation is general, as it extends the standard end-effector pose action representation and supports a diverse set of manipulation tasks in a unified manner. The oriented keypoint in our method enables natural generalization to objects with different shapes and sizes, while achieving sub-centimeter accuracy. Moreover, our formulation can easily handle multi-stage tasks, multi-modal robot behaviors, and deformable objects. Extensive simulated and hardware experiments demonstrate the effectiveness of our method.

cs.RO↗

arXiv2025

RollPacker: Mitigating Long-Tail Rollouts for Fast, Synchronous RL Post-Training

Reinforcement Learning (RL) is a pivotal post-training technique for enhancing the reasoning capabilities of Large Language Models (LLMs). However, synchronous RL post-training often suffers from significant GPU underutilization, referred to as bubbles, caused by imbalanced response lengths within rollout steps. Many RL systems attempt to alleviate this problem by relaxing synchronization, but this can compromise training accuracy. In this paper, we introduce tail batching, a novel rollout scheduling strategy for synchronous RL that systematically consolidates prompts leading to long-tail responses into a small subset of rollout steps (long rounds), while ensuring that the majority of steps (short rounds) involve only balanced, short rollouts. By excluding long responses from short rounds and rescheduling them into a few designated long rounds, tail batching effectively reduces GPU idle time during rollouts and significantly accelerates RL training without sacrificing accuracy. We present RollPacker, a system that fully harnesses the benefits of tail batching through holistic optimizations across all three RL stages: elastic parallelism adaptation for rollout, dynamic resource allocation and scheduling for reward, and stream-based training. Empirical results show that RollPacker achieves a 2.03x-2.56x end-to-end training time reduction compared to veRL and up to 2.24x speedup compared to RLHFuse for the Qwen2.5 family of LLMs on up to 128 H800 GPUs.

cs.DC↗

arXiv2025

Normal mode parameters estimation by a VLA in single-shooting

This paper proposes an orthogonality-constrained modal search (OCMS) method for estimating modal wavenumbers and modal depth functions using a vertical linear array (VLA). Under the assumption of a known sound speed profile, OCMS leverages the orthogonality of distinct modal depth functions to extract both the modal depth functions and their corresponding wavenumbers, even when the VLA and a monochromatic sound source remain stationary.The performance of OCMS is evaluated through numerical simulations under varying signal-to-noise ratios (SNRs), different VLA apertures, varying numbers of VLA elements, VLA tilt and sound speed profile (SSP) uncertainty. The results demonstrate that OCMS is robust against noise, VLA aperture variations, and changes in the number of VLA elements, meanwhile, the algorithm maintains reliable performance when SSP uncertainty < 1 m/s and VLA tilt angle <5°. Furthermore, the effectiveness of OCMS is validated using SwellEx96 experimental data. The relative error between the modal wavenumbers derived from experimental data and those computed via Kraken is on the order of $10^{-4}$.

eess.SP↗

arXiv2025

Efficient Geometry Compression and Communication for 3D Gaussian Splatting Point Clouds

Storage and transmission challenges in dynamic 3D scene representation based on the i3DV platform, With increasing scene complexity, the explosive growth of 3D Gaussian data volume causes excessive storage space occupancy. To address this issue, we propose adopting the AVS PCRM reference software for efficient compression of Gaussian point cloud geometry data. The strategy deeply integrates the advanced encoding capabilities of AVS PCRM into the i3DV platform, forming technical complementarity with the original rate-distortion optimization mechanism based on binary hash tables. On one hand, the hash table efficiently caches inter-frame Gaussian point transformation relationships, which allows for high-fidelity transmission within a 40 Mbps bandwidth constraint. On the other hand, AVS PCRM performs precise compression on geometry data. Experimental results demonstrate that the joint framework maintains the advantages of fast rendering and high-quality synthesis in 3D Gaussian technology while achieving significant 10\%-25\% bitrate savings on universal test sets. It provides a superior rate-distortion tradeoff solution for the storage, transmission, and interaction of 3D volumetric video.

cs.MM↗

arXiv2025

Learned Rate Control for Frame-Level Adaptive Neural Video Compression via Dynamic Neural Network

Neural Video Compression (NVC) has achieved remarkable performance in recent years. However, precise rate control remains a challenge due to the inherent limitations of learning-based codecs. To solve this issue, we propose a dynamic video compression framework designed for variable bitrate scenarios. First, to achieve variable bitrate implementation, we propose the Dynamic-Route Autoencoder with variable coding routes, each occupying partial computational complexity of the whole network and navigating to a distinct RD trade-off. Second, to approach the target bitrate, the Rate Control Agent estimates the bitrate of each route and adjusts the coding route of DRA at run time. To encompass a broad spectrum of variable bitrates while preserving overall RD performance, we employ the Joint-Routes Optimization strategy, achieving collaborative training of various routes. Extensive experiments on the HEVC and UVG datasets show that the proposed method achieves an average BD-Rate reduction of 14.8% and BD-PSNR gain of 0.47dB over state-of-the-art methods while maintaining an average bitrate error of 1.66%, achieving Rate-Distortion-Complexity Optimization (RDCO) for various bitrate and bitrate-constrained applications. Our code is available at https://git.openi.org.cn/OpenAICoding/DynamicDVC.

cs.CV↗

arXiv2025

AdaDPCC: Adaptive Rate Control and Rate-Distortion-Complexity Optimization for Dynamic Point Cloud Compression

Dynamic point cloud compression (DPCC) is crucial in applications like autonomous driving and AR/VR. Current compression methods face challenges with complexity management and rate control. This paper introduces a novel dynamic coding framework that supports variable bitrate and computational complexities. Our approach includes a slimmable framework with multiple coding routes, allowing for efficient Rate-Distortion-Complexity Optimization (RDCO) within a single model. To address data sparsity in inter-frame prediction, we propose the coarse-to-fine motion estimation and compensation module that deconstructs geometric information while expanding the perceptive field. Additionally, we propose a precise rate control module that content-adaptively navigates point cloud frames through various coding routes to meet target bitrates. The experimental results demonstrate that our approach reduces the average BD-Rate by 5.81% and improves the BD-PSNR by 0.42 dB compared to the state-of-the-art method, while keeping the average bitrate error at 0.40%. Moreover, the average coding time is reduced by up to 44.6% compared to D-DPCC, underscoring its efficiency in real-time and bitrate-constrained DPCC scenarios. Our code is available at https://git.openi.org.cn/OpenPointCloud/Ada_DPCC.

cs.MM↗

arXiv2025

Model Accuracy and Data Heterogeneity Shape Uncertainty Quantification in Machine Learning Interatomic Potentials

Machine learning interatomic potentials (MLIPs) enable accurate atomistic modelling, but reliable uncertainty quantification (UQ) remains elusive. In this study, we investigate two UQ strategies, ensemble learning and D-optimality, within the atomic cluster expansion framework. It is revealed that higher model accuracy strengthens the correlation between predicted uncertainties and actual errors and improves novelty detection, with D-optimality yielding more conservative estimates. Both methods deliver well calibrated uncertainties on homogeneous training sets, yet they underpredict errors and exhibit reduced novelty sensitivity on heterogeneous datasets. To address this limitation, we introduce clustering-enhanced local D-optimality, which partitions configuration space into clusters during training and applies D-optimality within each cluster. This approach substantially improves the detection of novel atomic environments in heterogeneous datasets. Our findings clarify the roles of model fidelity and data heterogeneity in UQ performance and provide a practical route to robust active learning and adaptive sampling strategies for MLIP development.

cond-mat.mtrl-sci↗

arXiv2025

MMBERT: Scaled Mixture-of-Experts Multimodal BERT for Robust Chinese Hate Speech Detection under Cloaking Perturbations

Hate speech detection on Chinese social networks presents distinct challenges, particularly due to the widespread use of cloaking techniques designed to evade conventional text-based detection systems. Although large language models (LLMs) have recently improved hate speech detection capabilities, the majority of existing work has concentrated on English datasets, with limited attention given to multimodal strategies in the Chinese context. In this study, we propose MMBERT, a novel BERT-based multimodal framework that integrates textual, speech, and visual modalities through a Mixture-of-Experts (MoE) architecture. To address the instability associated with directly integrating MoE into BERT-based models, we develop a progressive three-stage training paradigm. MMBERT incorporates modality-specific experts, a shared self-attention mechanism, and a router-based expert allocation strategy to enhance robustness against adversarial perturbations. Empirical results in several Chinese hate speech datasets show that MMBERT significantly surpasses fine-tuned BERT-based encoder models, fine-tuned LLMs, and LLMs utilizing in-context learning approaches.

cs.CL↗

arXiv2025

Heterogeneous Ensemble Enables a Universal Uncertainty Metric for Atomistic Foundation Models

Universal machine learning interatomic potentials (uMLIPs) are reshaping atomistic simulation as foundation models, delivering near \textit{ab initio} accuracy at a fraction of the cost. Yet the lack of reliable, general uncertainty quantification limits their safe, wide-scale use. Here we introduce a unified, scalable uncertainty metric $U$ based on a heterogeneous model ensemble with reuse of pretrained uMLIPs. Across chemically and structurally diverse datasets, $U$ shows a strong correlation with the true prediction errors and provides a robust ranking of configuration-level risk. Leveraging this metric, we propose an uncertainty-aware model distillation framework to produce system-specific potentials: for W, an accuracy comparable to full-DFT training is achieved using only $4\%$ of the DFT labels; for MoNbTaW, no additional DFT calculations are required. Notably, by filtering numerical label noise, the distilled models can, in some cases, surpass the accuracy of the DFT reference labels. The uncertainty-aware approach offers a practical monitor of uMLIP reliability in deployment, and guides data selection and fine-tuning strategies, thereby advancing the construction and safe use of foundation models and enabling cost-efficient development of accurate, system-specific potentials.

cond-mat.mtrl-sci↗

arXiv2025

Geometrical portrait of Multipath error propagation in GNSS Direct Position Estimation

Direct Position Estimation (DPE) is a method that directly estimate position, velocity, and time (PVT) information from cross ambiguity function (CAF) of the GNSS signals, significantly enhancing receiver robustness in urban environments. However, there is still a lack of theoretical characterization on multipath errors in the context of DPE theory. Geometric observations highlight the unique characteristics of DPE errors stemming from multipath and thermal noise as estimation bias and variance respectively. Expanding upon the theoretical framework of DPE noise variance through geometric analysis, this paper focuses on a geometric representation of multipath errors by quantifying the deviations in CAF and PVT solutions caused by off-centering bias relative to the azimuth and elevation angles. A satellite circular multipath bias (SCMB) model is introduced, amalgamating CAF and PVT errors from multiple satellite channels. The boundaries for maximum or minimum PVT bias are established through discussions encompassing various multipath conditions. The correctness of the multipath geometrical portrait is confirmed through both Monte Carlo simulations and urban canyon tests. The findings indicate that the maximum PVT bias depends on the largest multipath errors observed across various satellite channels. Additionally, the PVT bias increases with satellite elevation angles, influenced by the CAF multipath bias projection. This serves as a reference for selecting DPE satellites from a geometric standpoint, underscoring the importance of choosing a balanced combination of high and low elevation angles to achieve an optimal satellite geometry configuration.

eess.SP↗

arXiv2025

Kinetics of Vacancy-Assisted Reversible Phase Transition in Monolayer MoTe$_2$

We investigate the kinetics of phase transition between the 2H and 1T$^\prime$ phases in monolayer MoTe$_2$ using atomistic simulations based on a machine learning interatomic potential trained on SCAN-DFT data, combined with mean field kinetic theory to interpret the underlying mechanisms. The transition is found to involve both diffusive and diffusionless mechanisms. Nucleation of 1T$^\prime$ phase is initiated by the coalescence of neighboring Te monovacancies into divacancies, which are found to be mobile and can interact with other Te vacancies to form small triangular 1T$^\prime$ islands. Growth of these islands proceeds either by incorporating pre-existing vacancies at the phase boundaries or, in their absence, by absorbing divacancies that migrate from the surrounding lattice. Once a critical island size is reached, vacancy-free growth becomes possible although with a higher activation barrier. Upon removal of external stimuli, the system reverts to 2H phase, during which Te vacancies reorganize into three-fold spoke-like vacancy lines at the island center. This reverse process and the subsequent 1T$^\prime$$\leftrightarrow$2H reversible transitions are diffusionless, rapid, do not require additional vacancies and can be driven by mild external stimuli. Although our analysis focuses on strain-induced transitions, the kinetic mechanisms are expected to be generalizable to other types of stimuli.

cond-mat.mtrl-sci↗

arXiv2025

On the Learning with Augmented Class via Forests

Decision trees and forests have achieved successes in various real applications, most working with all testing classes known in training data. In this work, we focus on learning with augmented class via forests, where an augmented class may appear in testing data yet not in training data. We incorporate information of augmented class into trees' splitting, that is, augmented Gini impurity, a new splitting criterion is introduced to exploit some unlabeled data from testing distribution. We then develop the Learning with Augmented Class via Forests (short for LACForest) approach, which constructs shallow forests according to the augmented Gini impurity and then splits forests with pseudo-labeled augmented instances for better performance. We also develop deep neural forests via an optimization objective based on our augmented Gini impurity, which essentially utilizes the representation power of neural networks for forests. Theoretically, we present the convergence analysis for our augmented Gini impurity, and we finally conduct experiments to evaluate our approaches. The code is available at https://github.com/nju-xuf/LACForest.

cs.LG↗

arXiv2025

ProGait: A Multi-Purpose Video Dataset and Benchmark for Transfemoral Prosthesis Users

Prosthetic legs play a pivotal role in clinical rehabilitation, allowing individuals with lower-limb amputations the ability to regain mobility and improve their quality of life. Gait analysis is fundamental for optimizing prosthesis design and alignment, directly impacting the mobility and life quality of individuals with lower-limb amputations. Vision-based machine learning (ML) methods offer a scalable and non-invasive solution to gait analysis, but face challenges in correctly detecting and analyzing prosthesis, due to their unique appearances and new movement patterns. In this paper, we aim to bridge this gap by introducing a multi-purpose dataset, namely ProGait, to support multiple vision tasks including Video Object Segmentation, 2D Human Pose Estimation, and Gait Analysis (GA). ProGait provides 412 video clips from four above-knee amputees when testing multiple newly-fitted prosthetic legs through walking trials, and depicts the presence, contours, poses, and gait patterns of human subjects with transfemoral prosthetic legs. Alongside the dataset itself, we also present benchmark tasks and fine-tuned baseline models to illustrate the practical application and performance of the ProGait dataset. We compared our baseline models against pre-trained vision models, demonstrating improved generalizability when applying the ProGait dataset for prosthesis-specific tasks. Our code is available at https://github.com/pittisl/ProGait and dataset at https://huggingface.co/datasets/ericyxy98/ProGait.

cs.CV↗