Hui Lu

Shulei Wang, Longhui Wei, Xin He et al.

Personalized generation models for a single subject have demonstrated remarkable effectiveness, highlighting their significant potential. However, when extended to multiple subjects, existing models often exhibit degraded performance, particularly in maintaining subject consistency and adhering to textual prompts. We attribute these limitations to the absence of high-quality multi-subject datasets and refined post-training strategies. To address these challenges, we propose a scalable multi-subject data generation pipeline that leverages powerful single-subject generation models to construct diverse and high-quality multi-subject training data. Through this dataset, we first enable single-subject personalization models to acquire knowledge of synthesizing multi-image and multi-subject scenarios. Furthermore, to enhance both subject consistency and text controllability, we design a set of Pairwise Subject-Consistency Rewards and general-purpose rewards, which are incorporated into a refined reinforcement learning stage. To comprehensively evaluate multi-subject personalization, we introduce a new benchmark that assesses model performance using seven subsets across three dimensions. Extensive experiments demonstrate the effectiveness of our approach in advancing multi-subject personalized image generation. Github Link: https://github.com/wang-shulei/PSR

Haohan Guo, Fenglong Xie, Xixin Wu et al.

This paper aims to enhance low-resource TTS by reducing training data requirements using compact speech representations. A Multi-Stage Multi-Codebook (MSMC) VQ-GAN is trained to learn the representation, MSMCR, and decode it to waveforms. Subsequently, we train the multi-stage predictor to predict MSMCRs from the text for TTS synthesis. Moreover, we optimize the training strategy by leveraging more audio to learn MSMCRs better for low-resource languages. It selects audio from other languages using speaker similarity metric to augment the training set, and applies transfer learning to improve training quality. In MOS tests, the proposed system significantly outperforms FastSpeech and VITS in standard and low-resource scenarios, showing lower data requirements. The proposed training strategy effectively enhances MSMCRs on waveform reconstruction. It improves TTS performance further, which wins 77% votes in the preference test for the low-resource TTS with only 15 minutes of paired data.

Chao-Wei Huang, Hui Lu, Hongyu Gong et al.

Large language models (LLMs), known for their exceptional reasoning capabilities, generalizability, and fluency across diverse domains, present a promising avenue for enhancing speech-related tasks. In this paper, we focus on integrating decoder-only LLMs to the task of speech-to-text translation (S2TT). We propose a decoder-only architecture that enables the LLM to directly consume the encoded speech representation and generate the text translation. Additionally, we investigate the effects of different parameter-efficient fine-tuning techniques and task formulation. Our model achieves state-of-the-art performance on CoVoST 2 and FLEURS among models trained without proprietary data. We also conduct analyses to validate the design choices of our proposed model and bring insights to the integration of LLMs to S2TT.

Hui Lu, Yi Yu, Song Xia et al.

Large-scale Video Foundation Models (VFMs) has significantly advanced various video-related tasks, either through task-specific models or Multi-modal Large Language Models (MLLMs). However, the open accessibility of VFMs also introduces critical security risks, as adversaries can exploit full knowledge of the VFMs to launch potent attacks. This paper investigates a novel and practical adversarial threat scenario: attacking downstream models or MLLMs fine-tuned from open-source VFMs, without requiring access to the victim task, training data, model query, and architecture. In contrast to conventional transfer-based attacks that rely on task-aligned surrogate models, we demonstrate that adversarial vulnerabilities can be exploited directly from the VFMs. To this end, we propose the Transferable Video Attack (TVA), a temporal-aware adversarial attack method that leverages the temporal representation dynamics of VFMs to craft effective perturbations. TVA integrates a bidirectional contrastive learning mechanism to maximize the discrepancy between the clean and adversarial features, and introduces a temporal consistency loss that exploits motion cues to enhance the sequential impact of perturbations. TVA avoids the need to train expensive surrogate models or access to domain-specific data, thereby offering a more practical and efficient attack strategy. Extensive experiments across 24 video-related tasks demonstrate the efficacy of TVA against downstream models and MLLMs, revealing a previously underexplored security vulnerability in the deployment of video models.

Hui Lu, Mia Chiquier, Carl Vondrick

We introduce a framework for navigating through cluttered environments by connecting multiple cameras together while simultaneously preserving privacy. Occlusions and obstacles in large environments are often challenging situations for navigation agents because the environment is not fully observable from a single camera view. Given multiple camera views of an environment, our approach learns to produce a multiview scene representation that can only be used for navigation, provably preventing one party from inferring anything beyond the output task. On a new navigation dataset that we will publicly release, experiments show that private multiparty representations allow navigation through complex scenes and around obstacles while jointly preserving privacy. Our approach scales to an arbitrary number of camera viewpoints. We believe developing visual representations that preserve privacy is increasingly important for many applications such as navigation.

Renjie Liang, Yiming Yang, Hui Lu et al.

Temporal Sentence Grounding in Videos (TSGV) aims to detect the event timestamps described by the natural language query from untrimmed videos. This paper discusses the challenge of achieving efficient computation in TSGV models while maintaining high performance. Most existing approaches exquisitely design complex architectures to improve accuracy with extra layers and loss, suffering from inefficiency and heaviness. Although some works have noticed that, they only make an issue of feature fusion layers, which can hardly enjoy the highspeed merit in the whole clunky network. To tackle this problem, we propose a novel efficient multi-teacher model (EMTM) based on knowledge distillation to transfer diverse knowledge from both heterogeneous and isomorphic networks. Specifically, We first unify different outputs of the heterogeneous models into one single form. Next, a Knowledge Aggregation Unit (KAU) is built to acquire high-quality integrated soft labels from multiple teachers. After that, the KAU module leverages the multi-scale video and global query information to adaptively determine the weights of different teachers. A Shared Encoder strategy is then proposed to solve the problem that the student shallow layers hardly benefit from teachers, in which an isomorphic teacher is collaboratively trained with the student to align their hidden states. Extensive experimental results on three popular TSGV benchmarks demonstrate that our method is both effective and efficient without bells and whistles.

Qinlong Wang, Tingfeng Lan, Yinghao Tang et al.

Deep learning recommendation models (DLRM) rely on large embedding tables to manage categorical sparse features. Expanding such embedding tables can significantly enhance model performance, but at the cost of increased GPU/CPU/memory usage. Meanwhile, tech companies have built extensive cloud-based services to accelerate training DLRM models at scale. In this paper, we conduct a deep investigation of the DLRM training platforms at AntGroup and reveal two critical challenges: low resource utilization due to suboptimal configurations by users and the tendency to encounter abnormalities due to an unstable cloud environment. To overcome them, we introduce DLRover-RM, an elastic training framework for DLRMs designed to increase resource utilization and handle the instability of a cloud environment. DLRover-RM develops a resource-performance model by considering the unique characteristics of DLRMs and a three-stage heuristic strategy to automatically allocate and dynamically adjust resources for DLRM training jobs for higher resource utilization. Further, DLRover-RM develops multiple mechanisms to ensure efficient and reliable execution of DLRM training jobs. Our extensive evaluation shows that DLRover-RM reduces job completion times by 31%, increases the job completion rate by 6%, enhances CPU usage by 15%, and improves memory utilization by 20%, compared to state-of-the-art resource scheduling frameworks. DLRover-RM has been widely deployed at AntGroup and processes thousands of DLRM training jobs on a daily basis. DLRover-RM is open-sourced and has been adopted by 10+ companies.

Hanzhang Zhou, Junlang Qian, Zijian Feng et al.

In this study, we investigate in-context learning (ICL) in document-level event argument extraction (EAE) to alleviate the dependency on large-scale labeled data for this task. We introduce the Heuristic-Driven Link-of-Analogy (HD-LoA) prompting to address the challenge of example selection and to develop a prompting strategy tailored for EAE. Specifically, we hypothesize and validate that LLMs learn task-specific heuristics from demonstrations via ICL. Building upon this hypothesis, we introduce an explicit heuristic-driven demonstration construction approach, which transforms the haphazard example selection process into a methodical method that emphasizes task heuristics. Additionally, inspired by the analogical reasoning of human, we propose the link-of-analogy prompting, which enables LLMs to process new situations by drawing analogies to known situations, enhancing their performance on unseen classes beyond limited ICL examples. Experiments show that our method outperforms existing prompting methods and few-shot supervised learning methods on document-level EAE datasets. Additionally, the HD-LoA prompting shows effectiveness in diverse tasks like sentiment analysis and natural language inference, demonstrating its broad adaptability.

Mingyuan Li, Tong Jia, Han Gu et al.

Overlapping object perception aims to decouple the randomly overlapping foreground-background features, extracting foreground features while suppressing background features, which holds significant application value in fields such as security screening and medical auxiliary diagnosis. Despite some research efforts to tackle the challenge of overlapping object perception, most solutions are confined to the spatial domain. Through frequency domain analysis, we observe that the degradation of contours and textures due to the overlapping phenomenon can be intuitively reflected in the magnitude spectrum. Based on this observation, we propose a general Frequency-Optimized Anti-Overlapping Framework (FOAM) to assist the model in extracting more texture and contour information, thereby enhancing the ability for anti-overlapping object perception. Specifically, we design the Frequency Spatial Transformer Block (FSTB), which can simultaneously extract features from both the frequency and spatial domains, helping the network capture more texture features from the foreground. In addition, we introduce the Hierarchical De-Corrupting (HDC) mechanism, which aligns adjacent features in the separately constructed base branch and corruption branch using a specially designed consistent loss during the training phase. This mechanism suppresses the response to irrelevant background features of FSTBs, thereby improving the perception of foreground contour. We conduct extensive experiments to validate the effectiveness and generalization of the proposed FOAM, which further improves the accuracy of state-of-the-art models on four datasets, specifically for the three overlapping object perception tasks: Prohibited Item Detection, Prohibited Item Segmentation, and Pneumonia Detection. The code will be open source once the paper is accepted.

Yiming Yang, Yangyang Guo, Hui Lu et al.

Recently, Large Vision-Language Models (LVLMs) have made significant strides across diverse multimodal tasks and benchmarks. This paper reveals a largely under-explored problem from existing video-involved LVLMs - language bias, where models tend to prioritize language over video and thus result in incorrect responses. To address this research gap, we first collect a Video Language Bias Evaluation Benchmark, which is specifically designed to assess the language bias in video-involved LVLMs through two key tasks: ambiguous video contrast and interrogative question probing. Accordingly, we design accompanied evaluation metrics that aim to penalize LVLMs being biased by language. In addition, we also propose Multi-branch Contrastive Decoding (MCD), introducing two expert branches to simultaneously counteract language bias potentially generated by the amateur text-only branch. Our experiments demonstrate that i) existing video-involved LVLMs, including both proprietary and open-sourced, are largely limited by the language bias problem; ii) our MCD can effectively mitigate this issue and maintain general-purpose capabilities in various video-involved LVLMs without any additional retraining or alteration to model architectures.

Mingyuan Li, Tong Jia, Hao Wang et al.

Prohibited item detection based on X-ray images is one of the most effective security inspection methods. However, the foreground-background feature coupling caused by the overlapping phenomenon specific to X-ray images makes general detectors designed for natural images perform poorly. To address this issue, we propose a Category Semantic Prior Contrastive Learning (CSPCL) mechanism, which aligns the class prototypes perceived by the classifier with the content queries to correct and supplement the missing semantic information responsible for classification, thereby enhancing the model sensitivity to foreground features. To achieve this alignment, we design a specific contrastive loss, CSP loss, which comprises the Intra-Class Truncated Attraction (ITA) loss and the Inter-Class Adaptive Repulsion (IAR) loss, and outperforms classic contrastive losses. Specifically, the ITA loss leverages class prototypes to attract intra-class content queries and preserves essential intra-class diversity via a gradient truncation function. The IAR loss employs class prototypes to adaptively repel inter-class content queries, with the repulsion strength scaled by prototype-prototype similarity, thereby improving inter-class discriminability, especially among similar categories. CSPCL is general and can be easily integrated into Deformable DETR-based models. Extensive experiments on the PIXray, OPIXray, PIDray, and CLCXray datasets demonstrate that CSPCL significantly enhances the performance of various state-of-the-art models without increasing inference complexity. The code is publicly available at https://github.com/Limingyuan001/CSPCL.

Shuhan Ye, Yi Yu, Qixin Zhang et al.

SNNs promise energy-efficient and low-latency inference, but their performance still trails that of ANNs. ANN-to-SNN knowledge distillation helps narrow this gap, yet the original training data are often unavailable in practical deployment settings. Existing data-free knowledge distillation (DFKD) methods synthesize surrogate data by matching teacher-side priors, especially BN statistics, but these ANN-oriented constraints mainly regularize mean and variance and therefore remain under-constrained for SNN students whose responses depend on threshold-crossing dynamics. In this paper, we propose Spike Tail-Aware Relational Synthesis (STARS), a plug-and-play method for ANN-to-SNN DFKD that augments standard BN-guided synthesis with two complementary objectives: Relational Consistency Alignment, which preserves cross-sample relational consistency between teacher and student, and Tail-Aware Regularization, which regularizes threshold-relevant tail probabilities through soft exceedance over teacher-derived thresholds. Together, these objectives generate synthetic batches that remain teacher-valid while becoming more informative for SNN students. Experiments on CIFAR-10, CIFAR-100, and Tiny-ImageNet across multiple ANN-SNN pairs show that our method consistently improves conventional DFKD baselines and even surpasses several KD methods, with gains of up to 4.6\% on CIFAR-10 and 6.7\% on CIFAR-100, highlighting the importance of complementing BN matching with relational and tail-aware constraints in SNN-oriented DFKD.

Hui Lu, Xueyuan Chen, Huimeng Wang et al.

Full-duplex spoken dialogue requires a model to keep listening while generating its own spoken response. This is challenging for large language models (LLMs), which are designed to extend a single coherent sequence and do not naturally support user input arriving during generation. We argue that how the user stream is routed into the LLM is therefore a key architectural question for full-duplex modeling. To study this question, we extend a text-only LLM into a unified full-duplex spoken dialogue system and compare two routing strategies under a shared training pipeline: (i) channel fusion, which injects the user stream directly into the LLM input, and (ii) cross-attention routing, which keeps the user stream as external memory accessed through cross-attention adapters. Experiments on spoken question answering and full-duplex interaction benchmarks reveal a clear tradeoff. Channel fusion yields stronger semantic grounding and consistently better question-answering performance. However, under semantically overlapping conditions such as user interruptions, it is more vulnerable to context corruption: if the model fails to stop in time, the overlapping user stream can interfere with ongoing generation and lead to semantically incoherent continuations. Cross-attention routing underperforms on question answering, but better preserves the LLM generation context and is more robust to this failure mode. These results establish user-stream routing as a central design axis in full-duplex spoken dialogue and offer practical guidance on the tradeoff between semantic integration and context robustness. We provide a demo page for qualitative inspection.

Hui Lu, Zheng Chai, Shipeng Bai et al.

Driven by scaling laws, recommender systems increasingly rely on large-scale models to capture complex feature interactions and user behaviors, but this trend also leads to prohibitive training and inference costs. While long-sequence models(e.g., LONGER) can reuse user-side computation through KV caching, such reuse is difficult in dense feature interaction architectures(e.g., RankMixer), where user and group (candidate item) features are deeply entangled across layers. In this work, we propose User-Group Separation (UG-Sep), a novel framework that enables reusable user-side computation in dense interaction models for the first time. UG-Sep introduces a masking mechanism that explicitly disentangles user-side and item-side information flows within token-mixing layers, ensuring that a subset of tokens to preserve purely user-side representations across layers. This design enables corresponding token computations to be reused across multiple samples, significantly reducing redundant inference cost. To compensate for potential expressiveness loss induced by masking, we further propose an Information Compensation strategy that adaptively reconstructs suppressed user-item interactions. Moreover, as UG-Sep substantially reduces user-side FLOPs and exposes memory-bound components, we incorporate W8A16 (8-bit weight, 16-bit activation) weight-only quantization to alleviate memory bandwidth bottlenecks and achieve additional acceleration. We conduct extensive offline evaluations and large-scale online A/B experiments at ByteDance, demonstrating that UG-Sep reduces inference latency by up to 20 percent without degrading online user experience or commercial metrics across multiple business scenarios, including feed recommendation and advertising systems.

Hui Lu, Yi Yu, Yiming Yang et al.

Targeted adversarial attacks on closed-source multimodal large language models (MLLMs) have been increasingly explored under black-box transfer, yet prior methods are predominantly sample-specific and offer limited reusability across inputs. We instead study a more stringent setting, Universal Targeted Transferable Adversarial Attacks (UTTAA), where a single perturbation must consistently steer arbitrary inputs toward a specified target across unknown commercial MLLMs. Naively adapting existing sample-wise attacks to this universal setting faces three core difficulties: (i) target supervision becomes high-variance due to target-crop randomness, (ii) token-wise matching is unreliable because universality suppresses image-specific cues that would otherwise anchor alignment, and (iii) few-source per-target adaptation is highly initialization-sensitive, which can degrade the attainable performance. In this work, we propose MCRMO-Attack, which stabilizes supervision via Multi-Crop Aggregation with an Attention-Guided Crop, improves token-level reliability through alignability-gated Token Routing, and meta-learns a cross-target perturbation prior that yields stronger per-target solutions. Across commercial MLLMs, we boost unseen-image attack success rate by +23.7\% on GPT-4o and +19.9\% on Gemini-2.0 over the strongest universal baseline.

Hui Lu, Albert Ali Salah, Ronald Poppe

A key challenge in continuous sign language recognition (CSLR) is to efficiently capture long-range spatial interactions over time from the video input. To address this challenge, we propose TCNet, a hybrid network that effectively models spatio-temporal information from Trajectories and Correlated regions. TCNet's trajectory module transforms frames into aligned trajectories composed of continuous visual tokens. In addition, for a query token, self-attention is learned along the trajectory. As such, our network can also focus on fine-grained spatio-temporal patterns, such as finger movements, of a specific region in motion. TCNet's correlation module uses a novel dynamic attention mechanism that filters out irrelevant frame regions. Additionally, it assigns dynamic key-value tokens from correlated regions to each query. Both innovations significantly reduce the computation cost and memory. We perform experiments on four large-scale datasets: PHOENIX14, PHOENIX14-T, CSL, and CSL-Daily, respectively. Our results demonstrate that TCNet consistently achieves state-of-the-art performance. For example, we improve over the previous state-of-the-art by 1.5% and 1.0% word error rate on PHOENIX14 and PHOENIX14-T, respectively.

Zhengmao Ye, Dengchun Li, Zetao Hu et al.

Transformer-based, pre-trained large language models (LLMs) have demonstrated outstanding performance across diverse domains, particularly in the emerging {\em pretrain-then-finetune} paradigm. Low-Rank Adaptation (LoRA), a parameter-efficient fine-tuning method, is commonly used to adapt a base LLM to multiple downstream tasks. Further, LLM platforms enable developers to fine-tune multiple models and develop various domain-specific applications simultaneously. However, existing model parallelism schemes suffer from high communication overhead and inefficient GPU utilization when training multiple LoRA tasks across GPUs and machines. In this paper, we present mLoRA, a parallelism-efficient fine-tuning system designed for training multiple LoRA across GPUs and machines. mLoRA introduces a novel LoRA-aware pipeline parallelism scheme that efficiently pipelines independent LoRA adapters and their distinct fine-tuning stages across GPUs and machines, along with a new LoRA-efficient operator to enhance GPU utilization during pipelined LoRA training. Our extensive evaluation shows that mLoRA can significantly reduce average fine-tuning task completion time, e.g., by 30\%, compared to state-of-the-art methods like FSDP. More importantly, mLoRA enables simultaneous fine-tuning of larger models, e.g., two Llama-2-13B models on four NVIDIA RTX A6000 48GB GPUs, which is not feasible for FSDP due to high memory requirements. Hence, mLoRA not only increases fine-tuning efficiency but also makes it more accessible on cost-effective GPUs. mLoRA has been deployed in AntGroup's production environment.

Shuhai Peng, Hui Lu, Jinjiang Liu et al.

While generative models have set new benchmarks for Target Speaker Extraction (TSE), their inherent reliance on global context precludes deployment in real-time applications. Direct adaptation to streaming scenarios often leads to catastrophic inference performance degradation due to the severe mismatch between training and streaming inference. To bridge this gap, we present the first autoregressive (AR) models tailored for streaming TSE. Our approach introduces a Chunk-wise Interleaved Splicing Paradigm that ensures highly efficient and stable streaming inference. To ensure the coherence between the extracted speech segments, we design a historical context refinement mechanism that mitigates boundary discontinuities by leveraging historical information. Experiments on Libri2Mix show that while AR generative baseline exhibits performance degradation at low latencies, our approach maintains 100% stability and superior intelligibility. Furthermore, our streaming results are comparable to or even surpass offline baselines. Additionally, our model achieves a Real-Time-Factor (RTF) of 0.248 on consumer-level GPUs. This work provides empirical evidence that AR generative backbones are viable for latency-sensitive applications through the Chunk-wise Interleaved Splicing Paradigm.

Hui Lu, Hu Jian, Ronald Poppe et al.

Owing to their ability to extract relevant spatio-temporal video embeddings, Vision Transformers (ViTs) are currently the best performing models in video action understanding. However, their generalization over domains or datasets is somewhat limited. In contrast, Visual Language Models (VLMs) have demonstrated exceptional generalization performance, but are currently unable to process videos. Consequently, they cannot extract spatio-temporal patterns that are crucial for action understanding. In this paper, we propose the Four-tiered Prompts (FTP) framework that takes advantage of the complementary strengths of ViTs and VLMs. We retain ViTs' strong spatio-temporal representation ability but improve the visual encodings to be more comprehensive and general by aligning them with VLM outputs. The FTP framework adds four feature processors that focus on specific aspects of human action in videos: action category, action components, action description, and context information. The VLMs are only employed during training, and inference incurs a minimal computation cost. Our approach consistently yields state-of-the-art performance. For instance, we achieve remarkable top-1 accuracy of 93.8% on Kinetics-400 and 83.4% on Something-Something V2, surpassing VideoMAEv2 by 2.8% and 2.6%, respectively.

Hui Lu, Albert ali Salah, Ronald Poppe

Diffusion models achieve remarkable quality in image generation, but at a cost. Iterative denoising requires many time steps to produce high fidelity images. We argue that the denoising process is crucially limited by an accumulation of the reconstruction error due to an initial inaccurate reconstruction of the target data. This leads to lower quality outputs, and slower convergence. To address this issue, we propose compensation sampling to guide the generation towards the target domain. We introduce a compensation term, implemented as a U-Net, which adds negligible computation overhead during training and, optionally, inference. Our approach is flexible and we demonstrate its application in unconditional generation, face inpainting, and face de-occlusion using benchmark datasets CIFAR-10, CelebA, CelebA-HQ, FFHQ-256, and FSG. Our approach consistently yields state-of-the-art results in terms of image quality, while accelerating the denoising process to converge during training by up to an order of magnitude.

Yinghao Tang, Tingfeng Lan, Xiuqi Huang et al.

Existing Large Language Model (LLM) serving systems prioritize maximum throughput. They often neglect Service Level Objectives (SLOs) such as Time to First Token (TTFT) and Time Per Output Token (TPOT), which leads to suboptimal SLO attainment. This paper introduces SCORPIO, an SLO-oriented LLM serving system designed to maximize system goodput and SLO attainment for workloads with heterogeneous SLOs. Our core insight is to exploit SLO heterogeneity for adaptive scheduling across admission control, queue management, and batch selection. SCORPIO features a TTFT Guard, which employs least-deadline-first reordering and rejects unattainable requests, and a TPOT Guard, which utilizes a VBS-based admission control and a novel credit-based batching mechanism. Both guards are supported by a predictive module. Evaluations demonstrate that SCORPIO improves system goodput by up to 14.4X and SLO adherence by up to 46.5% compared to state-of-the-art baselines.

Hui Lu, Yi Yu, Yiming Yang et al.

Vision-Language-Action (VLA) models are vulnerable to adversarial attacks, yet universal and transferable attacks remain underexplored, as most existing patches overfit to a single model and fail in black-box settings. To address this gap, we present a systematic study of universal, transferable adversarial patches against VLA-driven robots under unknown architectures, finetuned variants, and sim-to-real shifts. We introduce UPA-RFAS (Universal Patch Attack via Robust Feature, Attention, and Semantics), a unified framework that learns a single physical patch in a shared feature space while promoting cross-model transfer. UPA-RFAS combines (i) a feature-space objective with an $\ell_1$ deviation prior and repulsive InfoNCE loss to induce transferable representation shifts, (ii) a robustness-augmented two-phase min-max procedure where an inner loop learns invisible sample-wise perturbations and an outer loop optimizes the universal patch against this hardened neighborhood, and (iii) two VLA-specific losses: Patch Attention Dominance to hijack text$\to$vision attention and Patch Semantic Misalignment to induce image-text mismatch without labels. Experiments across diverse VLA models, manipulation suites, and physical executions show that UPA-RFAS consistently transfers across models, tasks, and viewpoints, exposing a practical patch-based attack surface and establishing a strong baseline for future defenses.

Hui Lu, Yi Yu, Shijian Lu et al.

Temporal Action Detection (TAD) aims to identify and localize actions by determining their starting and ending frames within untrimmed videos. Recent Structured State-Space Models such as Mamba have demonstrated potential in TAD due to their long-range modeling capability and linear computational complexity. On the other hand, structured state-space models often face two key challenges in TAD, namely, decay of temporal context due to recursive processing and self-element conflict during global visual context modeling, which become more severe while handling long-span action instances. Additionally, traditional methods for TAD struggle with detecting long-span action instances due to a lack of global awareness and inefficient detection heads. This paper presents MambaTAD, a new state-space TAD model that introduces long-range modeling and global feature detection capabilities for accurate temporal action detection. MambaTAD comprises two novel designs that complement each other with superior TAD performance. First, it introduces a Diagonal-Masked Bidirectional State-Space (DMBSS) module which effectively facilitates global feature fusion and temporal action detection. Second, it introduces a global feature fusion head that refines the detection progressively with multi-granularity features and global awareness. In addition, MambaTAD tackles TAD in an end-to-end one-stage manner using a new state-space temporal adapter(SSTA) which reduces network parameters and computation cost with linear complexity. Extensive experiments show that MambaTAD achieves superior TAD performance consistently across multiple public benchmarks.

Zhiyang Zhang, Ningcong Chen, Xin Zhang et al.

The widespread use of GPS devices has driven advances in spatiotemporal data mining, enabling machine learning models to simulate human decision making and generate realistic trajectories, addressing both data collection costs and privacy concerns. Recent studies have shown the promise of diffusion models for high-quality trajectory generation. However, most existing methods rely on convolution based architectures (e.g. UNet) to predict noise during the diffusion process, which often results in notable deviations and the loss of fine-grained street-level details due to limited model capacity. In this paper, we propose Trajectory Transformer, a novel model that employs a transformer backbone for both conditional information embedding and noise prediction. We explore two GPS coordinate embedding strategies, location embedding and longitude-latitude embedding, and analyze model performance at different scales. Experiments on two real-world datasets demonstrate that Trajectory Transformer significantly enhances generation quality and effectively alleviates the deviation issues observed in prior approaches.

Weishu Deng, Yujie Yang, Peiran Du et al.

Scaling inference for large language models (LLMs) is increasingly constrained by limited GPU memory, especially due to growing key-value (KV) caches required for long-context generation. While existing approaches offload KV caches to CPU memory or apply sparse attention to reduce GPU load, they often underutilize CPU compute resources and compromise accuracy. We present HGCA, a hybrid CPU-GPU attention mechanism that enables scalable, high-throughput LLM inference with near-full attention quality. HGCA performs dense attention on recently generated KV entries retained in GPU memory and parallel sparse attention on selected, salient KV entries in CPU memory. The attention outputs are efficiently merged using log-sum-exp fusion, minimizing PCIe transfer overhead. HGCA also introduces a finegrained, per-head sparsification strategy optimized for CPU execution, preserving contextual relevance while reducing computation. Our implementation seamlessly integrates into existing LLM frameworks without requiring model retraining. Experiments across diverse models and workloads show that HGCA achieves superior scalability, supports longer sequences and larger batch sizes, and outperforms existing sparse attention baselines in both performance and accuracy -- all on commodity GPU hardware.

Hui Lu, Albert Ali Salah, Ronald Poppe

Video understanding requires the extraction of rich spatio-temporal representations, which transformer models achieve through self-attention. Unfortunately, self-attention poses a computational burden. In NLP, Mamba has surfaced as an efficient alternative for transformers. However, Mamba's successes do not trivially extend to vision tasks, including those in video analysis. In this paper, we theoretically analyze the differences between self-attention and Mamba. We identify two limitations in Mamba's token processing: historical decay and element contradiction. We propose VideoMambaPro (VMP) that solves the identified limitations by adding masked backward computation and elemental residual connections to a VideoMamba backbone. Differently sized VideoMambaPro models surpass VideoMamba by 1.6-2.8% and 1.1-1.9% top-1 on Kinetics-400 and Something-Something V2, respectively. Even without extensive pre-training, our models present an increasingly attractive and efficient alternative to current transformer models. Moreover, our two solutions are orthogonal to recent advances in Vision Mamba models, and are likely to provide further improvements in future models.

Mingyuan Li, Tong Jia, Hui Lu et al.

Unlike natural images with occlusion-based overlap, X-ray images exhibit depth-induced superimposition and semi-transparent appearances, where objects at different depths overlap and their features blend together. These characteristics demand specialized mechanisms to disentangle mixed representations between target objects (e.g., prohibited items) and irrelevant backgrounds. While recent studies have explored adapting detection transformers (DETR) for anti-overlapping object detection, the importance of well-distributed content queries that represent object hypotheses remains underexplored. In this paper, we introduce a multi-class min-margin contrastive learning (MMCL) framework to correct the distribution of content queries, achieving balanced intra-class diversity and inter-class separability. The framework first groups content queries by object category and then applies two proposed complementary loss components: a multi-class exclusion loss to enhance inter-class separability, and a min-margin clustering loss to encourage intra-class diversity. We evaluate the proposed method on three widely used X-ray prohibited-item detection datasets, PIXray, OPIXray, and PIDray, using two backbone networks and four DETR variants. Experimental results demonstrate that MMCL effectively enhances anti-overlapping object detection and achieves state-of-the-art performance on both datasets. Code will be made publicly available on GitHub.

Yifan Peng, Ilia Kulikov, Yilin Yang et al.

There have been emerging research interest and advances in speech-to-speech translation (S2ST), translating utterances from one language to another. This work proposes Multitask Speech Language Model (MSLM), which is a decoder-only speech language model trained in a multitask setting. Without reliance on text training data, our model is able to support multilingual S2ST with speaker style preserved.

Yifan Peng, Ilia Kulikov, Yilin Yang et al.

Speech language models (LMs) are promising for high-quality speech synthesis through in-context learning. A typical speech LM takes discrete semantic units as content and a short utterance as prompt, and synthesizes speech which preserves the content's semantics but mimics the prompt's style. However, there is no systematic understanding on how the synthesized audio is controlled by the prompt and content. In this work, we conduct an empirical study of the widely used autoregressive (AR) and non-autoregressive (NAR) speech LMs and provide insights into the prompt design and content semantic units. Our analysis reveals that heterogeneous and nonstationary prompts hurt the audio quality in contrast to the previous finding that longer prompts always lead to better synthesis. Moreover, we find that the speaker style of the synthesized audio is also affected by the content in addition to the prompt. We further show that semantic units carry rich acoustic information such as pitch, tempo, volume and speech emphasis, which might be leaked from the content to the synthesized audio.

Shaojie Zhu, Zhaobin Wang, Chengxiang Zhuo et al.

CoT (Chain-of-Thought) is a way to solve reasoning problems for LLMs . Recently, many researches appear for improving the CoT capability of LLMs. In this work, we also proposed Olapa-MCoT, which is a LLMs based on llama2-13B PLM for finetuning and alignment learning. During the alignment training, we proposed the SimRRHF algorithm and Incorrect Data Relearning and mainly focused on optimizing the Chinese mathematical reasoning ability of Olapa-MCoT. The experiment achieved significant results, with the accuracy of Chinese mathematical reasoning up to 50%, 36% rise compared to llama2-13B. In addition, the accuracy of English reasoning ability also increased by nearly 4%.

Disong Wang, Songxiang Liu, Xixin Wu et al.

Dysarthric speech reconstruction (DSR), which aims to improve the quality of dysarthric speech, remains a challenge, not only because we need to restore the speech to be normal, but also must preserve the speaker's identity. The speaker representation extracted by the speaker encoder (SE) optimized for speaker verification has been explored to control the speaker identity. However, the SE may not be able to fully capture the characteristics of dysarthric speakers that are previously unseen. To address this research problem, we propose a novel multi-task learning strategy, i.e., adversarial speaker adaptation (ASA). The primary task of ASA fine-tunes the SE with the speech of the target dysarthric speaker to effectively capture identity-related information, and the secondary task applies adversarial training to avoid the incorporation of abnormal speaking patterns into the reconstructed speech, by regularizing the distribution of reconstructed speech to be close to that of reference speech with high quality. Experiments show that the proposed approach can achieve enhanced speaker similarity and comparable speech naturalness with a strong baseline approach. Compared with dysarthric speech, the reconstructed speech achieves 22.3% and 31.5% absolute word error rate reduction for speakers with moderate and moderate-severe dysarthria respectively. Our demo page is released here: https://wendison.github.io/ASA-DSR-demo/

Thomas Chen, Hui Lu, Teeramet Kunpittaya et al.

A zk-SNARK is a protocol that lets one party, the prover, prove to another party, the verifier, that a statement about some privately-held information is true without revealing the information itself. This paper describes technical foundations, current applications, and some novel applications of zk-SNARKs. Regarding technical foundations, we go over the Quadratic Arithmetic Program reduction and the Pinocchio protocol. We then go over financial security applications like Zcash and Tornado Cash, and zk-Rollup applications like zkEVM and Darkforest. We propose novel zk-SNARK protocols for private auctions and decentralized card games on the blockchain, providing code for the proposed applications. We conclude by touching on promising zk-SNARK innovations, such as zk-STARKs.

Xu Li, Xixin Wu, Hui Lu et al.

Existing approaches for anti-spoofing in automatic speaker verification (ASV) still lack generalizability to unseen attacks. The Res2Net approach designs a residual-like connection between feature groups within one block, which increases the possible receptive fields and improves the system's detection generalizability. However, such a residual-like connection is performed by a direct addition between feature groups without channel-wise priority. We argue that the information across channels may not contribute to spoofing cues equally, and the less relevant channels are expected to be suppressed before adding onto the next feature group, so that the system can generalize better to unseen attacks. This argument motivates the current work that presents a novel, channel-wise gated Res2Net (CG-Res2Net), which modifies Res2Net to enable a channel-wise gating mechanism in the connection between feature groups. This gating mechanism dynamically selects channel-wise features based on the input, to suppress the less relevant channels and enhance the detection generalizability. Three gating mechanisms with different structures are proposed and integrated into Res2Net. Experimental results conducted on ASVspoof 2019 logical access (LA) demonstrate that the proposed CG-Res2Net significantly outperforms Res2Net on both the overall LA evaluation set and individual difficult unseen attacks, which also outperforms other state-of-the-art single systems, depicting the effectiveness of our method.

Hui Lu, Zhiyong Wu, Xixin Wu et al.

This paper describes a variational auto-encoder based non-autoregressive text-to-speech (VAENAR-TTS) model. The autoregressive TTS (AR-TTS) models based on the sequence-to-sequence architecture can generate high-quality speech, but their sequential decoding process can be time-consuming. Recently, non-autoregressive TTS (NAR-TTS) models have been shown to be more efficient with the parallel decoding process. However, these NAR-TTS models rely on phoneme-level durations to generate a hard alignment between the text and the spectrogram. Obtaining duration labels, either through forced alignment or knowledge distillation, is cumbersome. Furthermore, hard alignment based on phoneme expansion can degrade the naturalness of the synthesized speech. In contrast, the proposed model of VAENAR-TTS is an end-to-end approach that does not require phoneme-level durations. The VAENAR-TTS model does not contain recurrent structures and is completely non-autoregressive in both the training and inference phases. Based on the VAE architecture, the alignment information is encoded in the latent variable, and attention-based soft alignment between the text and the latent variable is used in the decoder to reconstruct the spectrogram. Experiments show that VAENAR-TTS achieves state-of-the-art synthesis quality, while the synthesis speed is comparable with other NAR-TTS models.