arXiv subjects

Hui Chen

Publications and source records attributed to Hui Chen.

At least 91 records · Page 5Linked to original sources

arXiv2025

PruneHal: Reducing Hallucinations in Multi-modal Large Language Models through Adaptive KV Cache Pruning

While multi-modal large language models (MLLMs) have made significant progress in recent years, the issue of hallucinations remains a major challenge. To mitigate this phenomenon, existing solutions either introduce additional data for further training or incorporate external or internal information during inference. However, these approaches inevitably introduce extra computational costs. In this paper, we observe that hallucinations in MLLMs are strongly associated with insufficient attention allocated to visual tokens. In particular, the presence of redundant visual tokens disperses the model's attention, preventing it from focusing on the most informative ones. As a result, critical visual cues are often under-attended, which in turn exacerbates the occurrence of hallucinations. Building on this observation, we propose \textbf{PruneHal}, a training-free, simple yet effective method that leverages adaptive KV cache pruning to enhance the model's focus on critical visual information, thereby mitigating hallucinations. To the best of our knowledge, we are the first to apply token pruning for hallucination mitigation in MLLMs. Notably, our method don't require additional training and incurs nearly no extra inference cost. Moreover, PruneHal is model-agnostic and can be seamlessly integrated with different decoding strategies, including those specifically designed for hallucination mitigation. We evaluate PruneHal on several widely used hallucination evaluation benchmarks using four mainstream MLLMs, achieving robust and outstanding results that highlight the effectiveness and superiority of our method. Our code will be publicly available.

cs.CV↗

arXiv2025

Data-Adaptive Transformed Bilateral Tensor Low-Rank Representation for Clustering

Tensor low-rank representation (TLRR) has demonstrated significant success in image clustering. However, most existing methods rely on fixed transformations and suffer from poor robustness to noise. In this paper, we propose a novel transformed bilateral tensor low-rank representation model called TBTLRR, which introduces a data-adaptive tensor nuclear norm by learning arbitrary unitary transforms, allowing for more effective capture of global correlations. In addition, by leveraging the bilateral structure of latent tensor data, TBTLRR is able to exploit local correlations between image samples and features. Furthermore, TBTLRR integrates the $\ell_{1/2}$-norm and Frobenius norm regularization terms for better dealing with complex noise in real-world scenarios. To solve the proposed nonconvex model, we develop an efficient optimization algorithm inspired by the alternating direction method of multipliers (ADMM) and provide theoretical convergence. Extensive experiments validate its superiority over the state-of-the-art methods in clustering. The code will be available at https://github.com/xianchaoxiu/TBTLRR.

cs.CV↗

arXiv2025

AWSPNet: Attention-based Dual-Tree Wavelet Scattering Prototypical Network for MIMO Radar Target Recognition and Jamming Suppression

The increasing of digital radio frequency memory based electronic countermeasures poses a significant threat to the survivability and effectiveness of radar systems. These jammers can generate a multitude of deceptive false targets, overwhelming the radar's processing capabilities and masking targets. Consequently, the ability to robustly discriminate between true targets and complex jamming signals, especially in low signal-to-noise ratio (SNR) environments, is of importance. This paper introduces the attention-based dual-tree wavelet scattering prototypical network (AWSPNet), a deep learning framework designed for simultaneous radar target recognition and jamming suppression. The core of AWSPNet is the encoder that leverages the dual-tree complex wavelet transform to extract features that are inherently robust to noise and signal translations. These features are further refined by an attention mechanism and a pre-trained backbone network. To address the challenge of limited labeled data and enhance generalization, we employ a supervised contrastive learning strategy during the training phase. The classification is performed by a prototypical network, which is particularly effective in few-shot learning scenarios, enabling rapid adaptation to new signal types. We demonstrate the efficacy of our approach through extensive experiments. The results show that AWSPNet achieves 90.45\% accuracy at -6 dB SNR. Furthermore, we provide a physical interpretation of the network's inner workings through t-SNE visualizations, which analyze the feature separability at different stages of the model. Finally, by integrating AWSPNet with a time-domain sliding window approach, we present a complete algorithm capable of not only identifying but also effectively suppressing various types of jamming, thereby validating its potential for practical application in complex electromagnetic environments.

eess.SP↗

arXiv2025

CAIT: Triple-Win Compression towards High Accuracy, Fast Inference, and Favorable Transferability For ViTs

Vision Transformers (ViTs) have emerged as state-of-the-art models for various vision tasks recently. However, their heavy computation costs remain daunting for resource-limited devices. To address this, researchers have dedicated themselves to compressing redundant information in ViTs for acceleration. However, existing approaches generally sparsely drop redundant image tokens by token pruning or brutally remove channels by channel pruning, leading to a sub-optimal balance between model performance and inference speed. Moreover, they struggle when transferring compressed models to downstream vision tasks that require the spatial structure of images, such as semantic segmentation. To tackle these issues, we propose CAIT, a joint \underline{c}ompression method for ViTs that achieves a harmonious blend of high \underline{a}ccuracy, fast \underline{i}nference speed, and favorable \underline{t}ransferability to downstream tasks. Specifically, we introduce an asymmetric token merging (ATME) strategy to effectively integrate neighboring tokens. It can successfully compress redundant token information while preserving the spatial structure of images. On top of it, we further design a consistent dynamic channel pruning (CDCP) strategy to dynamically prune unimportant channels in ViTs. Thanks to CDCP, insignificant channels in multi-head self-attention modules of ViTs can be pruned uniformly, significantly enhancing the model compression. Extensive experiments on multiple benchmark datasets show that our proposed method can achieve state-of-the-art performance across various ViTs.

cs.CV↗

arXiv2025

PrefixKV: Adaptive Prefix KV Cache is What Vision Instruction-Following Models Need for Efficient Generation

Recently, large vision-language models (LVLMs) have rapidly gained popularity for their strong generation and reasoning capabilities given diverse multimodal inputs. However, these models incur significant computational and memory overhead during inference, which greatly hinders the efficient deployment in practical scenarios. The extensive key-value (KV) cache, necessitated by the lengthy input and output sequences, notably contributes to the high inference cost. Based on this, recent works have investigated ways to reduce the KV cache size for higher efficiency. Although effective, they generally overlook the distinct importance distributions of KV vectors across layers and maintain the same cache size for each layer during the next token prediction. This results in the significant contextual information loss for certain layers, leading to notable performance decline. To address this, we present PrefixKV, where "Prefix" means the top-ranked KV based on importance rather than position in the original sequence. It reframes the challenge of determining KV cache sizes for all layers into the task of searching for the optimal global prefix configuration. With an adaptive layer-wise KV retention recipe based on binary search, the maximum contextual information can thus be preserved in each layer, facilitating the generation. Extensive experiments demonstrate that our method achieves the state-of-the-art performance compared with others. It exhibits superior inference efficiency and generation quality trade-offs, showing promising potential for practical applications. Code is available at https://github.com/THU-MIG/PrefixKV.

cs.CV↗

arXiv2025

YOLOE: Real-Time Seeing Anything

Object detection and segmentation are widely employed in computer vision applications, yet conventional models like YOLO series, while efficient and accurate, are limited by predefined categories, hindering adaptability in open scenarios. Recent open-set methods leverage text prompts, visual cues, or prompt-free paradigm to overcome this, but often compromise between performance and efficiency due to high computational demands or deployment complexity. In this work, we introduce YOLOE, which integrates detection and segmentation across diverse open prompt mechanisms within a single highly efficient model, achieving real-time seeing anything. For text prompts, we propose Re-parameterizable Region-Text Alignment (RepRTA) strategy. It refines pretrained textual embeddings via a re-parameterizable lightweight auxiliary network and enhances visual-textual alignment with zero inference and transferring overhead. For visual prompts, we present Semantic-Activated Visual Prompt Encoder (SAVPE). It employs decoupled semantic and activation branches to bring improved visual embedding and accuracy with minimal complexity. For prompt-free scenario, we introduce Lazy Region-Prompt Contrast (LRPC) strategy. It utilizes a built-in large vocabulary and specialized embedding to identify all objects, avoiding costly language model dependency. Extensive experiments show YOLOE's exceptional zero-shot performance and transferability with high inference efficiency and low training cost. Notably, on LVIS, with 3$\times$ less training cost and 1.4$\times$ inference speedup, YOLOE-v8-S surpasses YOLO-Worldv2-S by 3.5 AP. When transferring to COCO, YOLOE-v8-L achieves 0.6 AP$^b$ and 0.4 AP$^m$ gains over closed-set YOLOv8-L with nearly 4$\times$ less training time. Code and models are available at https://github.com/THU-MIG/yoloe.

cs.CV↗

arXiv2025

VerlTool: Towards Holistic Agentic Reinforcement Learning with Tool Use

Reinforcement Learning with Verifiable Rewards (RLVR) has demonstrated success in enhancing LLM reasoning capabilities, but remains limited to single-turn interactions without tool integration. While recent Agentic Reinforcement Learning with Tool use (ARLT) approaches have emerged to address multi-turn tool interactions, existing works develop task-specific codebases that suffer from fragmentation, synchronous execution bottlenecks, and limited extensibility across domains. These inefficiencies hinder broader community adoption and algorithmic innovation. We introduce VerlTool, a unified and modular framework that addresses these limitations through systematic design principles. VerlTool provides four key contributions: (1) upstream alignment with VeRL ensuring compatibility and simplified maintenance, (2) unified tool management via standardized APIs supporting diverse modalities including code execution, search, SQL databases, and vision processing, (3) asynchronous rollout execution achieving near 2$\times$ speedup by eliminating synchronization bottlenecks, and (4) comprehensive evaluation demonstrating competitive performance across 6 ARLT domains. Our framework formalizes ARLT as multi-turn trajectories with multi-modal observation tokens (text/image/video), extending beyond single-turn RLVR paradigms. We train and evaluate models on mathematical reasoning, knowledge QA, SQL generation, visual reasoning, web search, and software engineering tasks, achieving results comparable to specialized systems while providing unified training infrastructure. The modular plugin architecture enables rapid tool integration requiring only lightweight Python definitions, significantly reducing development overhead and providing a scalable foundation for tool-augmented RL research. Our code is open-sourced at https://github.com/TIGER-AI-Lab/verl-tool.

cs.AI↗

arXiv2025

Source-Free Object Detection with Detection Transformer

Source-Free Object Detection (SFOD) enables knowledge transfer from a source domain to an unsupervised target domain for object detection without access to source data. Most existing SFOD approaches are either confined to conventional object detection (OD) models like Faster R-CNN or designed as general solutions without tailored adaptations for novel OD architectures, especially Detection Transformer (DETR). In this paper, we introduce Feature Reweighting ANd Contrastive Learning NetworK (FRANCK), a novel SFOD framework specifically designed to perform query-centric feature enhancement for DETRs. FRANCK comprises four key components: (1) an Objectness Score-based Sample Reweighting (OSSR) module that computes attention-based objectness scores on multi-scale encoder feature maps, reweighting the detection loss to emphasize less-recognized regions; (2) a Contrastive Learning with Matching-based Memory Bank (CMMB) module that integrates multi-level features into memory banks, enhancing class-wise contrastive learning; (3) an Uncertainty-weighted Query-fused Feature Distillation (UQFD) module that improves feature distillation through prediction quality reweighting and query feature fusion; and (4) an improved self-training pipeline with a Dynamic Teacher Updating Interval (DTUI) that optimizes pseudo-label quality. By leveraging these components, FRANCK effectively adapts a source-pre-trained DETR model to a target domain with enhanced robustness and generalization. Extensive experiments on several widely used benchmarks demonstrate that our method achieves state-of-the-art performance, highlighting its effectiveness and compatibility with DETR-based SFOD models.

cs.CV↗

arXiv2025

Modern Base Station Architecture: Enabling Passive Beamforming with Beyond Diagonal RISs

Beamforming plays a crucial role in millimeter wave (mmWave) communication systems to mitigate the severe attenuation inherent to this spectrum. However, the use of large active antenna arrays in conventional architectures often results in high implementation costs and excessive power consumption, limiting their practicality. As an alternative, deploying large arrays at transceivers using passive devices, such as reconfigurable intelligent surfaces (RISs), offers a more cost-effective and energy-efficient solution. In this paper, we investigate a promising base station (BS) architecture that integrates a beyond diagonal RIS (BD-RIS) within the BS to enable passive beamforming. By utilizing Takagi's decomposition and leveraging the effective beamforming vector, the RIS profile can be designed to enable passive beamforming directed toward the target. Through the beamforming analysis, we reveal that BD-RIS provides robust beamforming performance across various system configurations, whereas the traditional diagonal RIS (D-RIS) exhibits instability with increasing RIS size and decreasing BS-RIS separation-two critical factors in optimizing RIS-assisted systems. Comprehensive computer simulation results across various aspects validate the superiority of the proposed BS-integrated BD-RIS over conventional D-RIS architectures, showcasing performance comparable to active analog beamforming antenna arrays.

eess.SP↗

arXiv2025

Long Context Scaling: Divide and Conquer via Multi-Agent Question-driven Collaboration

Processing long contexts has become a critical capability for modern large language models (LLMs). Existing works leverage agent-based divide-and-conquer methods for processing long contexts. But these methods face crucial limitations, including prohibitive accumulated latency and amplified information loss from excessive agent invocations, and the disruption of inherent textual dependencies by immoderate partitioning. In this paper, we propose a novel multi-agent framework XpandA (Expand-Agent) coupled with question-driven workflow and dynamic partitioning for robust long-context processing. XpandA overcomes these limitations through: 1) dynamic partitioning of long texts, which adaptively modulates the filling rate of context windows for input sequences of vastly varying lengths; 2) question-guided protocol to update flat information ensembles within centralized shared memory, constructing consistent inter-agent knowledge across partitions; and 3) selectively replaying specific partitions based on the state-tracking of question-information couples to promote the resolution of inverted-order structures across partitions (e.g., flashbacks). We perform a comprehensive evaluation of XpandA on multiple long-context benchmarks with length varying from 1k to 1M, demonstrating XpandA's feasibility for processing ultra-long sequences and its significant effectiveness in enhancing the long-context capabilities of various LLMs by achieving 20\% improvements and 1.5x inference speedup over baselines of full-context, RAG and previous agent-based methods.

cs.CL↗

arXiv2025

Learning to Infer Unseen Single-/Multi-Attribute-Object Compositions with Graph Networks

Inferring the unseen attribute-object composition is critical to make machines learn to decompose and compose complex concepts like people. Most existing methods are limited to the composition recognition of single-attribute-object, and can hardly learn relations between the attributes and objects. In this paper, we propose an attribute-object semantic association graph model to learn the complex relations and enable knowledge transfer between primitives. With nodes representing attributes and objects, the graph can be constructed flexibly, which realizes both single- and multi-attribute-object composition recognition. In order to reduce mis-classifications of similar compositions (e.g., scratched screen and broken screen), driven by the contrastive loss, the anchor image feature is pulled closer to the corresponding label feature and pushed away from other negative label features. Specifically, a novel balance loss is proposed to alleviate the domain bias, where a model prefers to predict seen compositions. In addition, we build a large-scale MultiAttribute Dataset (MAD) with 116,099 images and 8,030 label categories for inferring unseen multi-attribute-object compositions. Along with MAD, we propose two novel metrics Hard and Soft to give a comprehensive evaluation in the multi-attribute setting. Experiments on MAD and two other single-attribute-object benchmarks (MIT-States and UT-Zappos50K) demonstrate the effectiveness of our approach.

cs.CV↗

arXiv2025

Joint Active RIS Configuration and User Power Control for Localization: A Neuroevolution-Based Approach

This paper studies user localization aided by a Reconfigurable Intelligent Surface (RIS). A feedback link from the Base Station (BS) to the user is adopted to enable dynamic power control of the user pilot transmissions in the uplink. A novel multi-agent algorithm for the joint control of the RIS phase configuration and the user transmit power is presented, which is based on a hybrid approach integrating NeuroEvolution (NE) and supervised learning. The proposed scheme requires only single-bit feedback messages for the uplink power control, supports RIS elements with discrete responses, and is numerically shown to outperform fingerprinting, deep reinforcement learning baselines and backpropagation-based position estimators.

cs.NI↗

arXiv2025

Integrated Cellular and LEO-based Positioning and Synchronization under User Mobility

This paper investigates the localization, synchronization, and speed estimation of a mobile user equipment (UE) leveraging integrated terrestrial and non-terrestrial networks (NTNs), in particular low Earth orbit (LEO) satellites. We focus on a minimal setup in which the UE received signal from only one base station (BS) and one LEO satellite. We derive a generic signal model accounting for mobility, clock and frequency offsets, based on which a hierarchy of simplified models are proposed and organized by computational complexity. Estimation algorithms are developed for each model to facilitate efficient and accurate parameter recovery. Rigorous simulations validate the effectiveness of the proposed models, demonstrating their suitability across diverse scenarios. The findings highlight how the trade-off between complexity and performance can be optimized for varying deployment environments and application requirements, offering valuable insights for 6G positioning and synchronization systems under user mobility.

eess.SP↗

arXiv2025

Temporal Scaling Law for Large Language Models

Recently, Large Language Models (LLMs) have been widely adopted in a wide range of tasks, leading to increasing attention towards the research on how scaling LLMs affects their performance. Existing works, termed Scaling Laws, have discovered that the final test loss of LLMs scales as power-laws with model size, computational budget, and dataset size. However, the temporal change of the test loss of an LLM throughout its pre-training process remains unexplored, though it is valuable in many aspects, such as selecting better hyperparameters \textit{directly} on the target LLM. In this paper, we propose the novel concept of Temporal Scaling Law, studying how the test loss of an LLM evolves as the training steps scale up. In contrast to modeling the test loss as a whole in a coarse-grained manner, we break it down and dive into the fine-grained test loss of each token position, and further develop a dynamic hyperbolic-law. Afterwards, we derive the much more precise temporal scaling law by studying the temporal patterns of the parameters in the dynamic hyperbolic-law. Results on both in-distribution (ID) and out-of-distribution (OOD) validation datasets demonstrate that our temporal scaling law accurately predicts the test loss of LLMs across training steps. Our temporal scaling law has broad practical applications. First, it enables direct and efficient hyperparameter selection on the target LLM, such as data mixture proportions. Secondly, viewing the LLM pre-training dynamics from the token position granularity provides some insights to enhance the understanding of LLM pre-training.

cs.CL↗

arXiv2025

Improving Data Curation of Software Vulnerability Patches through Uncertainty Quantification

The changesets (or patches) that fix open source software vulnerabilities form critical datasets for various machine learning security-enhancing applications, such as automated vulnerability patching and silent fix detection. These patch datasets are derived from extensive collections of historical vulnerability fixes, maintained in databases like the Common Vulnerabilities and Exposures list and the National Vulnerability Database. However, since these databases focus on rapid notification to the security community, they contain significant inaccuracies and omissions that have a negative impact on downstream software security quality assurance tasks. In this paper, we propose an approach employing Uncertainty Quantification (UQ) to curate datasets of publicly-available software vulnerability patches. Our methodology leverages machine learning models that incorporate UQ to differentiate between patches based on their potential utility. We begin by evaluating a number of popular UQ techniques, including Vanilla, Monte Carlo Dropout, and Model Ensemble, as well as homoscedastic and heteroscedastic models of noise. Our findings indicate that Model Ensemble and heteroscedastic models are the best choices for vulnerability patch datasets. Based on these UQ modeling choices, we propose a heuristic that uses UQ to filter out lower quality instances and select instances with high utility value from the vulnerability dataset. Using our approach, we observe an improvement in predictive performance and a significant reduction of model training time (i.e., energy consumption) for a state-of-the-art vulnerability prediction model.

cs.SE↗

arXiv2025

Active Inference Framework for Closed-Loop Sensing, Communication, and Control in UAV Systems

Integrated sensing and communication (ISAC) is a core technology for 6G, and its application to closed-loop sensing, communication, and control (SCC) enables various services. Existing SCC solutions often treat sensing and control separately, leading to suboptimal performance and resource usage. In this work, we introduce the active inference framework (AIF) into SCC-enabled unmanned aerial vehicle (UAV) systems for joint state estimation, control, and sensing resource allocation. By formulating a unified generative model, the problem reduces to minimizing variational free energy for inference and expected free energy for action planning. Simulation results show that both control cost and sensing cost are reduced relative to baselines.

eess.SP↗

arXiv2025

DSMoE: Matrix-Partitioned Experts with Dynamic Routing for Computation-Efficient Dense LLMs

As large language models continue to scale, computational costs and resource consumption have emerged as significant challenges. While existing sparsification methods like pruning reduce computational overhead, they risk losing model knowledge through parameter removal. This paper proposes DSMoE (Dynamic Sparse Mixture-of-Experts), a novel approach that achieves sparsification by partitioning pre-trained FFN layers into computational blocks. We implement adaptive expert routing using sigmoid activation and straight-through estimators, enabling tokens to flexibly access different aspects of model knowledge based on input complexity. Additionally, we introduce a sparsity loss term to balance performance and computational efficiency. Extensive experiments on LLaMA models demonstrate that under equivalent computational constraints, DSMoE achieves superior performance compared to existing pruning and MoE approaches across language modeling and downstream tasks, particularly excelling in generation tasks. Analysis reveals that DSMoE learns distinctive layerwise activation patterns, providing new insights for future MoE architecture design.

cs.CL↗

arXiv2025

Suppression of pair beam instabilities in a laboratory analogue of blazar pair cascades

The generation of dense electron-positron pair beams in the laboratory can enable direct tests of theoretical models of $γ$-ray bursts and active galactic nuclei. We have successfully achieved this using ultra-relativistic protons accelerated by the Super Proton Synchrotron at CERN. In the first application of this experimental platform, the stability of the pair beam is studied as it propagates through a metre-length plasma, analogous to TeV $γ$-ray induced pair cascades in the intergalactic medium. It has been argued that pair beam instabilities disrupt the cascade, thus accounting for the observed lack of reprocessed GeV emission from TeV blazars. If true this would remove the need for a moderate strength intergalactic magnetic field to explain the observations. We find that the pair beam instability is suppressed if the beam is not perfectly collimated or monochromatic, hence the lower limit to the intergalactic magnetic field inferred from $γ$-ray observations of blazars is robust.

astro-ph.HE↗