arXiv subjects

Xin Lai

Publications and source records attributed to Xin Lai.

At least 37 records · Page 2Linked to original sources

arXiv2025

VideoZoomer: Reinforcement-Learned Temporal Focusing for Long Video Reasoning

Multimodal Large Language Models (MLLMs) have achieved remarkable progress in vision-language tasks yet remain limited in long video understanding due to the limited context window. Consequently, prevailing approaches tend to rely on uniform frame sampling or static pre-selection, which might overlook critical evidence and unable to correct its initial selection error during its reasoning process. To overcome these limitations, we propose VideoZoomer, a novel agentic framework that enables MLLMs to dynamically control their visual focus during reasoning. Starting from a coarse low-frame-rate overview, VideoZoomer invokes a temporal zoom tool to obtain high-frame-rate clips at autonomously chosen moments, thereby progressively gathering fine-grained evidence in a multi-turn interactive manner. Accordingly, we adopt a two-stage training strategy: a cold-start supervised fine-tuning phase on a curated dataset of distilled exemplar and reflection trajectories, followed by reinforcement learning to further refine the agentic policy. Extensive experiments demonstrate that our 7B model delivers diverse and complex reasoning patterns, yielding strong performance across a broad set of long video understanding and reasoning benchmarks. These emergent capabilities allow it to consistently surpass existing open-source models and even rival proprietary systems on challenging tasks, while achieving superior efficiency under reduced frame budgets.

cs.CV↗

arXiv2025

Making Every Head Count: Sparse Attention Without the Speed-Performance Trade-off

The design of Large Language Models (LLMs) has long been hampered by a fundamental conflict within their core attention mechanism: its remarkable expressivity is built upon a computational complexity of O(H N^2) that grows quadratically with the context size (N) and linearly with the number of heads (H). This standard implementation harbors significant computational redundancy, as all heads independently compute attention over the same sequence space. Existing sparse methods, meanwhile, often trade information integrity for computational efficiency. To resolve this efficiency-performance trade-off, we propose SPAttention, whose core contribution is the introduction of a new paradigm we term Principled Structural Sparsity. SPAttention does not merely drop connections but instead reorganizes the computational task by partitioning the total attention workload into balanced, non-overlapping distance bands, assigning each head a unique segment. This approach transforms the multi-head attention mechanism from H independent O(N^2) computations into a single, collaborative O(N^2) computation, fundamentally reducing complexity by a factor of H. The structured inductive bias compels functional specialization among heads, enabling a more efficient allocation of computational resources from redundant modeling to distinct dependencies across the entire sequence span. Our work demonstrates that thoughtfully designed structural sparsity can serve as an effective inductive bias that simultaneously improves both computational efficiency and model performance, opening a new avenue for the architectural design of next-generation, high-performance LLMs.

cs.LG↗

arXiv2025

Adaptive Lighting Control in Visible Light Systems: An Integrated Sensing, Communication, and Illumination Framework

Indoor visible light communication (VLC) is a promising sixth-generation (6G) technology, as its directional and sensitive optical signals are naturally suited for integrated sensing and communication (ISAC). However, current research mainly focuses on maximizing data rates and sensing accuracy, creating a conflict between high performance, high energy consumption, and user visual comfort. This paper proposes an adaptive integrated sensing, communication, and illumination (ISCI) framework that resolves this conflict by treating energy savings as a primary objective. The framework's mechanism first partitions the receiving plane using a geometric methodology, defining an activity area and a surrounding non-activity area to match distinct user requirements. User location, determined using non-line-of-sight (NLOS) sensing, then acts as a dynamic switch for the system's optimization objective. The system adaptively shifts between minimizing total transmit power while guaranteeing communication and illumination performance in the activity area and maximizing signal-to-noise ratio (SNR) uniformity in the non-activity area. Numerical results confirm that this adaptive ISCI approach achieves 53.59% energy savings over a non-adaptive system and improves SNR uniformity by 57.79%, while satisfying all illumination constraints and maintaining a mean localization error of 0.071 m.

eess.SY↗

arXiv2025

Mosaic Pruning: A Hierarchical Framework for Generalizable Pruning of Mixture-of-Experts Models

Sparse Mixture-of-Experts (SMoE) architectures have enabled a new frontier in scaling Large Language Models (LLMs), offering superior performance by activating only a fraction of their total parameters during inference. However, their practical deployment is severely hampered by substantial static memory overhead, as all experts must be loaded into memory. Existing post-training pruning methods, while reducing model size, often derive their pruning criteria from a single, general-purpose corpus. This leads to a critical limitation: a catastrophic performance degradation when the pruned model is applied to other domains, necessitating a costly re-pruning for each new domain. To address this generalization gap, we introduce Mosaic Pruning (MoP). The core idea of MoP is to construct a functionally comprehensive set of experts through a structured ``cluster-then-select" process. This process leverages a similarity metric that captures expert performance across different task domains to functionally cluster the experts, and subsequently selects the most representative expert from each cluster based on our proposed Activation Variability Score. Unlike methods that optimize for a single corpus, our proposed Mosaic Pruning ensures that the pruned model retains a functionally complementary set of experts, much like the tiles of a mosaic that together form a complete picture of the original model's capabilities, enabling it to handle diverse downstream tasks.Extensive experiments on various MoE models demonstrate the superiority of our approach. MoP significantly outperforms prior work, achieving a 7.24\% gain on general tasks and 8.92\% on specialized tasks like math reasoning and code generation.

cs.LG↗

arXiv2025

Self-Training with Dynamic Weighting for Robust Gradual Domain Adaptation

In this paper, we propose a new method called Self-Training with Dynamic Weighting (STDW), which aims to enhance robustness in Gradual Domain Adaptation (GDA) by addressing the challenge of smooth knowledge migration from the source to the target domain. Traditional GDA methods mitigate domain shift through intermediate domains and self-training but often suffer from inefficient knowledge migration or incomplete intermediate data. Our approach introduces a dynamic weighting mechanism that adaptively balances the loss contributions of the source and target domains during training. Specifically, we design an optimization framework governed by a time-varying hyperparameter $\varrho$ (progressing from 0 to 1), which controls the strength of domain-specific learning and ensures stable adaptation. The method leverages self-training to generate pseudo-labels and optimizes a weighted objective function for iterative model updates, maintaining robustness across intermediate domains. Experiments on rotated MNIST, color-shifted MNIST, portrait datasets, and the Cover Type dataset demonstrate that STDW outperforms existing baselines. Ablation studies further validate the critical role of $\varrho$'s dynamic scheduling in achieving progressive adaptation, confirming its effectiveness in reducing domain bias and improving generalization. This work provides both theoretical insights and a practical framework for robust gradual domain adaptation, with potential applications in dynamic real-world scenarios. The code is available at https://github.com/Dramwig/STDW.

cs.LG↗

arXiv2025

Ricci Flow on Weighted Digraphs with Balancing Factor

Ricci curvature and Ricci flow have proven to be powerful tools for analyzing the geometry of discrete structures, particularly on undirected graphs, where they have been applied to tasks ranging from community detection to graph representation learning. However, their development on directed graphs remains limited, with Ricci flow being especially underexplored. In this work, we introduce a rigorous formulation of Ricci flow on directed weighted graphs, which evolves edge weights while preserving distances, and establish both the existence and uniqueness of its solutions. To capture the essence of asymmetry in directed networks and to enhance the capability of modeling more flexible structures, we incorporate a node-wise balancing factor that regulates between outflow and inflow. Building on the continuous Ricci flow evolution framework, we propose a discrete Ricci flow algorithm that is applicable to numerical computing. Numerical studies on various directed graph examples demonstrate the capacity of the proposed flow to reveal structural asymmetry and dynamic evolutions.

math.DG↗

arXiv2025

Mini-o3: Scaling Up Reasoning Patterns and Interaction Turns for Visual Search

Recent advances in large multimodal models have leveraged image-based tools with reinforcement learning to tackle visual problems. However, existing open-source approaches often exhibit monotonous reasoning patterns and allow only a limited number of interaction turns, making them inadequate for difficult tasks that require trial-and-error exploration. In this work, we address this limitation by scaling up tool-based interactions and introduce Mini-o3, a system that executes deep, multi-turn reasoning -- spanning tens of steps -- and achieves state-of-the-art performance on challenging visual search tasks. Our recipe for reproducing OpenAI o3-style behaviors comprises three key components. First, we construct the Visual Probe Dataset, a collection of thousands of challenging visual search problems designed for exploratory reasoning. Second, we develop an iterative data collection pipeline to obtain cold-start trajectories that exhibit diverse reasoning patterns, including depth-first search, trial-and-error, and goal maintenance. Third, we propose an over-turn masking strategy that prevents penalization of over-turn responses (those that hit the maximum number of turns) during reinforcement learning, thereby balancing training-time efficiency with test-time scalability. Despite training with an upper bound of only six interaction turns, our model generates trajectories that naturally scale to tens of turns at inference time, with accuracy improving as the number of turns increases. Extensive experiments demonstrate that Mini-o3 produces rich reasoning patterns and deep thinking paths, effectively solving challenging visual search problems.

cs.CV↗

arXiv2025

VisionThink: Smart and Efficient Vision Language Model via Reinforcement Learning

Recent advancements in vision-language models (VLMs) have improved performance by increasing the number of visual tokens, which are often significantly longer than text tokens. However, we observe that most real-world scenarios do not require such an extensive number of visual tokens. While the performance drops significantly in a small subset of OCR-related tasks, models still perform accurately in most other general VQA tasks with only 1/4 resolution. Therefore, we propose to dynamically process distinct samples with different resolutions, and present a new paradigm for visual token compression, namely, VisionThink. It starts with a downsampled image and smartly decides whether it is sufficient for problem solving. Otherwise, the model could output a special token to request the higher-resolution image. Compared to existing Efficient VLM methods that compress tokens using fixed pruning ratios or thresholds, VisionThink autonomously decides whether to compress tokens case by case. As a result, it demonstrates strong fine-grained visual understanding capability on OCR-related tasks, and meanwhile saves substantial visual tokens on simpler tasks. We adopt reinforcement learning and propose the LLM-as-Judge strategy to successfully apply RL to general VQA tasks. Moreover, we carefully design a reward function and penalty mechanism to achieve a stable and reasonable image resize call ratio. Extensive experiments demonstrate the superiority, efficiency, and effectiveness of our method. Our code is available at https://github.com/dvlab-research/VisionThink.

cs.CV↗

arXiv2025

An improved point-to-surface contact algorithm with penalty method for peridynamics

It is significantly challenging to obtain accurate contact forces in peridynamics (PD) simulations due to the difficulty of surface particles identification, particularly for complex geometries. Here, an improved point-to-surface contact model is proposed for PD with high accuracy. First, the outer surface is identified using the eigenvalue method and then we construct a Verlet list to identify potential contact particle pairs efficiently. Subsequently, a point-to-surface contact search algorithm is utilized to determine precise contact locations with the penalty function method calculating the contact force. Finally, the accuracy of this point-to-surface contact model is validated through several representative contact examples. The results demonstrate that the point-to-surface contact model model can predict contact forces and deformations with high accuracy, aligning well with the classical Hertz contact theory solutions. This work presents a contact model for PD that automatically recognizes external surface particles and accurately calculates the contact force, which provides guidance for the study of multi-body contact as well as complex contact situations.

cs.CE↗

arXiv2025

LHQ-SVC: Lightweight and High Quality Singing Voice Conversion Modeling

Singing Voice Conversion (SVC) has emerged as a significant subfield of Voice Conversion (VC), enabling the transformation of one singer's voice into another while preserving musical elements such as melody, rhythm, and timbre. Traditional SVC methods have limitations in terms of audio quality, data requirements, and computational complexity. In this paper, we propose LHQ-SVC, a lightweight, CPU-compatible model based on the SVC framework and diffusion model, designed to reduce model size and computational demand without sacrificing performance. We incorporate features to improve inference quality, and optimize for CPU execution by using performance tuning tools and parallel computing frameworks. Our experiments demonstrate that LHQ-SVC maintains competitive performance, with significant improvements in processing speed and efficiency across different devices. The results suggest that LHQ-SVC can meet

cs.SD↗

arXiv2024

CausalVE: Face Video Privacy Encryption via Causal Video Prediction

Advanced facial recognition technologies and recommender systems with inadequate privacy technologies and policies for facial interactions increase concerns about bioprivacy violations. With the proliferation of video and live-streaming websites, public-face video distribution and interactions pose greater privacy risks. Existing techniques typically address the risk of sensitive biometric information leakage through various privacy enhancement methods but pose a higher security risk by corrupting the information to be conveyed by the interaction data, or by leaving certain biometric features intact that allow an attacker to infer sensitive biometric information from them. To address these shortcomings, in this paper, we propose a neural network framework, CausalVE. We obtain cover images by adopting a diffusion model to achieve face swapping with face guidance and use the speech sequence features and spatiotemporal sequence features of the secret video for dynamic video inference and prediction to obtain a cover video with the same number of frames as the secret video. In addition, we hide the secret video by using reversible neural networks for video hiding so that the video can also disseminate secret data. Numerous experiments prove that our CausalVE has good security in public video dissemination and outperforms state-of-the-art methods from a qualitative, quantitative, and visual point of view.

cs.CV↗

arXiv2024

Improved Genetic Algorithm Based on Greedy and Simulated Annealing Ideas for Vascular Robot Ordering Strategy

This study presents a comprehensive approach for optimizing the acquisition, utilization, and maintenance of ABLVR vascular robots in healthcare settings. Medical robotics, particularly in vascular treatments, necessitates precise resource allocation and optimization due to the complex nature of robot and operator maintenance. Traditional heuristic methods, though intuitive, often fail to achieve global optimization. To address these challenges, this research introduces a novel strategy, combining mathematical modeling, a hybrid genetic algorithm, and ARIMA time series forecasting. Considering the dynamic healthcare environment, our approach includes a robust resource allocation model for robotic vessels and operators. We incorporate the unique requirements of the adaptive learning process for operators and the maintenance needs of robotic components. The hybrid genetic algorithm, integrating simulated annealing and greedy approaches, efficiently solves the optimization problem. Additionally, ARIMA time series forecasting predicts the demand for vascular robots, further enhancing the adaptability of our strategy. Experimental results demonstrate the superiority of our approach in terms of optimization, transparency, and convergence speed from other state-of-the-art methods.

cs.NE↗

arXiv2024

Step-DPO: Step-wise Preference Optimization for Long-chain Reasoning of LLMs

Mathematical reasoning presents a significant challenge for Large Language Models (LLMs) due to the extensive and precise chain of reasoning required for accuracy. Ensuring the correctness of each reasoning step is critical. To address this, we aim to enhance the robustness and factuality of LLMs by learning from human feedback. However, Direct Preference Optimization (DPO) has shown limited benefits for long-chain mathematical reasoning, as models employing DPO struggle to identify detailed errors in incorrect answers. This limitation stems from a lack of fine-grained process supervision. We propose a simple, effective, and data-efficient method called Step-DPO, which treats individual reasoning steps as units for preference optimization rather than evaluating answers holistically. Additionally, we have developed a data construction pipeline for Step-DPO, enabling the creation of a high-quality dataset containing 10K step-wise preference pairs. We also observe that in DPO, self-generated data is more effective than data generated by humans or GPT-4, due to the latter's out-of-distribution nature. Our findings demonstrate that as few as 10K preference data pairs and fewer than 500 Step-DPO training steps can yield a nearly 3% gain in accuracy on MATH for models with over 70B parameters. Notably, Step-DPO, when applied to Qwen2-72B-Instruct, achieves scores of 70.8% and 94.0% on the test sets of MATH and GSM8K, respectively, surpassing a series of closed-source models, including GPT-4-1106, Claude-3-Opus, and Gemini-1.5-Pro. Our code, data, and models are available at https://github.com/dvlab-research/Step-DPO.

cs.LG↗

arXiv2024

LISA: Reasoning Segmentation via Large Language Model

Although perception systems have made remarkable advancements in recent years, they still rely on explicit human instruction or pre-defined categories to identify the target objects before executing visual recognition tasks. Such systems cannot actively reason and comprehend implicit user intention. In this work, we propose a new segmentation task -- reasoning segmentation. The task is designed to output a segmentation mask given a complex and implicit query text. Furthermore, we establish a benchmark comprising over one thousand image-instruction-mask data samples, incorporating intricate reasoning and world knowledge for evaluation purposes. Finally, we present LISA: large Language Instructed Segmentation Assistant, which inherits the language generation capabilities of multimodal Large Language Models (LLMs) while also possessing the ability to produce segmentation masks. We expand the original vocabulary with a token and propose the embedding-as-mask paradigm to unlock the segmentation capability. Remarkably, LISA can handle cases involving complex reasoning and world knowledge. Also, it demonstrates robust zero-shot capability when trained exclusively on reasoning-free datasets. In addition, fine-tuning the model with merely 239 reasoning segmentation data samples results in further performance enhancement. Both quantitative and qualitative experiments show our method effectively unlocks new reasoning segmentation capabilities for multimodal LLMs. Code, models, and data are available at https://github.com/dvlab-research/LISA.

cs.CV↗

arXiv2024

LongLoRA: Efficient Fine-tuning of Long-Context Large Language Models

We present LongLoRA, an efficient fine-tuning approach that extends the context sizes of pre-trained large language models (LLMs), with limited computation cost. Typically, training LLMs with long context sizes is computationally expensive, requiring extensive training hours and GPU resources. For example, training on the context length of 8192 needs 16x computational costs in self-attention layers as that of 2048. In this paper, we speed up the context extension of LLMs in two aspects. On the one hand, although dense global attention is needed during inference, fine-tuning the model can be effectively and efficiently done by sparse local attention. The proposed shifted sparse attention effectively enables context extension, leading to non-trivial computation saving with similar performance to fine-tuning with vanilla attention. Particularly, it can be implemented with only two lines of code in training, while being optional in inference. On the other hand, we revisit the parameter-efficient fine-tuning regime for context expansion. Notably, we find that LoRA for context extension works well under the premise of trainable embedding and normalization. LongLoRA combines this improved LoRA with S^2-Attn. LongLoRA demonstrates strong empirical results on various tasks on Llama2 models from 7B/13B to 70B. LongLoRA extends Llama2 7B from 4k context to 100k, or Llama2 70B to 32k on a single 8x A100 machine. LongLoRA extends models' context while retaining their original architectures, and is compatible with most existing techniques, like Flash-Attention2. In addition, we further conduct supervised fine-tuning with LongLoRA and our long instruction-following LongAlpaca dataset.

cs.CL↗

arXiv2024

LISA++: An Improved Baseline for Reasoning Segmentation with Large Language Model

While LISA effectively bridges the gap between segmentation and large language models to enable reasoning segmentation, it poses certain limitations: unable to distinguish different instances of the target region, and constrained by the pre-defined textual response formats. In this work, we introduce LISA++, an update to the existing LISA model, focusing on improving core functionalities while keeping the base architecture intact. The main enhancements in LISA++ include: \textbf{1) Enhanced Segmentation}: The instance segmentation ability has been added, providing a more detailed scene analysis along with the existing multi-region semantic segmentation. \textbf{2) More Natural Conversation}: Improved capability for multi-turn dialogue, with the ability to incorporate segmentation results directly into text responses, i.e., Segmentation in Dialogue (SiD). These improvements are achieved by curating the existing samples of generic segmentation datasets, aimed specifically at enhancing the segmentation and conversational skills without structural change and additional data sources. Comparative analysis with the original LISA model shows significant advancements in these areas, positioning LISA++ as a notable upgrade in visual understanding and interaction. LISA++'s adaptability and improved features highlight the versatility of the mask-as-embedding paradigm proposed by LISA, and the potential as a foundational model for diverse applications.

cs.CV↗

arXiv2023

An Updated Lagrangian Particle Hydrodynamics (ULPH)-NOSBPD Coupling Approach for Modeling Fluid-Structure Interaction Problems

In order to solve the fluid-structure interaction problem of Newtonian fluid, a fluid-structure interaction approach is proposed based on Non-ordinary State-based Peridynamics (NOSB-PD) and Updated Lagrangian particle Hydrodynamics (ULPH), to simulate the fluid-structure interaction problem in which large geometric deformation and material failure are considered. In the coupled framework, the NOSB-PD theory is used to describe the deformation and fracture of the solid material structure. ULPH is applied to describe the flow of Newtonian fluids due to its advantages in computational accuracy. The framework effectively utilizes the advantages of NOSB-PD theory for solving discontinuous problems and ULPH theory for solving fluid problems and has good computational stability and robustness. To deal with the interface of fluid structure, a fluid-structure coupling algorithm using pressure as the transmission medium is established. The dynamic model of solid structure and the PD-ULPH fluid-structure interaction model involving large deformation are verified by several numerical validations, which are in good agreement with the analytical solution, the available experimental data and other numerical results, that demonstrates the accuracy and effectiveness of the proposed method in solving the fluid-structure interaction problem. Overall, the fluid-structure interaction model based on ULPH and NOSB-PD established in this paper provides a new idea for the numerical solution of fluid-structure interaction and a promising approach for engineering design and experimental prediction.

physics.flu-dyn↗

arXiv2023

An Updated Lagrangian particle hydrodynamics (ULPH) implementation of heat conduction model in weakly-compressive fluid

Heat conduction is quite common in natural, industrial, and military applications. In this work, the updated Lagrangian particle hydrodynamics (ULPH) theory, is utilized and applied to solve heat conduction problems. Since heat conduction is a second-order problem, the high-order ULPH theory is employed to establish the governing equations of heat conduction in ULPH, which is then validated using various numerical simulations. In this work, numerical simulations have been carried out to solve both static heat conduction problems and dynamic heat convection problems. The results show good accuracy and capability of the ULPH heat conduction model, suggesting promising prospects of the ULPH theory in multiphysics problems. The findings of this paper suggest that ULPH is effective in addressing convective heat transfer problems.

physics.flu-dyn↗