arXiv subjects

Xu Han

Publications and source records attributed to Xu Han.

At least 109 records · Page 6Linked to original sources

arXiv2025

Process Reinforcement through Implicit Rewards

Dense process rewards have proven a more effective alternative to the sparse outcome-level rewards in the inference-time scaling of large language models (LLMs), particularly in tasks requiring complex multi-step reasoning. While dense rewards also offer an appealing choice for the reinforcement learning (RL) of LLMs since their fine-grained rewards have the potential to address some inherent issues of outcome rewards, such as training efficiency and credit assignment, this potential remains largely unrealized. This can be primarily attributed to the challenges of training process reward models (PRMs) online, where collecting high-quality process labels is prohibitively expensive, making them particularly vulnerable to reward hacking. To address these challenges, we propose PRIME (Process Reinforcement through IMplicit rEwards), which enables online PRM updates using only policy rollouts and outcome labels through implict process rewards. PRIME combines well with various advantage functions and forgoes the dedicated reward model training phrase that existing approaches require, substantially reducing the development overhead. We demonstrate PRIME's effectiveness on competitional math and coding. Starting from Qwen2.5-Math-7B-Base, PRIME achieves a 15.1% average improvement across several key reasoning benchmarks over the SFT model. Notably, our resulting model, Eurus-2-7B-PRIME, surpasses Qwen2.5-Math-7B-Instruct on seven reasoning benchmarks with 10% of its training data.

cs.LG↗

arXiv2025

Cost-Optimal Grouped-Query Attention for Long-Context Modeling

Grouped-Query Attention (GQA) is a widely adopted strategy for reducing the computational cost of attention layers in large language models (LLMs). However, current GQA configurations are often suboptimal because they overlook how context length influences inference cost. Since inference cost grows with context length, the most cost-efficient GQA configuration should also vary accordingly. In this work, we analyze the relationship among context length, model size, GQA configuration, and model loss, and introduce two innovations: (1) we decouple the total head size from the hidden size, enabling more flexible control over attention FLOPs; and (2) we jointly optimize the model size and the GQA configuration to arrive at a better allocation of inference resources between attention layers and other components. Our analysis reveals that commonly used GQA configurations are highly suboptimal for long-context scenarios. More importantly, we propose a recipe for deriving cost-optimal GQA configurations. Our results show that for long-context scenarios, one should use fewer attention heads while scaling up model size. Configurations selected by our recipe can reduce both memory usage and FLOPs by more than 50% compared to Llama-3's GQA, with *no degradation in model capabilities*. Our findings offer valuable insights for designing efficient long-context LLMs. The code is available at https://www.github.com/THUNLP/cost-optimal-gqa .

cs.CL↗

arXiv2025

BurstEngine: an Efficient Distributed Framework for Training Transformers on Extremely Long Sequences of over 1M Tokens

Existing methods for training LLMs on long-sequence data, such as Tensor Parallelism and Context Parallelism, exhibit low Model FLOPs Utilization as sequence lengths and number of GPUs increase, especially when sequence lengths exceed 1M tokens. To address these challenges, we propose BurstEngine, an efficient framework designed to train LLMs on long-sequence data. BurstEngine introduces BurstAttention, an optimized distributed attention with lower communication cost than RingAttention. BurstAttention leverages topology-aware ring communication to fully utilize network bandwidth and incorporates fine-grained communication-computation overlap. Furthermore, BurstEngine introduces sequence-level selective checkpointing and fuses the language modeling head with the loss function to reduce memory cost. Additionally, BurstEngine introduces workload balance optimization for various types of attention masking. By integrating these optimizations, BurstEngine achieves a $1.2\times$ speedup with much lower memory overhead than the state-of-the-art baselines when training LLMs on extremely long sequences of over 1M tokens. We have made our code publicly available on GitHub: https://github.com/thunlp/BurstEngine.

cs.DC↗

arXiv2025

On the algebraic stretching dynamics of variable-density mixing in shock-bubble interaction

The mixing mechanism within a single-vortex has been a theoretical focus for decades, while remains unclear especially under variable-density (VD) scenario. This study investigates canonical single-vortex VD mixing in shock-bubble interactions (SBI) through high-resolution numerical simulations. Special attention is paid to examine the stretching dynamics and its impact on VD mixing within a single-vortex, and this problem is investigated by quantitatively characterizing the scalar dissipation rate (SDR), namely mixing rate, and its time integral, referred to as mixedness. We first examine single-vortex passive-scalar (PS) mixing with the absence of density difference. Under several assumptions, the single-vortex stretching rate illustrates an algebraic growth of the length of scalar strips. By incorporating diffusion process through the solution of the advection-diffusion equation along these stretched strips, a PS mixing model for SDR is proposed. Within this framework, density-gradient effects from two perspectives of stretching dynamics and diffusion process are discovered to challenge the extension of PS mixing model to VD mixing. First, the secondary baroclinic effect increases the VD stretching rate by the additional secondary baroclinic principal strain. Second, the density source effect suppresses the diffusion process. By accounting for both the secondary baroclinic effect on stretching and the density source effect on diffusion, a VD mixing model for SBI is further modified. This model establishes a quantitative relationship between the stretching dynamics and the evolution of the mixing rate and mixedness for single-vortex VD mixing over a broad range of Mach numbers. Furthermore, the essential role of stretching dynamics on the mixing rate is demonstrated by the derived dependence of the time-averaged mixing rate on the Peclet number Pe.

physics.flu-dyn↗

arXiv2025

MiniCPM-V 4.5: Cooking Efficient MLLMs via Architecture, Data, and Training Recipe

Multimodal Large Language Models (MLLMs) are undergoing rapid progress and represent the frontier of AI development. However, their training and inference efficiency have emerged as a core bottleneck in making MLLMs more accessible and scalable. To address the challenges, we present MiniCPM-V 4.5, an 8B parameter model designed for high efficiency and strong performance. We introduce three core improvements in model architecture, data strategy and training method: a unified 3D-Resampler model architecture for highly compact encoding over images and videos, a unified learning paradigm for document knowledge and text recognition without heavy data engineering, and a hybrid reinforcement learning strategy for proficiency in both short and long reasoning modes. Comprehensive experimental results in OpenCompass evaluation show that MiniCPM-V 4.5 surpasses widely used proprietary models such as GPT-4o-latest, and significantly larger open-source models such as Qwen2.5-VL 72B. Notably, the strong performance is achieved with remarkable efficiency. For example, on the widely adopted VideoMME benchmark, MiniCPM-V 4.5 achieves state-of-the-art performance among models under 30B size, using just 46.7\% GPU memory cost and 8.7\% inference time of Qwen2.5-VL 7B.

cs.LG↗

arXiv2025

PATIMT-Bench: A Multi-Scenario Benchmark for Position-Aware Text Image Machine Translation in Large Vision-Language Models

Text Image Machine Translation (TIMT) aims to translate texts embedded within an image into another language. Current TIMT studies primarily focus on providing translations for all the text within an image, while neglecting to provide bounding boxes and covering limited scenarios. In this work, we extend traditional TIMT into position-aware TIMT (PATIMT), aiming to support fine-grained and layoutpreserving translation, which holds great practical value but remains largely unexplored. This task comprises two key sub-tasks: regionspecific translation and full-image translation with grounding. To support existing models on PATIMT and conduct fair evaluation, we construct the PATIMT benchmark (PATIMTBench), which consists of 10 diverse real-world scenarios. Specifically, we introduce an Adaptive Image OCR Refinement Pipeline, which adaptively selects appropriate OCR tools based on scenario and refines the results of text-rich images. To ensure evaluation reliability, we further construct a test set, which contains 1,200 high-quality instances manually annotated and reviewed by human experts. After fine-tuning on our data, compact Large Vision-Language Models (LVLMs) achieve state-of-the-art performance on both sub-tasks. Experimental results also highlight the scalability and generalizability of our training data

cs.CV↗

arXiv2025

MiniCPM4: Ultra-Efficient LLMs on End Devices

This paper introduces MiniCPM4, a highly efficient large language model (LLM) designed explicitly for end-side devices. We achieve this efficiency through systematic innovation in four key dimensions: model architecture, training data, training algorithms, and inference systems. Specifically, in terms of model architecture, we propose InfLLM v2, a trainable sparse attention mechanism that accelerates both prefilling and decoding phases for long-context processing. Regarding training data, we propose UltraClean, an efficient and accurate pre-training data filtering and generation strategy, and UltraChat v2, a comprehensive supervised fine-tuning dataset. These datasets enable satisfactory model performance to be achieved using just 8 trillion training tokens. Regarding training algorithms, we propose ModelTunnel v2 for efficient pre-training strategy search, and improve existing post-training methods by introducing chunk-wise rollout for load-balanced reinforcement learning and data-efficient tenary LLM, BitCPM. Regarding inference systems, we propose CPM.cu that integrates sparse attention, model quantization, and speculative sampling to achieve efficient prefilling and decoding. To meet diverse on-device requirements, MiniCPM4 is available in two versions, with 0.5B and 8B parameters, respectively. Furthermore, we construct a hybrid reasoning model, MiniCPM4.1, which can be used in both deep reasoning mode and non-reasoning mode. Evaluation results demonstrate that MiniCPM4 and MiniCPM4.1 outperform similar-sized open-source models across benchmarks, with the 8B variants showing significant speed improvements on long sequence understanding and generation.

cs.CL↗

arXiv2025

Optimized Lattice-Structured Flexible EIT Sensor for Tactile Reconstruction and Classification

Flexible electrical impedance tomography (EIT) offers a promising alternative to traditional tactile sensing approaches, enabling low-cost, scalable, and deformable sensor designs. Here, we propose an optimized lattice-structured flexible EIT tactile sensor incorporating a hydrogel-based conductive layer, systematically designed through three-dimensional coupling field simulations to optimize structural parameters for enhanced sensitivity and robustness. By tuning the lattice channel width and conductive layer thickness, we achieve significant improvements in tactile reconstruction quality and classification performance. Experimental results demonstrate high-quality tactile reconstruction with correlation coefficients up to 0.9275, peak signal-to-noise ratios reaching 29.0303 dB, and structural similarity indexes up to 0.9660, while maintaining low relative errors down to 0.3798. Furthermore, the optimized sensor accurately classifies 12 distinct tactile stimuli with an accuracy reaching 99.6%. These results highlight the potential of simulation-guided structural optimization for advancing flexible EIT-based tactile sensors toward practical applications in wearable systems, robotics, and human-machine interfaces.

cs.RO↗

arXiv2025

Spin decoherence dynamics of Er$^{3+}$ in CeO$_2$ film

Developing telecom-compatible spin-photon interfaces is essential towards scalable quantum networks. Erbium ions (Er$^{3+}$) exhibit a unique combination of a telecom (1.5 $μ$m) optical transition and an effective spin-$1/2$ ground state, but identifying a host that enables heterogeneous device integration while preserving long optical and spin coherence remains an open challenge. We explore a new platform of Er$^{3+}$:CeO$_2$ films on silicon, offering low nuclear spin density and the potential for on-chip integration. We demonstrate a 38.8 $μ$s spin coherence, which can be extended to 176.4\nobreakspace $μ$s with dynamical decoupling. Pairing experiments with cluster correlation expansion calculations, we identify spectral diffusion-induced Er$^{3+}$ spin flips as the dominant decoherence mechanism and provide pathways to millisecond-scale coherence.

quant-ph↗

arXiv2025

PointDreamer: Zero-shot 3D Textured Mesh Reconstruction from Colored Point Cloud

Faithfully reconstructing textured meshes is crucial for many applications. Compared to text or image modalities, leveraging 3D colored point clouds as input (colored-PC-to-mesh) offers inherent advantages in comprehensively and precisely replicating the target object's 360° characteristics. While most existing colored-PC-to-mesh methods suffer from blurry textures or require hard-to-acquire 3D training data, we propose PointDreamer, a novel framework that harnesses 2D diffusion prior for superior texture quality. Crucially, unlike prior 2D-diffusion-for-3D works driven by text or image inputs, PointDreamer successfully adapts 2D diffusion models to 3D point cloud data by a novel project-inpaint-unproject pipeline. Specifically, it first projects the point cloud into sparse 2D images and then performs diffusion-based inpainting. After that, diverging from most existing 3D reconstruction or generation approaches that predict texture in 3D/UV space thus often yielding blurry texture, PointDreamer achieves high-quality texture by directly unprojecting the inpainted 2D images to the 3D mesh. Furthermore, we identify for the first time a typical kind of unprojection artifact appearing in occlusion borders, which is common in other multiview-image-to-3D pipelines but less-explored. To address this, we propose a novel solution named the Non-Border-First (NBF) unprojection strategy. Extensive qualitative and quantitative experiments on various synthetic and real-scanned datasets demonstrate that PointDreamer, though zero-shot, exhibits SoTA performance (30% improvement on LPIPS score from 0.118 to 0.068), and is robust to noisy, sparse, or even incomplete input data. Code at: https://github.com/YuQiao0303/PointDreamer.

cs.CV↗

arXiv2025

Fancy123: One Image to High-Quality 3D Mesh Generation via Plug-and-Play Deformation

Generating 3D meshes from a single image is an important but ill-posed task. Existing methods mainly adopt 2D multiview diffusion models to generate intermediate multiview images, and use the Large Reconstruction Model (LRM) to create the final meshes. However, the multiview images exhibit local inconsistencies, and the meshes often lack fidelity to the input image or look blurry. We propose Fancy123, featuring two enhancement modules and an unprojection operation to address the above three issues, respectively. The appearance enhancement module deforms the 2D multiview images to realign misaligned pixels for better multiview consistency. The fidelity enhancement module deforms the 3D mesh to match the input image. The unprojection of the input image and deformed multiview images onto LRM's generated mesh ensures high clarity, discarding LRM's predicted blurry-looking mesh colors. Extensive qualitative and quantitative experiments verify Fancy123's SoTA performance with significant improvement. Also, the two enhancement modules are plug-and-play and work at inference time, allowing seamless integration into various existing single-image-to-3D methods. Code at: https://github.com/YuQiao0303/Fancy123

cs.CV↗

arXiv2025

BlockFFN: Towards End-Side Acceleration-Friendly Mixture-of-Experts with Chunk-Level Activation Sparsity

To alleviate the computational burden of large language models (LLMs), architectures with activation sparsity, represented by mixture-of-experts (MoE), have attracted increasing attention. However, the non-differentiable and inflexible routing of vanilla MoE hurts model performance. Moreover, while each token activates only a few parameters, these sparsely-activated architectures exhibit low chunk-level sparsity, indicating that the union of multiple consecutive tokens activates a large ratio of parameters. Such a sparsity pattern is unfriendly for acceleration under low-resource conditions (e.g., end-side devices) and incompatible with mainstream acceleration techniques (e.g., speculative decoding). To address these challenges, we introduce a novel MoE architecture, BlockFFN, as well as its efficient training and deployment techniques. Specifically, we use a router integrating ReLU activation and RMSNorm for differentiable and flexible routing. Next, to promote both token-level sparsity (TLS) and chunk-level sparsity (CLS), CLS-aware training objectives are designed, making BlockFFN more acceleration-friendly. Finally, we implement efficient acceleration kernels, combining activation sparsity and speculative decoding for the first time. The experimental results demonstrate the superior performance of BlockFFN over other MoE baselines, achieving over 80% TLS and 70% 8-token CLS. Our kernels achieve up to 3.67$\times$ speedup on real end-side devices than dense models. All codes and checkpoints are available publicly (https://github.com/thunlp/BlockFFN).

cs.LG↗

arXiv2025

HypKG: Hypergraph-based Knowledge Graph Contextualization for Precision Healthcare

Knowledge graphs (KGs) are important products of the semantic web, which are widely used in various application domains. Healthcare is one of such domains where KGs are intensively used, due to the high requirement for knowledge accuracy and interconnected nature of healthcare data. However, KGs storing general factual information often lack the ability to account for important contexts of the knowledge such as the status of specific patients, which are crucial in precision healthcare. Meanwhile, electronic health records (EHRs) provide rich personal data, including various diagnoses and medications, which provide natural contexts for general KGs. In this paper, we propose HypKG, a framework that integrates patient information from EHRs into KGs to generate contextualized knowledge representations for accurate healthcare predictions. Using advanced entity-linking techniques, we connect relevant knowledge from general KGs with patient information from EHRs, and then utilize a hypergraph model to "contextualize" the knowledge with the patient information. Finally, we employ hypergraph transformers guided by downstream prediction tasks to jointly learn proper contextualized representations for both KGs and patients, fully leveraging existing knowledge in KGs and patient contexts in EHRs. In experiments using a large biomedical KG and two real-world EHR datasets, HypKG demonstrates significant improvements in healthcare prediction tasks across multiple evaluation metrics. Additionally, by integrating external contexts, HypKG can learn to adjust the representations of entities and relations in KG, potentially improving the quality and real-world utility of knowledge.

cs.AI↗

arXiv2025

CDA-SimBoost: A Unified Framework Bridging Real Data and Simulation for Infrastructure-Based CDA Systems

Cooperative Driving Automation (CDA) has garnered increasing research attention, yet the role of intelligent infrastructure remains insufficiently explored. Existing solutions offer limited support for addressing long-tail challenges, real-synthetic data fusion, and heterogeneous sensor management. This paper introduces CDA-SimBoost, a unified framework that constructs infrastructure-centric simulation environments from real-world data. CDA-SimBoost consists of three main components: a Digital Twin Builder for generating high-fidelity simulator assets based on sensor and HD map data, OFDataPip for processing both online and offline data streams, and OpenCDA-InfraX, a high-fidelity platform for infrastructure-focused simulation. The system supports realistic scenario construction, rare event synthesis, and scalable evaluation for CDA research. With its modular architecture and standardized benchmarking capabilities, CDA-SimBoost bridges real-world dynamics and virtual environments, facilitating reproducible and extensible infrastructure-driven CDA studies. All resources are publicly available at https://github.com/zhz03/CDA-SimBoost

eess.SY↗

arXiv2025

AutoTriton: Automatic Triton Programming with Reinforcement Learning in LLMs

Kernel development in deep learning requires optimizing computational units across hardware while balancing memory management, parallelism, and hardware-specific optimizations through extensive empirical tuning. Although domain-specific languages like Triton simplify GPU programming by abstracting low-level details, developers must still manually tune critical parameters such as tile sizes and memory access patterns through iterative experimentation, creating substantial barriers to optimal performance and wider adoption. In this work, we introduce AutoTriton, the first model dedicated to Triton programming powered by reinforcement learning (RL). AutoTriton performs supervised fine-tuning (SFT) to be equipped with essential Triton programming expertise using a high-quality data gathering pipeline, and conducts RL with Group Relative Policy Optimization (GRPO) algorithm, combining a rule-based reward and an execution-based reward to further improve Triton programming ability, sequentially. Experiments across five evaluation channels of TritonBench and KernelBench illustrate that our 8B model AutoTriton achieves performance comparable to mainstream large models, including Claude-4-Sonnet and DeepSeek-R1-0528. Further experimental analysis demonstrates the crucial role of each module within AutoTriton, including the SFT stage, the RL stage, and the reward design strategy. These findings underscore the promise of RL for automatically generating high-performance kernels, and since high-performance kernels are core components of AI systems, this breakthrough establishes an important foundation for building more efficient AI systems. The model and code will be available at https://github.com/AI9Stars/AutoTriton.

cs.LG↗

arXiv2025

PLUS: Plug-and-Play Enhanced Liver Lesion Diagnosis Model on Non-Contrast CT Scans

Focal liver lesions (FLL) are common clinical findings during physical examination. Early diagnosis and intervention of liver malignancies are crucial to improving patient survival. Although the current 3D segmentation paradigm can accurately detect lesions, it faces limitations in distinguishing between malignant and benign liver lesions, primarily due to its inability to differentiate subtle variations between different lesions. Furthermore, existing methods predominantly rely on specialized imaging modalities such as multi-phase contrast-enhanced CT and magnetic resonance imaging, whereas non-contrast CT (NCCT) is more prevalent in routine abdominal imaging. To address these limitations, we propose PLUS, a plug-and-play framework that enhances FLL analysis on NCCT images for arbitrary 3D segmentation models. In extensive experiments involving 8,651 patients, PLUS demonstrated a significant improvement with existing methods, improving the lesion-level F1 score by 5.66%, the malignant patient-level F1 score by 6.26%, and the benign patient-level F1 score by 4.03%. Our results demonstrate the potential of PLUS to improve malignant FLL screening using widely available NCCT imaging substantially.

eess.IV↗

arXiv2025

Significance of secondary baroclinic hydrodynamic instability on mixing enhancement in shock-bubble interaction

Different strength of hydrodynamic instability can be induced by the variations in the initial diffusion of shock bubble interaction (SBI), while the influence of hydrodynamic instability on variable-density mixing in SBI remains unclear. The present study aims to investigate the hydrodynamic instability of SBI through high-resolution numerical simulations. To isolate each factor within this instability, a circulation control method is employed to ensure consistent Reynolds number Re and Peclect number Pe. An examination of the morphology of the bubbles and vorticity dynamics reveals that the hydrodynamic instability can be characterized by positive circulation. Through vorticity budget analysis, the positive circulation is dominated by the baroclinic torque. Therefore, the identified hydrodynamic instability is labeled as secondary baroclinic hydrodynamic instability (SBHI). Based on the dimensional analysis of vorticity transport equation, a new dimensionless parameter, the secondary baroclinic vorticity (SBV) number, is proposed to characterize the strength of SBHI. Regarding mixing characteristics, cases with stronger SBHI exhibit higher mixing rates. Indicated by the temporal-averaged mixing rate with different SBV numbers, a scaling behavior is revealed: the mixing rate is increased proportionally to the square of SBV numbers. It is widely recognized that unstable flow can also be induced by a high Reynolds number Re. The distinction and connection of SBHI and high Re unstable flow are further studied. The scaling behavior of the mixing rate in SBHI shows distinct from the Reynolds number Re scaling: the mixing can hardly be altered effectively in the limit of large Re, the mechanisms of which are inherently different with respect to the stretch term, closely relating to the principal strain and the alignment angle between the scalar gradient and the principal strain axis.

physics.flu-dyn↗

arXiv2025

Sparsing Law: Towards Large Language Models with Greater Activation Sparsity

Activation sparsity denotes the existence of substantial weakly-contributed elements within activation outputs that can be eliminated, benefiting many important applications concerned with large language models (LLMs). Although promoting greater activation sparsity within LLMs deserves deep studies, existing works lack comprehensive and quantitative research on the correlation between activation sparsity and potentially influential factors. In this paper, we present a comprehensive study on the quantitative scaling properties and influential factors of the activation sparsity within decoder-only Transformer-based LLMs. Specifically, we propose PPL-$p\%$ sparsity, a precise and performance-aware activation sparsity metric that is applicable to any activation function. Through extensive experiments, we find several important phenomena. Firstly, different activation functions exhibit comparable performance but opposite training-time sparsity trends. The activation ratio (i.e., $1-\mathrm{sparsity\ ratio}$) evolves as a convergent increasing power-law and decreasing logspace power-law with the amount of training data for SiLU-activated and ReLU-activated LLMs, respectively. These demonstrate that ReLU is more efficient as the activation function than SiLU and can leverage more training data to improve activation sparsity. Secondly, the activation ratio linearly increases with the width-depth ratio below a certain bottleneck point, indicating the potential advantage of a deeper architecture at a fixed parameter scale. Finally, at similar width-depth ratios, we surprisingly find that the limit value of activation sparsity varies weakly with the parameter scale, i.e., the activation patterns within LLMs are insensitive to the parameter scale. These empirical laws towards LLMs with greater activation sparsity have important implications for making LLMs more efficient and interpretable.

cs.LG↗