Searcharxiv⌕ Search

arXiv subjects

Chenghao Zhang

Publications and source records attributed to Chenghao Zhang.

At least 37 records · Page 2Linked to original sources
arXiv2025

NVIDIA Nemotron 3: Efficient and Open Intelligence

We introduce the Nemotron 3 family of models - Nano, Super, and Ultra. These models deliver strong agentic, reasoning, and conversational capabilities. The Nemotron 3 family uses a Mixture-of-Experts hybrid Mamba-Transformer architecture to provide best-in-class throughput and context lengths of up to 1M tokens. Super and Ultra models are trained with NVFP4 and incorporate LatentMoE, a novel approach that improves model quality. The two larger models also include MTP layers for faster text generation. All Nemotron 3 models are post-trained using multi-environment reinforcement learning enabling reasoning, multi-step tool use, and support granular reasoning budget control. Nano, the smallest model, outperforms comparable models in accuracy while remaining extremely cost-efficient for inference. Super is optimized for collaborative agents and high-volume workloads such as IT ticket automation. Ultra, the largest model, provides state-of-the-art accuracy and reasoning performance. Nano is released together with its technical report and this white paper, while Super and Ultra will follow in the coming months. We will openly release the model weights, pre- and post-training software, recipes, and all data for which we hold redistribution rights.

cs.CL↗
arXiv2025

Nemotron 3 Nano: Open, Efficient Mixture-of-Experts Hybrid Mamba-Transformer Model for Agentic Reasoning

We present Nemotron 3 Nano 30B-A3B, a Mixture-of-Experts hybrid Mamba-Transformer language model. Nemotron 3 Nano was pretrained on 25 trillion text tokens, including more than 3 trillion new unique tokens over Nemotron 2, followed by supervised fine tuning and large-scale RL on diverse environments. Nemotron 3 Nano achieves better accuracy than our previous generation Nemotron 2 Nano while activating less than half of the parameters per forward pass. It achieves up to 3.3x higher inference throughput than similarly-sized open models like GPT-OSS-20B and Qwen3-30B-A3B-Thinking-2507, while also being more accurate on popular benchmarks. Nemotron 3 Nano demonstrates enhanced agentic, reasoning, and chat abilities and supports context lengths up to 1M tokens. We release both our pretrained Nemotron 3 Nano 30B-A3B Base and post-trained Nemotron 3 Nano 30B-A3B checkpoints on Hugging Face.

cs.CL↗
arXiv2025

Towards Mixed-Modal Retrieval for Universal Retrieval-Augmented Generation

Retrieval-Augmented Generation (RAG) has emerged as a powerful paradigm for enhancing large language models (LLMs) by retrieving relevant documents from an external corpus. However, existing RAG systems primarily focus on unimodal text documents, and often fall short in real-world scenarios where both queries and documents may contain mixed modalities (such as text and images). In this paper, we address the challenge of Universal Retrieval-Augmented Generation (URAG), which involves retrieving and reasoning over mixed-modal information to improve vision-language generation. To this end, we propose Nyx, a unified mixed-modal to mixed-modal retriever tailored for URAG scenarios. To mitigate the scarcity of realistic mixed-modal data, we introduce a four-stage automated pipeline for generation and filtering, leveraging web documents to construct NyxQA, a dataset comprising diverse mixed-modal question-answer pairs that better reflect real-world information needs. Building on this high-quality dataset, we adopt a two-stage training framework for Nyx: we first perform pre-training on NyxQA along with a variety of open-source retrieval datasets, followed by supervised fine-tuning using feedback from downstream vision-language models (VLMs) to align retrieval outputs with generative preferences. Experimental results demonstrate that Nyx not only performs competitively on standard text-only RAG benchmarks, but also excels in the more general and realistic URAG setting, significantly improving generation quality in vision-language tasks.

cs.CL↗
arXiv2025

RoSe: Robust Self-supervised Stereo Matching under Adverse Weather Conditions

Recent self-supervised stereo matching methods have made significant progress, but their performance significantly degrades under adverse weather conditions such as night, rain, and fog. We identify two primary weaknesses contributing to this performance degradation. First, adverse weather introduces noise and reduces visibility, making CNN-based feature extractors struggle with degraded regions like reflective and textureless areas. Second, these degraded regions can disrupt accurate pixel correspondences, leading to ineffective supervision based on the photometric consistency assumption. To address these challenges, we propose injecting robust priors derived from the visual foundation model into the CNN-based feature extractor to improve feature representation under adverse weather conditions. We then introduce scene correspondence priors to construct robust supervisory signals rather than relying solely on the photometric consistency assumption. Specifically, we create synthetic stereo datasets with realistic weather degradations. These datasets feature clear and adverse image pairs that maintain the same semantic context and disparity, preserving the scene correspondence property. With this knowledge, we propose a robust self-supervised training paradigm, consisting of two key steps: robust self-supervised scene correspondence learning and adverse weather distillation. Both steps aim to align underlying scene results from clean and adverse image pairs, thus improving model disparity estimation under adverse weather effects. Extensive experiments demonstrate the effectiveness and versatility of our proposed solution, which outperforms existing state-of-the-art self-supervised methods. Codes are available at \textcolor{blue}{https://github.com/cocowy1/RoSe-Robust-Self-supervised-Stereo-Matching-under-Adverse-Weather-Conditions}.

cs.CV↗
arXiv2025

Generate the browsing process for short-video recommendation

This paper proposes a generative method to dynamically simulate users' short video watching journey for watch time prediction in short video recommendation. Unlike existing methods that rely on multimodal features for video content understanding, our method simulates users' sustained interest in watching short videos by learning collaborative information, using interest changes from existing positive and negative feedback videos and user interaction behaviors to implicitly model users' video watching journey. By segmenting videos based on duration and adopting a Transformer-like architecture, our method can capture sequential dependencies between segments while mitigating duration bias. Extensive experiments on industrial-scale and public datasets demonstrate that our method achieves state-of-the-art performance on watch time prediction tasks. The method has been deployed on Kuaishou Lite, achieving a significant improvement of +0.13\% in APP duration, and reaching an XAUC of 83\% for single video watch time prediction on industrial-scale streaming training sets, far exceeding other methods. The proposed method provides a scalable and effective solution for video recommendation through segment-level modeling and user engagement feedback.

cs.IR↗
arXiv2025

S-BEVLoc: BEV-based Self-supervised Framework for Large-scale LiDAR Global Localization

LiDAR-based global localization is an essential component of simultaneous localization and mapping (SLAM), which helps loop closure and re-localization. Current approaches rely on ground-truth poses obtained from GPS or SLAM odometry to supervise network training. Despite the great success of these supervised approaches, substantial cost and effort are required for high-precision ground-truth pose acquisition. In this work, we propose S-BEVLoc, a novel self-supervised framework based on bird's-eye view (BEV) for LiDAR global localization, which eliminates the need for ground-truth poses and is highly scalable. We construct training triplets from single BEV images by leveraging the known geographic distances between keypoint-centered BEV patches. Convolutional neural network (CNN) is used to extract local features, and NetVLAD is employed to aggregate global descriptors. Moreover, we introduce SoftCos loss to enhance learning from the generated triplets. Experimental results on the large-scale KITTI and NCLT datasets show that S-BEVLoc achieves state-of-the-art performance in place recognition, loop closure, and global localization tasks, while offering scalability that would require extra effort for supervised approaches.

cs.CV↗
arXiv2025

CITE: A Comprehensive Benchmark for Heterogeneous Text-Attributed Graphs on Catalytic Materials

Text-attributed graphs(TAGs) are pervasive in real-world systems,where each node carries its own textual features. In many cases these graphs are inherently heterogeneous, containing multiple node types and diverse edge types. Despite the ubiquity of such heterogeneous TAGs, there remains a lack of large-scale benchmark datasets. This shortage has become a critical bottleneck, hindering the development and fair comparison of representation learning methods on heterogeneous text-attributed graphs. In this paper, we introduce CITE - Catalytic Information Textual Entities Graph, the first and largest heterogeneous text-attributed citation graph benchmark for catalytic materials. CITE comprises over 438K nodes and 1.2M edges, spanning four relation types. In addition, we establish standardized evaluation procedures and conduct extensive benchmarking on the node classification task, as well as ablation experiments on the heterogeneous and textual properties of CITE. We compare four classes of learning paradigms, including homogeneous graph models, heterogeneous graph models, LLM(Large Language Model)-centric models, and LLM+Graph models. In a nutshell, we provide (i) an overview of the CITE dataset, (ii) standardized evaluation protocols, and (iii) baseline and ablation experiments across diverse modeling paradigms.

cs.LG↗
arXiv2025

AddressVLM: Cross-view Alignment Tuning for Image Address Localization using Large Vision-Language Models

Large visual language models (LVLMs) have demonstrated impressive performance in coarse-grained geo-localization at the country or city level, but they struggle with fine-grained street-level localization within urban areas. In this paper, we explore integrating city-wide address localization capabilities into LVLMs, facilitating flexible address-related question answering using street-view images. A key challenge is that the street-view visual question-and-answer (VQA) data provides only microscopic visual cues, leading to subpar performance in fine-tuned models. To tackle this issue, we incorporate perspective-invariant satellite images as macro cues and propose cross-view alignment tuning including a satellite-view and street-view image grafting mechanism, along with an automatic label generation mechanism. Then LVLM's global understanding of street distribution is enhanced through cross-view matching. Our proposed model, named AddressVLM, consists of two-stage training protocols: cross-view alignment tuning and address localization tuning. Furthermore, we have constructed two street-view VQA datasets based on image address localization datasets from Pittsburgh and San Francisco. Qualitative and quantitative evaluations demonstrate that AddressVLM outperforms counterpart LVLMs by over 9% and 12% in average address localization accuracy on these two datasets, respectively.

cs.CV↗
arXiv2025

Learning Robust Stereo Matching in the Wild with Selective Mixture-of-Experts

Recently, learning-based stereo matching networks have advanced significantly. However, they often lack robustness and struggle to achieve impressive cross-domain performance due to domain shifts and imbalanced disparity distributions among diverse datasets. Leveraging Vision Foundation Models (VFMs) can intuitively enhance the model's robustness, but integrating such a model into stereo matching cost-effectively to fully realize their robustness remains a key challenge. To address this, we propose SMoEStereo, a novel framework that adapts VFMs for stereo matching through a tailored, scene-specific fusion of Low-Rank Adaptation (LoRA) and Mixture-of-Experts (MoE) modules. SMoEStereo introduces MoE-LoRA with adaptive ranks and MoE-Adapter with adaptive kernel sizes. The former dynamically selects optimal experts within MoE to adapt varying scenes across domains, while the latter injects inductive bias into frozen VFMs to improve geometric feature extraction. Importantly, to mitigate computational overhead, we further propose a lightweight decision network that selectively activates MoE modules based on input complexity, balancing efficiency with accuracy. Extensive experiments demonstrate that our method exhibits state-of-the-art cross-domain and joint generalization across multiple benchmarks without dataset-specific adaptation. The code is available at \textcolor{red}{https://github.com/cocowy1/SMoE-Stereo}.

cs.CV↗
arXiv2025

Re-ranking Reasoning Context with Tree Search Makes Large Vision-Language Models Stronger

Recent advancements in Large Vision Language Models (LVLMs) have significantly improved performance in Visual Question Answering (VQA) tasks through multimodal Retrieval-Augmented Generation (RAG). However, existing methods still face challenges, such as the scarcity of knowledge with reasoning examples and erratic responses from retrieved knowledge. To address these issues, in this study, we propose a multimodal RAG framework, termed RCTS, which enhances LVLMs by constructing a Reasoning Context-enriched knowledge base and a Tree Search re-ranking method. Specifically, we introduce a self-consistent evaluation mechanism to enrich the knowledge base with intrinsic reasoning patterns. We further propose a Monte Carlo Tree Search with Heuristic Rewards (MCTS-HR) to prioritize the most relevant examples. This ensures that LVLMs can leverage high-quality contextual reasoning for better and more consistent responses. Extensive experiments demonstrate that our framework achieves state-of-the-art performance on multiple VQA datasets, significantly outperforming In-Context Learning (ICL) and Vanilla-RAG methods. It highlights the effectiveness of our knowledge base and re-ranking method in improving LVLMs. Our code is available at https://github.com/yannqi/RCTS-RAG.

cs.CV↗
arXiv2025

SSHNet: Unsupervised Cross-modal Homography Estimation via Problem Reformulation and Split Optimization

We propose a novel unsupervised cross-modal homography estimation learning framework, named Split Supervised Homography estimation Network (SSHNet). SSHNet reformulates the unsupervised cross-modal homography estimation into two supervised sub-problems, each addressed by its specialized network: a homography estimation network and a modality transfer network. To realize stable training, we introduce an effective split optimization strategy to train each network separately within its respective sub-problem. We also formulate an extra homography feature space supervision to enhance feature consistency, further boosting the estimation accuracy. Moreover, we employ a simple yet effective distillation training technique to reduce model parameters and improve cross-domain generalization ability while maintaining comparable performance. The training stability of SSHNet enables its cooperation with various homography estimation architectures. Experiments reveal that the SSHNet using IHN as homography estimation network, namely SSHNet-IHN, outperforms previous unsupervised approaches by a significant margin. Even compared to supervised approaches MHN and LocalTrans, SSHNet-IHN achieves 47.4% and 85.8% mean average corner errors (MACEs) reduction on the challenging OPT-SAR dataset.

cs.CV↗
arXiv2025

VenusMutHub: A systematic evaluation of protein mutation effect predictors on small-scale experimental data

In protein engineering, while computational models are increasingly used to predict mutation effects, their evaluations primarily rely on high-throughput deep mutational scanning (DMS) experiments that use surrogate readouts, which may not adequately capture the complex biochemical properties of interest. Many proteins and their functions cannot be assessed through high-throughput methods due to technical limitations or the nature of the desired properties, and this is particularly true for the real industrial application scenario. Therefore, the desired testing datasets, will be small-size (~10-100) experimental data for each protein, and involve as many proteins as possible and as many properties as possible, which is, however, lacking. Here, we present VenusMutHub, a comprehensive benchmark study using 905 small-scale experimental datasets curated from published literature and public databases, spanning 527 proteins across diverse functional properties including stability, activity, binding affinity, and selectivity. These datasets feature direct biochemical measurements rather than surrogate readouts, providing a more rigorous assessment of model performance in predicting mutations that affect specific molecular functions. We evaluate 23 computational models across various methodological paradigms, such as sequence-based, structure-informed and evolutionary approaches. This benchmark provides practical guidance for selecting appropriate prediction methods in protein engineering applications where accurate prediction of specific functional properties is crucial.

q-bio.QM↗
arXiv2025

FlashRAG: A Modular Toolkit for Efficient Retrieval-Augmented Generation Research

With the advent of large language models (LLMs) and multimodal large language models (MLLMs), the potential of retrieval-augmented generation (RAG) has attracted considerable research attention. Various novel algorithms and models have been introduced to enhance different aspects of RAG systems. However, the absence of a standardized framework for implementation, coupled with the inherently complex RAG process, makes it challenging and time-consuming for researchers to compare and evaluate these approaches in a consistent environment. Existing RAG toolkits, such as LangChain and LlamaIndex, while available, are often heavy and inflexibly, failing to meet the customization needs of researchers. In response to this challenge, we develop \ours{}, an efficient and modular open-source toolkit designed to assist researchers in reproducing and comparing existing RAG methods and developing their own algorithms within a unified framework. Our toolkit has implemented 16 advanced RAG methods and gathered and organized 38 benchmark datasets. It has various features, including a customizable modular framework, multimodal RAG capabilities, a rich collection of pre-implemented RAG works, comprehensive datasets, efficient auxiliary pre-processing scripts, and extensive and standard evaluation metrics. Our toolkit and resources are available at https://github.com/RUC-NLPIR/FlashRAG.

cs.CL↗
arXiv2025

CoL3D: Collaborative Learning of Single-view Depth and Camera Intrinsics for Metric 3D Shape Recovery

Recovering the metric 3D shape from a single image is particularly relevant for robotics and embodied intelligence applications, where accurate spatial understanding is crucial for navigation and interaction with environments. Usually, the mainstream approaches achieve it through monocular depth estimation. However, without camera intrinsics, the 3D metric shape can not be recovered from depth alone. In this study, we theoretically demonstrate that depth serves as a 3D prior constraint for estimating camera intrinsics and uncover the reciprocal relations between these two elements. Motivated by this, we propose a collaborative learning framework for jointly estimating depth and camera intrinsics, named CoL3D, to learn metric 3D shapes from single images. Specifically, CoL3D adopts a unified network and performs collaborative optimization at three levels: depth, camera intrinsics, and 3D point clouds. For camera intrinsics, we design a canonical incidence field mechanism as a prior that enables the model to learn the residual incident field for enhanced calibration. Additionally, we incorporate a shape similarity measurement loss in the point cloud space, which improves the quality of 3D shapes essential for robotic applications. As a result, when training and testing on a single dataset with in-domain settings, CoL3D delivers outstanding performance in both depth estimation and camera calibration across several indoor and outdoor benchmark datasets, which leads to remarkable 3D shape quality for the perception capabilities of robots.

cs.CV↗
arXiv2024

Progressive Multimodal Reasoning via Active Retrieval

Multi-step multimodal reasoning tasks pose significant challenges for multimodal large language models (MLLMs), and finding effective ways to enhance their performance in such scenarios remains an unresolved issue. In this paper, we propose AR-MCTS, a universal framework designed to progressively improve the reasoning capabilities of MLLMs through Active Retrieval (AR) and Monte Carlo Tree Search (MCTS). Our approach begins with the development of a unified retrieval module that retrieves key supporting insights for solving complex reasoning problems from a hybrid-modal retrieval corpus. To bridge the gap in automated multimodal reasoning verification, we employ the MCTS algorithm combined with an active retrieval mechanism, which enables the automatic generation of step-wise annotations. This strategy dynamically retrieves key insights for each reasoning step, moving beyond traditional beam search sampling to improve the diversity and reliability of the reasoning space. Additionally, we introduce a process reward model that aligns progressively to support the automatic verification of multimodal reasoning tasks. Experimental results across three complex multimodal reasoning benchmarks confirm the effectiveness of the AR-MCTS framework in enhancing the performance of various multimodal models. Further analysis demonstrates that AR-MCTS can optimize sampling diversity and accuracy, yielding reliable multimodal reasoning.

cs.CL↗
arXiv2024

Understand What LLM Needs: Dual Preference Alignment for Retrieval-Augmented Generation

Retrieval-augmented generation (RAG) has demonstrated effectiveness in mitigating the hallucination problem of large language models (LLMs). However, the difficulty of aligning the retriever with the diverse LLMs' knowledge preferences inevitably poses an inevitable challenge in developing a reliable RAG system. To address this issue, we propose DPA-RAG, a universal framework designed to align diverse knowledge preferences within RAG systems. Specifically, we initially introduce a preference knowledge construction pipline and incorporate five novel query augmentation strategies to alleviate preference data scarcity. Based on preference data, DPA-RAG accomplishes both external and internal preference alignment: 1) It jointly integrate pair-wise, point-wise, and contrastive preference alignment abilities into the reranker, achieving external preference alignment among RAG components. 2) It further introduces a pre-aligned stage before vanilla Supervised Fine-tuning (SFT), enabling LLMs to implicitly capture knowledge aligned with their reasoning preferences, achieving LLMs' internal alignment. Experimental results across four knowledge-intensive QA datasets demonstrate that DPA-RAG outperforms all baselines and seamlessly integrates both black-box and open-sourced LLM readers. Further qualitative analysis and discussions also provide empirical guidance for achieving reliable RAG systems. Our code is publicly available at https://github.com/dongguanting/DPA-RAG.

cs.CL↗
arXiv2024

AddressCLIP: Empowering Vision-Language Models for City-wide Image Address Localization

In this study, we introduce a new problem raised by social media and photojournalism, named Image Address Localization (IAL), which aims to predict the readable textual address where an image was taken. Existing two-stage approaches involve predicting geographical coordinates and converting them into human-readable addresses, which can lead to ambiguity and be resource-intensive. In contrast, we propose an end-to-end framework named AddressCLIP to solve the problem with more semantics, consisting of two key ingredients: i) image-text alignment to align images with addresses and scene captions by contrastive learning, and ii) image-geography matching to constrain image features with the spatial distance in terms of manifold learning. Additionally, we have built three datasets from Pittsburgh and San Francisco on different scales specifically for the IAL problem. Experiments demonstrate that our approach achieves compelling performance on the proposed datasets and outperforms representative transfer learning methods for vision-language models. Furthermore, extensive ablations and visualizations exhibit the effectiveness of the proposed method. The datasets and source code are available at https://github.com/xsx1001/AddressCLIP.

cs.CV↗
arXiv2024

INTERS: Unlocking the Power of Large Language Models in Search with Instruction Tuning

Large language models (LLMs) have demonstrated impressive capabilities in various natural language processing tasks. Despite this, their application to information retrieval (IR) tasks is still challenging due to the infrequent occurrence of many IR-specific concepts in natural language. While prompt-based methods can provide task descriptions to LLMs, they often fall short in facilitating a comprehensive understanding and execution of IR tasks, thereby limiting LLMs' applicability. To address this gap, in this work, we explore the potential of instruction tuning to enhance LLMs' proficiency in IR tasks. We introduce a novel instruction tuning dataset, INTERS, encompassing 20 tasks across three fundamental IR categories: query understanding, document understanding, and query-document relationship understanding. The data are derived from 43 distinct datasets with manually written templates. Our empirical results reveal that INTERS significantly boosts the performance of various publicly available LLMs, such as LLaMA, Mistral, and Phi, in IR tasks. Furthermore, we conduct extensive experiments to analyze the effects of instruction design, template diversity, few-shot demonstrations, and the volume of instructions on performance. We make our dataset and the fine-tuned models publicly accessible at https://github.com/DaoD/INTERS.

cs.CL↗