Explore · 每周精选
发现最受关注的研究论文,追踪研究趋势,订阅感兴趣的期刊与关键词。
50 篇论文 · 按点赞排序
Pengxiang Li, Lu Yin, Shiwei Liu
Mix-LN normalization technique improves deep layer training in LLMs by combining Pre-LN and Post-LN, enhancing gradient norms and overall performance during pre-training and fine-tuning.
Jaehyeon Kim, Taehong Moon, Keon Lee +1 authors
ResGen, an RVQ-based discrete diffusion model, achieves high-fidelity generation with faster inference speeds, outperforming autoregressive models in conditional image generation and zero-shot text-to-speech synthesis.
Yipeng Zhang, Yifan Liu, Zonghao Guo +9 authors
LLaVA-UHD v2, using a Hierarchical Window Transformer, improves performance in multimodal large language models by capturing diverse visual granularity, leading to superior results on benchmarks.
Jiale Cheng, Xiao Liu, Cunxiang Wang +7 authors
SPaR, a self-play framework using tree-search refinement, enhances instruction-following capabilities of LLMs by generating relevant preference pairs, leading to improved performance on benchmarks without sacrificing generalization.
Tiehan Fan, Kepan Nan, Rui Xie +6 authors
A novel structured caption framework improves video generation fidelity by converting videos into instance-level descriptions, leading to more precise and hallucination-free captions.
Sihui Ji, Yiyang Wang, Xi Chen +3 authors
FashionComposer generates personalized fashion images using multi-modal input, featuring a universal framework, asset library, reference UNet, and subject-binding attention.
Hanwen Liang, Junli Cao, Vidit Goel +6 authors
A novel pipeline using latents from a video diffusion model to predict 3D scenes from single images achieves high quality and efficiency, outperforming existing methods.
Hanlin Wang, Hao Ouyang, Qiuyu Wang +5 authors
A 3D drag-based interaction method that uses object masks and depth information to control trajectories in image-to-video synthesis, enhancing creativity and precision.
Haoge Deng, Ting Pan, Haiwen Diao +6 authors
A new autoregressive video generation model, NOVA, achieves high efficiency, surpassing prior models and diffusion methods in various tasks with fewer parameters.
Manan Suri, Puneet Mathur, Franck Dernoncourt +3 authors
VisDoMRAG, a multimodal Retrieval Augmented Generation model combining text and visual RAG, enhances document-grounded question answering in multimodal rich content settings by aligning reasoning processes across modalities.
Seungwook Han, Jinyeop Song, Jeff Gore +1 authors
Transformers improve in-context learning through a concept encoding-decoding mechanism, forming and utilizing internal abstractions in their representations.
Jixuan He, Wanhua Li, Ye Liu +3 authors
The Mask-Aware Dual Diffusion (MADD) model enables seamless object insertion into scenes by explicitly modeling the insertion mask in the diffusion process, addressing data limitations and generalizing well to real-world images.
Zihan Liu, Yang Chen, Mohammad Shoeybi +2 authors
In this paper, we introduce AceMath, a suite of frontier math models that excel in solving complex math problems, along with highly effective reward models capable of evaluating generated solutions and reliably identifying the correct ones. To develop the instruction-tuned math models, we propose a supervised fine-tuning (SFT) process that first achieves competitive performance across general domains, followed by targeted fine-tuning for the math domain using a carefully curated set of prompts and synthetically generated responses. The resulting model, AceMath-72B-Instruct greatly outperforms Qwen2.5-Math-72B-Instruct, GPT-4o and Claude-3.5 Sonnet. To develop math-specialized reward model, we first construct AceMath-RewardBench, a comprehensive and robust benchmark for evaluating math reward models across diverse problems and difficulty levels. After that, we present a systematic approach to build our math reward models. The resulting model, AceMath-72B-RM, consistently outperforms state-of-the-art reward models. Furthermore, when combining AceMath-72B-Instruct with AceMath-72B-RM, we achieve the highest average rm@8 score across the math reasoning benchmarks. We will release model weights, training data, and evaluation benchmarks at: https://research.nvidia.com/labs/adlr/acemath
Mark Endo, Xiaohan Wang, Serena Yeung-Levy
FEATHER, a novel acceleration technique for Vision-Language Models, improves localization performance significantly by addressing flaws in early-layer pruning and incorporating uniform sampling.
Hao Shao, Shulun Wang, Yang Zhou +7 authors
A diffusion-based framework with a VidFaceVAE and Attribute-Identity Disentanglement Triplet Dataset improves video face swapping through enhanced temporal consistency, identity preservation, and robustness to occlusions and pose variations.
Xinli Xu, Wenhang Ge, Dicong Qiu +8 authors
GaussianProperty is a training-free framework that assigns physical properties to 3D Gaussians using SAM and GPT-4V, enabling applications in physics-based dynamic simulation and robotic grasping.
Haotong Lin, Sida Peng, Jingxiao Chen +7 authors
Prompt Depth Anything integrates LiDAR data into depth foundation models for high-resolution depth estimation, achieving state-of-the-art results on ARKitScenes and ScanNet++ datasets and improving downstream applications.
Yifei Zhou, Qianlan Yang, Kaixiang Lin +5 authors
A learning system (Proposer-Agent-Evaluator) enables foundation models to autonomously discover and practice skills by proposing tasks, attempting them, and being evaluated by a success evaluator using reinforcement learning.
Zhibing Li, Tong Wu, Jing Tan +3 authors
IDArb, a diffusion-based model, accurately decomposes geometric and material information from multiple images under varying lights, improving consistency and performance compared to traditional methods with the help of ARB-Objaverse dataset.
Guillaume Astruc, Nicolas Gonthier, Clement Mallet +1 authors
Anysat uses joint embedding predictive architecture and resolution-adaptive spatial encoders to train a unified multimodal model on diverse Earth observation datasets, achieving state-of-the-art performance across various environment monitoring tasks.
北京市昌平区探索星信息技术及软件开发工作室
京ICP备2026059466号