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。
Jiawei Zhao, Zhenyu Zhang, Beidi Chen +3 authors
Gradient Low-Rank Projection (GaLore) improves memory efficiency for training large language models without sacrificing performance, enabling pre-training of 7B models on consumer GPUs.
48 篇论文 · 按点赞排序
Chenglei Si, Yanzhe Zhang, Zhengyuan Yang +2 authors
Evaluates the performance of multimodal LLMs, including GPT-4V and Gemini Pro Vision, on converting visual designs to code using a new Design2Code task and benchmarks.
Pierre Colombo, Telmo Pessoa Pires, Malik Boudiaf +8 authors
SaulLM-7B, a large language model with 7 billion parameters, excels in legal text comprehension and generation using instructional fine-tuning on a legal corpus.
Patrick Esser, Sumith Kulal, Andreas Blattmann +14 authors
Improved noise sampling techniques and a transformer-based architecture enhance rectified flow models for high-resolution text-to-image synthesis, outperforming existing methods.
Xin Men, Mingyu Xu, Qingyu Zhang +5 authors
A new layer pruning method called ShortGPT significantly improves model pruning performance by removing redundant layers based on Block Influence scores.
01. AI, Alex Young, Bei Chen +28 authors
The Yi model family, based on transformer architecture, showcases strong performance across benchmarks and modalities through optimized data and scalable infrastructure.
Alex Havrilla, Yuqing Du, Sharath Chandra Raparthy +6 authors
Multiple RLHF algorithms, including Expert Iteration, PPO, and Return-Conditioned RL, similarly improve LLM reasoning with similar sample complexity, with Expert Iteration slightly outperforming others.
Xiangxiang Chu, Jianlin Su, Bo Zhang +1 authors
A LLaMA-like vision transformer, VisionLLaMA, is introduced for processing 2D images and achieves state-of-the-art performance in various vision tasks.
Wei-Lin Chiang, Lianmin Zheng, Ying Sheng +8 authors
Chatbot Arena, an open platform, evaluates LLMs based on human preferences using pairwise comparisons and crowdsourced votes, confirming its credibility through diverse and accurate data analysis.
Junsong Chen, Chongjian Ge, Enze Xie +7 authors
PixArt-\Sigma, a Diffusion Transformer model, generates high-quality 4K images more efficiently than existing models using improved data and novel token compression techniques.
Zeqian Ju, Yuancheng Wang, Kai Shen +16 authors
NaturalSpeech 3, using factorized diffusion models, outperforms other TTS systems by generating disentangled speech attributes in a zero-shot manner.
Quentin Garrido, Mahmoud Assran, Nicolas Ballas +3 authors
Image World Models extend Joint-Embedding Predictive Architecture to predict global photometric transformations, offering adaptable and high-performance representations through fine-tuning.
Yuhao Xu, Tao Gu, Weifeng Chen +1 authors
OOTDiffusion uses latent diffusion models and a novel outfitting UNet to generate high-fidelity and controllable virtual try-on images with preserved garment details.
Zhende Song, Chenchen Wang, Jiamu Sheng +4 authors
MovieLLM generates synthetic, high-quality data for long videos using GPT-4 and text-to-image models, improving multimodal models' performance on complex narratives.
Yutao Cui, Xiaotong Zhao, Guozhen Zhang +3 authors
StableDrag, a drag-based image editing framework, enhances dragging performance through precise point tracking and confident latent enhancement, applied to both GAN and diffusion models.
Suyuchen Wang, Ivan Kobyzev, Peng Lu +2 authors
Resonance RoPE enhances LLM generalization in TSTL scenarios by improving RoPE interpolation for OOD token positions, outperforming existing methods on both synthetic and real-world tasks.
Litong Gong, Yiran Zhu, Weijie Li +4 authors
AtomoVideo generates high-fidelity videos from images using multi-granularity injection, achieving superior motion intensity and temporal stability.
Boshi Wang, Hao Fang, Jason Eisner +2 authors
A biologically inspired method called simulated trial and error (STE) improves the accuracy of tool usage in large language models through imagination, trial and error, and memory, reaching better performance than existing models like GPT-4.
Yizhe Zhang, He Bai, Ruixiang Zhang +4 authors
VLMs show deficiencies in visual deductive reasoning, particularly in multi-hop relational tasks, despite strong performance in text-based reasoning.
Shannon Zejiang Shen, Hunter Lang, Bailin Wang +2 authors
Multiple large language models collaborate by interleaving token-level generation decisions, optimizing performance across tasks without direct supervision.
Weizhi Wang, Khalil Mrini, Linjie Yang +4 authors
A new framework using fine-tuned Multimodal Language Models (MLMs) improves the precision of image-text data filtering compared to CLIPScore, enhancing model performance across different foundation models and tasks.
Wei He, Kai Han, Yehui Tang +4 authors
DenseSSM enhances state space model architectures by integrating shallow-layer hidden states into deeper layers, improving performance without increasing model size.
Chen Li, Weiqi Wang, Jingcheng Hu +5 authors
The LLaMA-2 7B model shows strong mathematical abilities through pre-training, and scaling synthetic data can enhance its performance on math benchmarks.
Yuan Gao, Kunyu Shi, Pengkai Zhu +7 authors
Strongly Supervised pre-training with ScreenShots (S4) enhances Vision-Language Models by using web screenshots with tree-structured HTML elements and spatial localization, resulting in significant improvements in diverse downstream tasks.
Zhengyang Tang, Xingxing Zhang, Benyou Wan +1 authors
MathScale generates a large dataset of mathematical reasoning questions by building concept graphs from seed questions, enhancing open-source LLMs' performance on Math Word Problems.
Buyun Zhang, Liang Luo, Yuxin Chen +12 authors
Wukong, a network architecture using stacked factorization machines with a synergistic upscaling strategy, establishes a scaling law in recommendation systems, outperforming state-of-the-art models and maintaining quality across significant increases in complexity.
Dmitry Tochilkin, David Pankratz, Zexiang Liu +7 authors
TripoSR, a 3D reconstruction model using transformer architecture, generates 3D meshes quickly from single images and outperforms alternatives.
Haogeng Liu, Quanzeng You, Xiaotian Han +7 authors
InfiMM-HD, a novel architecture with cross-attention and visual windows, enhances multimodal large language models for high-resolution image processing with efficient training and low computational overhead.
Mengqi Huang, Zhendong Mao, Mingcong Liu +2 authors
RealCustom disentangles similarity and controllability in text-to-image customization by using a decoupled train-inference framework with adaptive scoring and mask guidance strategies.
Hitesh Kandala, Jianfeng Gao, Jianwei Yang
Pix2Gif, a motion-guided diffusion model, generates image-to-GIF videos using text and motion prompts, ensuring coherence and consistency through a perceptual loss and motion-guided warping module.
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