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。
Binyuan Hui, Jian Yang, Zeyu Cui +14 authors
Qwen2.5-Coder series demonstrates state-of-the-art code generation, completion, reasoning, and repair capabilities using the Qwen2.5 architecture with over 5.5 trillion tokens of training data.
50 篇论文 · 按点赞排序
Aviral Kumar, Vincent Zhuang, Rishabh Agarwal +15 authors
The multi-turn online reinforcement learning approach SCoRe improves LLM self-correction using self-generated data and regularization, achieving state-of-the-art performance.
Shitao Xiao, Yueze Wang, Junjie Zhou +6 authors
OmniGen is a unified diffusion model for image generation that supports diverse tasks without additional modules, emphasizing simplicity, knowledge transfer, and reasoning capabilities.
Peng Wang, Shuai Bai, Sinan Tan +16 authors
The Qwen2-VL Series uses Naive Dynamic Resolution and Multimodal Rotary Position Embedding to enhance visual processing and achieves competitive performance on multimodal benchmarks.
Wenliang Dai, Nayeon Lee, Boxin Wang +7 authors
NVML 1.0, a family of multimodal large language models, achieves state-of-the-art results in vision-language tasks by combining text-only and multimodal training, utilizing a new architecture and dataset strategy.
Ye Bai, Haonan Chen, Jitong Chen +35 authors
Seed-Music generates high-quality music with style control using a unified framework combining auto-regressive language modeling and diffusion approaches, supporting vocal music generation and post-production editing.
Xiaotian Han, Yiren Jian, Xuefeng Hu +8 authors
InfiMM-WebMath-40B, a large multimodal dataset, significantly improves mathematical reasoning capabilities in LLMs, achieving state-of-the-art performance on multimodal benchmarks.
Yi Lu, Jing Nathan Yan, Songlin Yang +6 authors
A controlled evaluation protocol for long-context models reveals that exact fine-tuning methods are effective, while approximate attention methods underperform, and extrapolation remains challenging.
Xingyi Yang, Xinchao Wang
The Kolmogorov-Arnold Transformer replaces MLP layers with Kolmogorov-Arnold Network layers to enhance transformers, overcoming challenges related to inference speed, computation efficiency, and weight initialization through rational basis, group learning, and variance-preserving techniques.
Di Liu, Meng Chen, Baotong Lu +11 authors
RetrievalAttention is a training-free method that accelerates attention computation in Transformer-based LLMs by using attention-aware vector search to reduce retrieval costs and GPU memory usage.
Zayne Sprague, Fangcong Yin, Juan Diego Rodriguez +7 authors
Chain-of-thought prompting enhances large language model performance primarily in math or logic tasks by improving symbolic reasoning, but it is less beneficial in other domains and can be selectively applied to save inference costs.
Dongzhi Jiang, Renrui Zhang, Ziyu Guo +10 authors
MMSearch-Engine enables multimodal search capabilities in Large Multimodal Models (LMMs) using a comprehensive benchmark and demonstrates improved performance over commercial products.
Jiongnan Liu, Yutao Zhu, Shuting Wang +6 authors
A novel model constructs user-specific embeddings using a lightweight plug-in module to personalize LLM outputs without fine-tuning, improving performance on various tasks.
Saba Sturua, Isabelle Mohr, Mohammad Kalim Akram +9 authors
jina-embeddings-v3, a large-scale text embedding model, achieves state-of-the-art performance in multilingual and long-context retrieval tasks using Low-Rank Adaptation and Matryoshka Representation Learning.
Joonghyuk Shin, Daehyeon Choi, Jaesik Park
InstantDrag, an optimization-free pipeline with two networks, performs fast, photo-realistic drag-based image editing without requiring masks or text prompts.
Gonzalo Martin Garcia, Karim Abou Zeid, Christian Schmidt +3 authors
Efficient depth estimation is achieved by optimizing the inference pipeline of diffusion models and fine-tuning them for specific tasks, outperforming existing methods.
Mouxiang Chen, Zhongxin Liu, He Tao +4 authors
The research defines an optimal code selection strategy using Bayesian posterior probability within an integer programming framework and proposes an efficient approximation to outperform heuristics in selecting code solutions with LLM-generated tests.
Zhenwei Wang, Tengfei Wang, Zexin He +3 authors
Phidias, a diffusion-based generative model, uses a 3D reference to improve 3D generation quality, generalization, and controllability through meta-ControlNet, dynamic reference routing, and self-reference augmentations.
Zuyan Liu, Yuhao Dong, Ziwei Liu +3 authors
Oryx, a unified multimodal architecture, uses a pre-trained OryxViT model and a dynamic compressor to efficiently process diverse visual inputs of varying resolutions and lengths, enhancing spatial-temporal understanding and multimodal capabilities.
Zhitong Huang, Mohan Zhang, Jing Liao
A video diffusion framework for lineart colorization uses Sketch-guided ControlNet, Reference Attention, and sequential sampling with Overlapped Blending Module and Prev-Reference Attention to produce temporally consistent, high-quality animation videos from lineart drawings.
Emanuele Vivoli, Andrey Barsky, Mohamed Ali Souibgui +3 authors
A comprehensive survey addresses the challenges of comics understanding, introducing a new taxonomy and reviewing methods within the context of vision-language models.
Orion Weller, Benjamin Van Durme, Dawn Lawrie +3 authors
Promptriever, a retrieval model trained to respond to prompts, achieves state-of-the-art performance on instruction-following tasks and demonstrates improved robustness and hyperparameter optimization capacity.
Mohammad Samragh, Iman Mirzadeh, Keivan Alizadeh Vahid +5 authors
HyperCloning expands smaller pre-trained language models to larger ones, retaining their accuracy and reducing training time.
Zhaoxi Chen, Jiaxiang Tang, Yuhao Dong +10 authors
3DTopia-XL, a scalable 3D generative model using a new primitive-based representation and Diffusion Transformer framework, generates high-quality 3D assets with PBR materials and textures from textual or visual inputs.
Genta Indra Winata, Hanyang Zhao, Anirban Das +4 authors
Preference tuning in deep generative models integrates human feedback across modalities using reinforcement learning to align model outputs with human preferences, with a focus on methodologies and future directions.
Jie Zhou, Chufeng Xiao, Miu-Ling Lam +1 authors
A new system, DrawingSpinUp, reconstructs 3D models from 2D character drawings and applies motion data to generate high-quality 3D animations.
Jiarui Hai, Yong Xu, Hao Zhang +4 authors
EzAudio, a transformer-based diffusion model operating in the latent space of a 1D waveform VAE, enhances text-to-audio generation by improving quality, efficiency, and data utilization through various optimizations.
Liyuan Liu, Young Jin Kim, Shuohang Wang +14 authors
GRIN enhances MoE models by using sparse gradient estimation for expert routing, enabling better scaling and performance compared to dense models.
Zhengguang Zhou, Jing Li, Huaxia Li +2 authors
StoryMaker enhances personalized image generation by maintaining holistic consistency of multiple characters in scenes, using a Positional-aware Perceiver Resampler and segment-specific cross-attention constraints.
Jemin Lee, Sihyeong Park, Jinse Kwon +2 authors
Various quantization methods are evaluated on large instruction-tuned LLMs across multiple benchmarks, revealing that quantized larger models often perform better than smaller FP16 models, except in some specific 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