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。
Dahyun Kim, Chanjun Park, Sanghoon Kim +15 authors
A novel technique, depth up-scaling (DUS), efficiently enhances large language models (LLMs) without complex changes, building SOLAR 10.7B that outperforms existing open-source models in various NLP tasks, including instruction-following.
47 篇论文 · 按点赞排序
Zhaojian Yu, Xin Zhang, Ning Shang +5 authors
A Generator-Discriminator framework using LLMs generates high-quality instruction data, enhancing the fine-tuning of code-related tasks and model generalization.
Sondos Mahmoud Bsharat, Aidar Myrzakhan, Zhiqiang Shen
Guiding principles streamline querying and prompting of large language models, enhancing understanding and effectiveness across different scales.
Zhengqing Yuan, Zhaoxu Li, Lichao Sun
TinyGPT-V, built on Phi-2 with vision modules from BLIP-2 or CLIP, offers high performance with low computational requirements, enabling efficient multimodal large language modeling.
Jiasen Lu, Christopher Clark, Sangho Lee +5 authors
A unified autoregressive multimodal model handles image, text, audio, and action by tokenizing inputs into a shared space and processing them with a transformer architecture, achieving state-of-the-art performance across various benchmarks.
Zengzhi Wang, Rui Xia, Pengfei Liu
A new large-scale, high-quality math-centric corpus named MathPile has been introduced to enhance mathematical reasoning abilities of language models.
Jianqiang Ren, Chao He, Lin Liu +10 authors
The Make-A-Character (Mach) framework uses large language and vision models to create realistic, 3D avatars from text descriptions efficiently, enabling customization and seamless integration with existing CG pipelines.
Xiangxiang Chu, Limeng Qiao, Xinyang Lin +8 authors
MobileVLM, a small-scale multimodal vision language model, achieves competitive performance and superior inference speed on mobile devices compared to larger models.
Zhe Chen, Jiannan Wu, Wenhai Wang +12 authors
A large-scale vision-language foundation model with 6 billion parameters achieves state-of-the-art performance on various visual and vision-language tasks and can integrate with LLMs to develop multi-modal dialogue systems.
Dantong Niu, Xudong Wang, Xinyang Han +3 authors
A novel unsupervised model, U2Seg, achieves high performance in instance, semantic, and panoptic segmentation through self-supervised learning and clustering.
Jiannan Wu, Yi Jiang, Bin Yan +3 authors
UniRef++ unifies four reference-based object segmentation tasks using a single architecture with a UniFusion module and a unified Transformer, achieving state-of-the-art performance on RIS and RVOS and competitive results on FSS and VOS.
Jiawei Ren, Liang Pan, Jiaxiang Tang +4 authors
DreamGaussian4D efficiently generates 4D content with reduced optimization time, controllable motion, and high-quality animated meshes using Gaussian Splatting representation.
Lu Ling, Yichen Sheng, Zhi Tu +17 authors
A large-scale dataset, DL3DV-10K, featuring diverse real-world 3D scenes enhances benchmarks for novel view synthesis and neural radiance field learning.
Zhangyang Qi, Ye Fang, Mengchen Zhang +6 authors
A comparative study evaluates Google's Gemini and OpenAI's GPT-4V across vision-language capabilities, human interaction, temporal understanding, and intelligence/emotional quotient, revealing their unique strengths and exploring benefits of their combination.
Weiwen Xu, Deng Cai, Zhisong Zhang +2 authors
A novel framework, Contrastive Unlikelihood Training (CUT), effectively aligns large language models using language feedback, outperforming existing methods with less data and iterative improvements.
Jitesh Jain, Jianwei Yang, Humphrey Shi
VCoder, a versatile vision encoder, enhances multimodal LLMs' perception and reasoning abilities, particularly in object identification and counting, using specialized perception modalities and a newly created dataset.
Kaiwen Song, Juyong Zhang
City-on-Web achieves real-time rendering of large-scale 3D scenes on the web through partitioning and level-of-detail management, ensuring high fidelity and efficient resource usage.
Minghan Qin, Wanhua Li, Jiawei Zhou +2 authors
LangSplat constructs a 3D language field using a splatting technique with hierarchical semantics, enabling precise and efficient open-vocabulary querying in 3D spaces while significantly outperforming existing methods.
Apoorv Vyas, Bowen Shi, Matthew Le +21 authors
Audiobox, a flow-matching based unified model, enhances controllability and efficiency in generating various audio modalities including speech and sound using description-based and example-based prompting.
Filippos Christianos, Georgios Papoudakis, Matthieu Zimmer +13 authors
A framework integrates structured reasoning into AI agents' policies using intrinsic and extrinsic functions, enhancing adaptability and performance over diverse tasks by combining prior knowledge with modular learning.
Xun Guo, Mingwu Zheng, Liang Hou +8 authors
I2V-Adapter enables the conversion of images to videos by integrating a lightweight adapter into existing T2I models, preserving motion modules and reducing parameter requirements.
Xiang Wang, Shiwei Zhang, Hangjie Yuan +6 authors
A novel text-to-video generation framework TF-T2V uses text-free videos to improve performance and generalizability, showing scalability and further gain with text labels.
Yin Luo, Qingchao Kong, Nan Xu +50 authors
YAYI 2, a multilingual large language model with 30 billion parameters, achieves superior performance in Chinese contexts after pre-training and fine-tuning compared to other open-source models.
Kellin Pelrine, Mohammad Taufeeque, Michał Zając +2 authors
Exploration of gray-box threats on GPT-4 API functionalities, including fine-tuning, function calling, and knowledge retrieval, reveals new vulnerabilities that can enable harmful outputs and unauthorized actions.
Mingwei Li, Jiachen Tao, Zongxin Yang +1 authors
Human101 is a framework that reconstructs high-fidelity dynamic 3D humans from single-view videos in real-time using 3D Gaussian Splatting and Human-centric Forward Gaussian Animation, achieving high rendering speeds and quality.
Sangmin Woo, Byeongjun Park, Hyojun Go +2 authors
HarmonyView, a diffusion sampling technique, addresses consistency and diversity in single-image 3D generation and introduces a CLIP-based evaluation metric for diversity.
Brian Nlong Zhao, Yuhang Xiao, Jiashu Xu +6 authors
A pretrained Text-to-Image diffusion model learns soft prompts to generate diverse images with specific attributes, adapting to text-to-3D and validated through quantitative and qualitative evaluations.
Yiqi Lin, Conghui He, Alex Jinpeng Wang +3 authors
CLIP models exhibit a text spotting bias in vision-language applications due to excessive reliance on visual text within images in datasets like LAION-2B.
Jiawei Ren, Mingyuan Zhang, Cunjun Yu +3 authors
InsActor, a generative framework using diffusion-based human motion models, produces high-quality physics-based animations guided by human instructions through diffusion policies and compact latent skill sequences.
Zhan Li, Zhang Chen, Zhong Li +1 authors
Spacetime Gaussian Feature Splatting enables high-resolution, real-time novel view synthesis of dynamic scenes with compact storage.
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