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。
Dustin Podell, Zion English, Kyle Lacey +5 authors
SDXL, a latent diffusion model using a larger UNet with additional text encoders and attention mechanisms, improves text-to-image synthesis significantly.
44 篇论文 · 按点赞排序
Jiayu Ding, Shuming Ma, Li Dong +4 authors
LongNet, a Transformer variant with dilated attention, enables scaling to over 1 billion tokens with linear computational complexity and maintains performance on shorter sequences.
Nelson F. Liu, Kevin Lin, John Hewitt +4 authors
Analysis reveals language models perform better when relevant information is at the ends of longer contexts and suggests new evaluation protocols for long-context models.
Guocheng Qian, Jinjie Mai, Abdullah Hamdi +8 authors
A two-stage neural approach generates high-quality textured 3D meshes from a single image using neural radiance fields, differentiable mesh representations, and combined 2D and 3D diffusion priors.
Yupeng Chang, Xu Wang, Jindong Wang +13 authors
A comprehensive review of evaluation methods for large language models (LLMs) focusing on tasks, methods, and future challenges.
Frano Rajič, Lei Ke, Yu-Wing Tai +3 authors
SAM-PT extends SAM for video object segmentation using robust and sparse point propagation, achieving strong zero-shot performance and demonstrating advantages on benchmarks like DAVIS, YouTube-VOS, MOSE, and UVO.
Chong Mou, Xintao Wang, Jiechong Song +2 authors
DragonDiffusion, a novel image editing method, enables precise manipulation of generated or real images using classifier guidance and multi-scale attention mechanisms without requiring fine-tuning.
Linoy Tsaban, Apolinário Passos
LEDITS combines Edit Friendly DDPM inversion and Semantic Guidance to enable versatile and efficient real-image editing using text prompts.
Guillaume Sanchez, Honglu Fan, Alexander Spangher +3 authors
Classifier-Free Guidance enhances performance across various language modeling tasks and improves the faithfulness and coherence of AI assistants, outperforming models with higher parameter counts.
Tan Wang, Linjie Li, Kevin Lin +5 authors
DISCO addresses the challenges of human-centric dance synthesis by introducing a model architecture for better faithfulness and compositionality, and pre-training for generalizability.
Deepanway Ghosal, Yew Ken Chia, Navonil Majumder +1 authors
Fine-tuning VICUNA on a customized instruction dataset called FLANMINI enhances its problem-solving skills, leading to improved performance across various benchmarks.
Mustafa Safdari, Greg Serapio-García, Clément Crepy +6 authors
A method quantifies and manipulates personality traits in large language models' outputs, finding reliability and validity enhanced in larger, instruction-fine-tuned models and exploring ethical considerations.
Junting Pan, Keqiang Sun, Yuying Ge +10 authors
A large-scale dataset called JourneyDB with 4 million generated images and text prompts is introduced to evaluate multi-modal models' performance in understanding generated images through four benchmarks.
Jiabo Ye, Anwen Hu, Haiyang Xu +10 authors
mPLUG-DocOwl, an extension of mPLUG-Owl, enhances OCR-free document understanding through instruction tuning and achieves superior performance across various visual-text tasks.
Alexander Wei, Nika Haghtalab, Jacob Steinhardt
Research identifies failure modes in safety training of large language models, leading to the creation of effective jailbreak attacks that exploit these vulnerabilities.
Mounica Maddela, Megan Ung, Jing Xu +3 authors
Existing language models can generate tailored cognitive practice material for well-being by using a new dataset of unhelpful thoughts and reframes, requiring minimal additional training.
Yan Zeng, Hanbo Zhang, Jiani Zheng +5 authors
A comprehensive study examines design choices impacting multi-modal capabilities of Large Language Models, introducing Lynx as the most accurate and versatile multi-modal model among existing GPT4-style models.
Takuma Yoneda, Jiading Fang, Peng Li +7 authors
Statler framework enhances LLMs' ability to reason over long time horizons by integrating an explicit world state memory using two LLM instances.
Szymon Tworkowski, Konrad Staniszewski, Mikołaj Pacek +3 authors
A new method called Focused Transformer (FoT) enhances large language models by extending their context length through contrastive learning, demonstrated with LongLLaMA models that handle up to 256k context length.
Hongxin Zhang, Weihua Du, Jiaming Shan +5 authors
LLMs enable effective multi-agent cooperation and communication without fine-tuning or few-shot prompting, surpassing strong planning methods and enhancing human trust.
Yuan Gong, Sameer Khurana, Leonid Karlinsky +1 authors
Whisper-AT, a unified model built upon Whisper, extends speech recognition to audio event tagging with minimal extra computational cost.
Luciano Del Corro, Allie Del Giorno, Sahaj Agarwal +3 authors
SkipDecode, a token-level early exit method, enhances batch inferencing and KV caching in autoregressive LLMs, achieving significant speedups while maintaining performance.
Allen Z. Ren, Anushri Dixit, Alexandra Bodrova +11 authors
KnowNo is a framework using conformal prediction to measure and align the uncertainty of LLM-based planners, improving efficiency and autonomy in multi-step planning for robots.
Shitao Tang, Fuyang Zhang, Jiacheng Chen +2 authors
MVDiffusion is a multi-view image generation method that uses a correspondence-aware attention mechanism to simultaneously create high-resolution images with global consistency and fine details.
Sai Vemprala, Rogerio Bonatti, Arthur Bucker +1 authors
ChatGPT can effectively perform various robotics tasks using natural language instructions and prompt engineering, supported by an open-sourced tool for collaborative development.
Yuxiang Zhang, Hongwen Zhang, Liangxiao Hu +3 authors
A learning-based system using sequential proxy-to-motion learning achieves real-time monocular full-body capture with accurate foot-ground contact and world space representation.
Manli Shu, Jiongxiao Wang, Chen Zhu +3 authors
AutoPoison, an automated data poisoning pipeline, allows adversaries to exploit instruction tuning by injecting specific examples into training data, thereby altering the model's behavior with minimal data and high stealth.
Zhen Qin, Rolf Jagerman, Kai Hui +8 authors
A pairwise ranking prompting technique for document ranking using large language models achieves state-of-the-art performance with moderate-sized, open-sourced models compared to larger, commercial models.
Shentong Mo, Enze Xie, Ruihang Chu +4 authors
A novel Diffusion Transformer architecture for 3D shape generation, DiT-3D, using voxelized point clouds achieves superior performance compared to U-Net methods, employing 3D positional embeddings and 3D window attention to reduce computational costs.
Ariel N. Lee, Sarah Adel Bargal, Janavi Kasera +3 authors
Patch Mixing augmentation allows CNNs to simulate patch selectivity and handle occlusions better, similar to ViTs.
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