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
Yao Shu, Wenyang Hu, See-Kiong Ng +2 authors
Ferret is a federated full-parameter tuning method for LLMs that achieves high efficiency, reduced communication overhead, and competitive accuracy using first-order methods and shared randomness.
Ayush Chopra, Shashank Kumar, Nurullah Giray-Kuru +2 authors
AgentTorch, a framework utilizing large language models as agents, scales agent-based modeling to millions of agents, enabling high-resolution simulations of complex systems like the COVID-19 pandemic.
Xiaofeng Mao, Zhengkai Jiang, Fu-Yun Wang +5 authors
A two-stage training framework combining consistency distillation and GAN training improves video diffusion model efficiency and performance.
DaDong Jiang, Xianghui Yang, Zibo Zhao +7 authors
FlexiTex generates high-quality textures using visual guidance to enhance text prompts and maintain consistency.
Yifan Zhang, Yang Yuan, Andrew Chi-Chih Yao
The Diagram of Thought (DoT) framework models iterative reasoning in LLMs as a DAG, leveraging auto-regressive prediction and Topos Theory to enhance reasoning consistency and efficiency.
Haoxuan Wang, Qingdong He, Jinlong Peng +3 authors
Mamba-YOLO-World improves open-vocabulary detection by introducing a linear complexity MambaFusion-PAN neck architecture, enhancing performance on COCO and LVIS benchmarks.
EverestAI, Sijin Chen, Yuan Feng +16 authors
Takin AudioLLM, a series of models including Takin TTS, Takin VC, and Takin Morphing, facilitates zero-shot speech production with high-quality and customizable outputs.
Yuxiang Yang, Guanya Shi, Changyi Lin +8 authors
A hierarchical learning and control framework enables a quadrupedal robot to perform agile, continuous, and terrain-adaptive jumping in complex environments using a combination of terrain perception, motion planning, and low-level control.
Sreyan Ghosh, Sonal Kumar, Chandra Kiran Reddy Evuru +3 authors
A method employing ReCLAP and prompt augmentation enhances zero-shot audio classification by using descriptive audio features in context-specific prompts.
Yohan Poirier-Ginter, Alban Gauthier, Julien Phillip +2 authors
The method creates relightable 3D radiance fields using single-illumination data by leveraging 2D diffusion model priors, achieving realistic relighting in full scenes.
Jiaxin Wen, Ruiqi Zhong, Akbir Khan +6 authors
Reinforcement Learning from Human Feedback (RLHF) enhances language models' ability to deceive human evaluators without improving task accuracy, leading to increased evaluation errors and highlighting shortcomings in current sophistry detection methods.
Yuheng Jiang, Zhehao Shen, Yu Hong +5 authors
A Gaussian-based method, DualGS, enables efficient real-time and high-fidelity volumetric video playback with low storage requirements, facilitating immersive VR experiences.
Helin Wang, Jiarui Hai, Yen-Ju Lu +3 authors
SoloAudio, a diffusion-based generative model using a Transformer architecture, achieves state-of-the-art results in target sound extraction, supporting both audio-oriented and language-oriented tasks with strong generalization.
Qiuhong Shen, Xingyi Yang, Michael Bi Mi +1 authors
Vista3D framework generates 3D models from 2D images using Gaussian Splatting and disentangled implicit functions, combining 2D and 3D diffusion priors for improved quality and consistency.
Kai Li, Yi Luo
Apollo, a generative model for high-sample-rate audio restoration, uses a frequency band split module to enhance restored audio quality while maintaining efficiency, outperforming existing SR-GAN models on MUSDB18-HQ and MoisesDB datasets.
Marko Mihajlovic, Sergey Prokudin, Siyu Tang +4 authors
The proposed optimization strategy enhances 3D Gaussian Splatting performance in sparse reconstruction settings by modeling splat features using implicit neural fields.
Abdullatif Köksal, Marion Thaler, Ayyoob Imani +3 authors
MURI generates instruction tuning datasets for low-resource languages without human annotators, improving model performance in NLU and open-ended generation.
Wei Shen, Chuheng Zhang
Policy Filtration for Proximal Policy Optimization (PF-PPO) improves RLHF by filtering unreliable reward samples, enhancing performance in code generation tasks compared to state-of-the-art methods.
Neelabh Sinha, Vinija Jain, Aman Chadha
A framework for evaluating Vision-Language Models in VQA tasks using a novel dataset and GoEval metric, highlighting performance variation across models and guiding selection based on task and resource constraints.
Manjie Xu, Chenxing Li, Xinyi Tu +4 authors
Diffusion-based Audio Captioning achieves state-of-the-art performance in caption quality and surpasses existing models in generation speed and diversity for audio captioning 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