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。
Andrey Gromov, Kushal Tirumala, Hassan Shapourian +2 authors
Layer pruning of pre-trained LLMs with parameter-efficient finetuning methods shows minimal performance degradation and significant resource savings in both finetuning and inference.
45 篇论文 · 按点赞排序
Kai Mei, Zelong Li, Shuyuan Xu +3 authors
AIOS, an operating system embedding large language models, addresses resource allocation, context switching, and concurrency challenges for intelligent agents, demonstrating reliability and efficiency.
Qihang Fan, Quanzeng You, Xiaotian Han +5 authors
ViTAR enhances Vision Transformers' scalability across resolutions through dynamic token integration and fuzzy positional encoding, improving accuracy and reducing computational costs.
Yanwei Li, Yuechen Zhang, Chengyao Wang +5 authors
Mini-Gemini enhances multi-modality Vision Language Models through high-resolution visual tokens, high-quality datasets, and VLM-guided generation, achieving leading performance in zero-shot benchmarks.
Dahyun Kim, Yungi Kim, Wonho Song +4 authors
A stepwise direct preference optimization approach improves the alignment of large language models with human preferences and enhances their performance.
Zheng Cai, Maosong Cao, Haojiong Chen +97 authors
InternLM2 is an open-source LLM that outperforms predecessors through innovative pre-training and optimization techniques, including Supervised Fine-Tuning and Conditional Online Reinforcement Learning from Human Feedback.
Binbin Huang, Zehao Yu, Anpei Chen +2 authors
2D Gaussian Splatting models and reconstructs geometrically accurate radiance fields from multi-view images, improving surface accuracy and rendering efficiency.
Akshay Krishnamurthy, Keegan Harris, Dylan J. Foster +2 authors
Experimentation with LLMs in multi-armed bandit environments reveals that robust exploration requires specific prompts, external summarization of history, or algorithmic interventions.
Daniel Winter, Matan Cohen, Shlomi Fruchter +3 authors
A method using a counterfactual dataset and bootstrap supervision enhances diffusion models for realistic object removal and insertion by accurately modeling scene effects.
Yi Wang, Kunchang Li, Xinhao Li +15 authors
InternVideo2, a state-of-the-art video foundation model, employs a progressive training approach using multiple self- or weakly-supervised learning paradigms to excel in various video and audio tasks.
Jerry Wei, Chengrun Yang, Xinying Song +8 authors
SAFE, a novel method using LLMs and Google Search, evaluates long-form factual accuracy in LLM responses more cheaply and accurately than human annotators.
Nicholas Lee, Thanakul Wattanawong, Sehoon Kim +6 authors
LLM2LLM enhances LLM performance in low-data scenarios by iteratively using a teacher model to augment the training dataset with synthetic data based on the student's prediction errors.
Nikolaos Sarafianos, Tuur Stuyck, Xiaoyu Xiang +3 authors
Garment3DGen synthesizes 3D textile assets from 2D images using diffusion methods and mesh deformation, enabling simulation-ready garments from real or generated images, including textual prompts.
Omer Dahary, Or Patashnik, Kfir Aberman +1 authors
Bounded Attention addresses semantic leakage in text-to-image diffusion models by restricting information flow between subjects, improving the alignment with given prompts and layouts.
Elliot Bolton, Abhinav Venigalla, Michihiro Yasunaga +8 authors
BioMedLM, a 2.7 billion parameter GPT-style model trained on PubMed, achieves competitive performance on biomedical NLP tasks with smaller computational requirements.
Shilong Zhang, Lianghua Huang, Xi Chen +7 authors
FlashFace personalizes photos using face identity encoding into feature maps and disentangled integration of text and image guidance.
Yuda Song, Zehao Sun, Xuanwu Yin
A dual approach using model miniaturization and reduced sampling steps improves diffusion models' inference speed through knowledge distillation and feature matching.
De-An Huang, Shijia Liao, Subhashree Radhakrishnan +4 authors
LITA addresses temporal localization in multimodal LLMs by introducing time tokens, SlowFast architecture, and a new task dataset, improving video-based text generation and temporal mIoU.
Bowen Zhang, Yiji Cheng, Jiaolong Yang +5 authors
GaussianCube, a structured Gaussian Splatting representation, enables efficient and high-quality generative modeling using diffusion models and U-Net architecture.
Qiuhong Shen, Xuanyu Yi, Zike Wu +4 authors
Gamba is an end-to-end 3D reconstruction model that uses 3D Gaussian splatting and a Mamba-based sequential network for efficient single-view 3D asset generation.
Oscar Mañas, Pietro Astolfi, Melissa Hall +6 authors
The OPT2I framework leverages a large language model to iteratively refine text-to-image prompts, enhancing consistency between prompts and generated images while maintaining image quality.
Kerui Ren, Lihan Jiang, Tao Lu +4 authors
Octree-GS, a novel LOD-structured 3D Gaussian approach, enhances real-time rendering efficiency and consistency in large, complex scenes.
Sherwin Bahmani, Xian Liu, Yifan Wang +9 authors
TC4D, a trajectory-conditioned text-to-4D generation method, improves motion realism and synthesis by separating global motion into rigid transformations and learning local deformations using text-to-video supervision.
Rakuten Group, Aaron Levine, Connie Huang +27 authors
A suite of Japanese-oriented language models achieve top performance on open benchmarks, with variations for instructions and chat.
Junyuan Hong, Jinhao Duan, Chenhui Zhang +12 authors
Compression techniques, particularly quantization, have varying impacts on the trustworthiness of large language models, with quantization within moderate bit ranges enhancing certain ethical dimensions while extreme quantization degrades trustworthiness.
Yanyu Li, Xian Liu, Anil Kag +6 authors
Fine-tuning the CLIP text encoder within diffusion-based text-to-image models using the TextCraftor approach improves performance and enables controllable image generation.
Shenhao Zhu, Junming Leo Chen, Zuozhuo Dai +5 authors
A method for human image animation using a 3D human parametric model within a latent diffusion framework enhances shape alignment and motion guidance by integrating depth images, normal maps, and semantic maps with self-attention mechanisms.
Zhenwei Wang, Tengfei Wang, Gerhard Hancke +2 authors
ThemeStation generates high-quality, theme-aware 3D assets from few exemplars using a dual score distillation loss in a two-stage framework.
Yujin Chen, Yinyu Nie, Benjamin Ummenhofer +4 authors
Mesh2NeRF derives ground-truth radiance fields from 3D meshes using an occupancy function and reflection function, improving accuracy for NeRF training and generation.
Zhongwei Zhang, Fuchen Long, Yingwei Pan +4 authors
TRIP enhances image-to-video generation using diffusion models with an image noise prior and temporal residual learning to maintain coherence and alignment.
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