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。
Dongsheng Wang, Natraj Raman, Mathieu Sibue +6 authors
DocLLM extends large language models for visual document understanding by focusing on bounding box information and disentangled matrices for multimodal attention, achieving state-of-the-art performance on document intelligence tasks.
50 篇论文 · 按点赞排序
Peiyuan Zhang, Guangtao Zeng, Tianduo Wang +1 authors
TinyLlama, a compact 1.1B language model, leverages FlashAttention to achieve high performance in downstream tasks with enhanced computational efficiency.
Liang Wang, Nan Yang, Xiaolong Huang +3 authors
A method using synthetic data and contrastive loss achieves high-quality text embeddings without labeled data, and sets state-of-the-art results when combined with labeled data.
Zixiang Chen, Yihe Deng, Huizhuo Yuan +2 authors
A novel self-play fine-tuning method (SPIN) enhances Large Language Models by generating its own training data and refining its policy, achieving performance on par with models trained using direct preference optimization and additional human data.
Yiheng Liu, Hao He, Tianle Han +18 authors
The paper reviews techniques for cost-efficient training and deployment of large language models, covering aspects like data preprocessing, parallel training, model fine-tuning, and inference optimizations including model compression and memory scheduling.
Chengyue Wu, Yukang Gan, Yixiao Ge +5 authors
A new post-pretraining method using expanded Transformer blocks for Large Language Models improves knowledge without catastrophic forgetting, yielding LLaMA Pro-8.3B that excels in general tasks, programming, and mathematics.
Jun Zhao, Zhihao Zhang, Qi Zhang +2 authors
A comprehensive study on transferring English-dominant LLMs to non-English languages using LLaMA demonstrates that comparable performance can be achieved with minimal pretraining data across various language capabilities and response quality metrics.
Rachit Bansal, Bidisha Samanta, Siddharth Dalmia +6 authors
CALM, a composition method for large language models, enhances capabilities by integrating cross-attention with specialized models, demonstrating improvements on translation and coding tasks.
Ming Yan, Ruihao Li, Hao Zhang +3 authors
LARP is a language agent framework for open-world games, incorporating memory and decision-making capabilities to enhance interaction and coherence in a diverse range of applications.
Zipeng Fu, Tony Z. Zhao, Chelsea Finn
Co-training of supervised behavior cloning with static and mobile manipulation datasets improves the success rates of mobile manipulation tasks using a whole-body teleoperation system.
Hexiang Hu, Kelvin C. K. Chan, Yu-Chuan Su +9 authors
Instruct-Imagen achieves robust and generalized image generation through fine-tuning a pre-trained text-to-image diffusion model with multi-modal instructions and retrieval-augmented training.
Suraj Patil, William Berman, Robin Rombach +1 authors
aMUSEd, a masked image model with 10% of MUSE's parameters, generates images quickly and efficiently, requiring fewer inference steps and offering easier fine-tuning compared to latent diffusion.
Nikhil Sardana, Jonathan Frankle
The modified Chinchilla scaling laws incorporate inference cost, suggesting that models for large inference demand should be smaller and trained longer than originally proposed.
Evonne Ng, Javier Romero, Timur Bagautdinov +4 authors
A framework combines vector quantization and diffusion models to generate diverse, dynamic, and photorealistic avatars responding to speech in a conversational setting using a novel multi-view dataset.
Hongye Jin, Xiaotian Han, Jingfeng Yang +5 authors
Bi-level attention mechanism extends the context window of LLMs without additional training or fine-tuning.
Boyuan Zheng, Boyu Gou, Jihyung Kil +2 authors
LMMs like GPT-4V demonstrate potential as generalist web agents by following natural language instructions to complete tasks on live websites, outperforming text-only LLMs and smaller models with a combination of HTML text and visuals for grounding, though challenges remain.
Terry Yue Zhuo, Armel Zebaze, Nitchakarn Suppattarachai +4 authors
Astraios, a suite of instruction-tuned models, investigates the cost-performance trade-offs of full-parameter fine-tuning and parameter-efficient fine-tuning across different model sizes and tasks, finding that LoRA offers the best trade-off and FFT generally performs best in downstream tasks.
Fuchen Long, Zhaofan Qiu, Ting Yao +1 authors
VideoDrafter generates high-quality, content-consistent multi-scene videos using Large Language Models and diffusion models, enhancing visual quality and user preference by leveraging comprehensive scripts and reference images.
Jiraphon Yenphraphai, Xichen Pan, Sainan Liu +2 authors
Image Sculpting integrates 3D geometry and graphics for precise 2D image editing, offering high-fidelity results through a conversion to 3D and re-rendering process.
Ningyu Zhang, Yunzhi Yao, Bozhong Tian +19 authors
Knowledge editing techniques for large language models are reviewed, categorized, and benchmarked to improve efficiency and performance with specific domain modifications.
Feng Liang, Bichen Wu, Jialiang Wang +8 authors
A proposed consistent V2V synthesis framework uses diffusion models combined with encoded optical flow to propagate edits across video frames, offering flexibility, efficiency, and high-quality results.
Yichen Zhu, Minjie Zhu, Ning Liu +3 authors
A compact multi-modal assistant, LLaVA-Phi, using a smaller language model, Phi-2, achieves high performance in visual and multi-modal dialog tasks with resource efficiency.
Mia Gaia Polansky, Charles Herrmann, Junhwa Hur +3 authors
A differentiable model using boundary attention accurately detects weak and noisy boundaries in images, offering scalable, adaptive, and sub-pixel precise performance.
Alex Jinpeng Wang, Linjie Li, Kevin Qinghong Lin +5 authors
A contrastive loss and unified multimodal framework improve vision-language pre-training models, particularly in tasks involving long textual contexts and extended video data, outperforming existing models and utilizing a novel interleaved video-text dataset.
David Junhao Zhang, Dongxu Li, Hung Le +3 authors
Moonshot, a new video generation model, uses multimodal conditioning with images and text, along with an optional pre-trained ControlNet for geometry, to improve video quality and temporal consistency.
Yonglong Tian, Lijie Fan, Kaifeng Chen +3 authors
SynCLR learns visual representations from synthetic images and captions using contrastive learning, outperforming other methods in image classification and semantic segmentation tasks.
Yunhe Wang, Hanting Chen, Yehui Tang +17 authors
The paper proposes a new efficient architecture for large language models by incorporating informed activation functions and augmented shortcuts to enhance nonlinearity, resulting in improved performance and speed compared to state-of-the-art models.
Yuqing Wang, Yun Zhao
Gemini, a multimodal large language model, demonstrates competitive commonsense reasoning abilities across a variety of datasets, highlighting the need for ongoing advancements in this area.
Wan-Duo Kurt Ma, J. P. Lewis, W. Bastiaan Kleijn
TrailBlazer enhances text-to-video generation controllability using bounding boxes for subject guidance without requiring neural training, fine-tuning, or pre-existing videos.
Alex Trevithick, Matthew Chan, Towaki Takikawa +5 authors
The paper proposes techniques to enhance neural volume rendering in 3D GANs for high-resolution 3D geometry generation and multiview consistency without post-processing.
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