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。
Xinlong Wang, Xiaosong Zhang, Zhengxiong Luo +22 authors
Emu3, a transformer-based multimodal model trained exclusively with next-token prediction, outperforms existing diffusion and compositional models in generation and perception tasks.
50 篇论文 · 按点赞排序
Zhengfeng Lai, Vasileios Saveris, Chen Chen +9 authors
A study examines the use of synthetic and AltText captions in multimodal models, finding that a hybrid approach using both types improves alignment and performance across various models.
Ming Zhong, Aston Zhang, Xuewei Wang +14 authors
LLMs show suboptimal performance in cross-capability tasks, constrained by their weakest individual capabilities, as demonstrated by the CrossEval benchmark.
Haotian Zhang, Mingfei Gao, Zhe Gan +20 authors
MM1.5 enhances multimodal language models with a data-centric approach, supporting diverse tasks like text-rich image understanding and video comprehension through careful data curation and training strategies.
Zekun Wang, King Zhu, Chunpu Xu +14 authors
MIO, a novel foundation model, achieves competitive performance in multimodal tasks through an end-to-end, autoregressive approach using causal multimodal modeling.
Jintao Zhang, Jia wei, Pengle Zhang +2 authors
A new quantization method for attention mechanisms, named SageAttention, enhances inference speed and accuracy without significant loss in performance across various models.
Aleksei Bochkovskii, Amaël Delaunoy, Hugo Germain +4 authors
Depth Pro is a fast and accurate zero-shot monocular depth estimation model that generates high-resolution depth maps using a multi-scale transformer and a combined real-synthetic training protocol.
Yuanhan Zhang, Jinming Wu, Wei Li +4 authors
A synthetic dataset, LLaVA-Video-178K, and a new video large multimodal model, LLaVA-Video, achieve strong performance in video instruction-following tasks.
Bin Wang, Chao Xu, Xiaomeng Zhao +15 authors
MinerU is an open-source tool that enhances document content extraction using fine-tuned models and pre/postprocessing rules across diverse document types.
Yuling Shi, Songsong Wang, Chengcheng Wan +1 authors
MGDebugger hierarchical debugger resolves code bugs at multiple granularities, improving debugging accuracy and success rates over existing systems.
Tianyi Xiong, Xiyao Wang, Dong Guo +5 authors
LLaVA-Critic, an open-source large multimodal model, effectively evaluates multimodal tasks and provides reliable scores, surpassing GPT models, and enhances preference learning for model alignment.
Hong-You Chen, Zhengfeng Lai, Haotian Zhang +7 authors
CLOC enhances CLIP's localization capabilities by introducing region-text contrastive loss and promptable embeddings, improving regional image representation for multimodal large language models.
Yuqing Wang, Tianwei Xiong, Daquan Zhou +5 authors
A novel autoregressive large language model (LLM)-based video generator, Loong, is proposed to produce minute-long videos by addressing challenges related to long video generation.
Dongwei Jiang, Guoxuan Wang, Yining Lu +5 authors
RATIONALYST, a process-supervised reasoning model pre-trained on vast web-scale rationale annotations, enhances reasoning accuracy across diverse tasks compared to larger models.
Seyedmorteza Sadat, Otmar Hilliges, Romann M. Weber
Adaptive projected guidance modifies classifier-free guidance to enhance image quality and reduce oversaturation by decomposing and rescaling update components, enabling higher guidance scales without artifacts.
Mike Ranzinger, Jon Barker, Greg Heinrich +3 authors
Heterogeneous multi-teacher knowledge distillation improves visual foundation models by optimizing activation statistics and alignment, with PHI Standardization shown to achieve the best student model quality.
Zonghang Li, Wenjiao Feng, Mohsen Guizani +1 authors
A tensor parallel inference system named TPI-LLM improves large language model (LLM) inference on edge devices by minimizing memory usage and reducing latency.
Oussama Zekri, Ambroise Odonnat, Abdelhakim Benechehab +3 authors
Theoretical analysis of large language models explores their inference capabilities and generalization through equivalence with Markov chains, providing pre-training and in-context bounds.
Yang Cao, Yuanliang Jv, Dan Xu
3D Gaussian Splatting is introduced for 3D object detection, improving NeRF's limitations through 2D Boundary Guidance and Box-Focused Sampling strategies, achieving superior performance on ScanNet and ARKITScenes datasets.
Siheng Li, Cheng Yang, Taiqiang Wu +12 authors
Honesty is a fundamental principle for aligning large language models (LLMs) with human values, requiring these models to recognize what they know and don't know and be able to faithfully express their knowledge. Despite promising, current LLMs still exhibit significant dishonest behaviors, such as confidently presenting wrong answers or failing to express what they know. In addition, research on the honesty of LLMs also faces challenges, including varying definitions of honesty, difficulties in distinguishing between known and unknown knowledge, and a lack of comprehensive understanding of related research. To address these issues, we provide a survey on the honesty of LLMs, covering its clarification, evaluation approaches, and strategies for improvement. Moreover, we offer insights for future research, aiming to inspire further exploration in this important area.
Arian Hosseini, Alessandro Sordoni, Daniel Toyama +2 authors
LLMs display a significant reasoning gap when solving dependent math problems, which is more pronounced in smaller and math-specialized models, and is attributed to distraction from additional context and poor second-hop reasoning.
Mengzhao Jia, Wenhao Yu, Kaixin Ma +6 authors
A model designed for multimodal tasks involving multiple text-rich images addresses dataset scarcity and visual feature balancing, outperforming existing models in specialized and general evaluations.
Defa Zhu, Hongzhi Huang, Zihao Huang +5 authors
Hyper-connections, an alternative to residual connections, improve performance in pre-training large language models and vision tasks by dynamically adjusting feature connections.
Shaowei Liu, Zhongzheng Ren, Saurabh Gupta +1 authors
PhysGen generates realistic, physically plausible videos from a single image and input conditions by integrating physical simulation with video diffusion.
Yifei Liu, Jicheng Wen, Yang Wang +5 authors
Vector Post-Training Quantization (VPTQ) uses Second-Order Optimization to achieve extremely low-bit quantization of Large Language Models (LLMs), improving accuracy and inference throughput.
Jiaming Li, Lei Zhang, Yunshui Li +5 authors
A novel approach, Ruler, uses Meta Length Tokens to improve large language models' adherence to specified response lengths in generated text.
Zhilin Wang, Alexander Bukharin, Olivier Delalleau +5 authors
A novel combined Bradley-Terry and Regression reward modeling approach outperforms other models on RewardBench and aligns models to follow instructions in RLHF.
Guokan Shang, Hadi Abdine, Yousef Khoubrane +9 authors
Atlas-Chat, a collection of large language models fine-tuned for Moroccan Arabic (Darija), outperforms state-of-the-art LLMs on Darija-specific tasks.
William Held, Ella Li, Michael Ryan +3 authors
A self-supervised training approach using text-only LLM responses improves Speech LLMs for various tasks with reduced computational resources and enhanced user satisfaction.
John X. Morris, Alexander M. Rush
Contextualized document embeddings that consider neighboring documents improve retrieval performance over traditional biencoders, achieving state-of-the-art results on the MTEB benchmark.
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