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
Alsu Sagirova, Mikhail Burtsev
The addition of symbolic working memory to Transformer decoders improves machine translation by storing key concepts, with memory content correlating to text complexity.
Shenghai Yuan, Jinfa Huang, Yongqi Xu +7 authors
ChronoMagic-Bench evaluates text-to-video models' ability to generate time-lapse videos with metamorphic amplitude and temporal coherence using novel metrics and a large dataset.
Yiduo Guo, Jie Fu, Huishuai Zhang +2 authors
Strategies for continual pre-training of large language models reduce performance drops and computational costs during adaptation to new domains.
Euiin Yi, Taehyeon Kim, Hongseok Jeung +2 authors
A speculative decoding approach with language-specific draft models speeds up inference time for large language models in multilingual settings.
Mohammad Salama, Jonathan Kahana, Eliahu Horwitz +1 authors
A method called DSiRe is introduced for estimating the dataset size used to fine-tune models using LoRA, leveraging norms and spectra of LoRA matrices.
Jianzong Wu, Xiangtai Li, Yanhong Zeng +5 authors
MotionBooth fine-tunes a text-to-video model using a few images to animate customized subjects with precise control over object and camera movements, enhancing performance with subject region loss, video preservation loss, and subject token cross-attention loss.
Xiangyu Zhao, Xiangtai Li, Haodong Duan +4 authors
MG-LLaVA enhances multimodal language models with a multi-granularity vision flow, including high-resolution and object-centric features, to improve visual understanding and performance on multimodal tasks.
Chao Lou, Zixia Jia, Zilong Zheng +1 authors
SPARSEK Attention addresses inefficiencies in autoregressive Transformers by providing linear time complexity and constant memory footprint, enabling efficient processing of long sequences and integration into large pre-trained models.
Xuan He, Dongfu Jiang, Ge Zhang +16 authors
MantisScore, trained on the VideoFeedback dataset, provides highly accurate automatic video quality assessment, outperforming existing metrics and serving as a reliable human-rater proxy.
Ka Man Lo, Zeyu Huang, Zihan Qiu +2 authors
A study on the inner workings of MoE-based large language models reveals insights into expert behavior and router selection, offering practical suggestions for practitioners.
Zackary Rackauckas, Arthur Câmara, Jakub Zavrel
A new evaluation framework using LLMs and an Elo-based system addresses challenges in automating the evaluation of RAG-based QA systems, showing RAG-Fusion outperforms RAG in completeness but not in precision.
Dachun Kai, Jiayao Lu, Yueyi Zhang +1 authors
A new video super-resolution method uses event signals to enhance texture recovery, achieving state-of-the-art results on several datasets.
Matthew Zheng, Enis Simsar, Hidir Yesiltepe +3 authors
A comprehensive dataset and series of tasks for text-to-image models explore diverse user-generated styles and enhance personalization methods, supporting unique artistic expressions.
Tianyu Fu, Haofeng Huang, Xuefei Ning +10 authors
Mixture of Attention (MoA) adapts sparse attention configurations for different heads and layers, enhancing context length, retrieval accuracy, and throughput in Large Language Models.
Minzheng Wang, Longze Chen, Cheng Fu +11 authors
A new long-context benchmark, Loong, evaluates large language models using extended multi-document question answering tasks that mimic real-world scenarios.
Markus Frohmann, Igor Sterner, Ivan Vulić +2 authors
A new model called Segment Any Text (SaT) is introduced for robust and efficient text segmentation across various domains, outperforming existing methods, especially in poorly formatted text.
Seungju Han, Kavel Rao, Allyson Ettinger +5 authors
WildGuard is an open, lightweight moderation tool for LLM safety that identifies malicious intent, detects safety risks, and assesses refusal rates, achieving state-of-the-art performance across multiple benchmarks.
Chia-Yu Hung, Navonil Majumder, Ambuj Mehrish +1 authors
An evolutionary approach decouples exploration and exploitation in decoding-time alignment to enhance LLM preferences without compromising performance.
Yeonsu Kwon, Jiho Kim, Gyubok Lee +6 authors
A new dataset and framework, EHRCon and CheckEHR, ensure consistency between structured and unstructured data in EHRs using manual annotations and large language models.
Xinrong Zhang, Yingfa Chen, Shengding Hu +6 authors
Duplex models enable real-time, simultaneous interactions for LLMs through time-division-multiplexing, improving user satisfaction and naturalness in AI conversations.
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