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
Mingyu Jin, Qinkai Yu, Dong shu +5 authors
Increasing the length of reasoning steps in Chain of Thought prompts improves large language models' reasoning abilities, even with incorrect rationales, and this benefit is task-dependent.
Dennis Ulmer, Elman Mansimov, Kaixiang Lin +3 authors
LLMs generate high-quality dialogue training data through self-simulated conversations, improving fine-tuning results and evaluated by automated metrics.
Yatong Bai, Utsav Garg, Apaar Shanker +10 authors
The Let's Go Shopping (LGS) dataset, containing 15 million image-caption pairs from clean, e-commerce websites, improves image classification and captioning performance, especially for vision-language tasks.
Paweł Budzianowski, Taras Sereda, Tomasz Cichy +1 authors
The Pheme model series achieves compact and high-quality voice generation with parallel processing, efficient training on smaller datasets, and improved voice quality through distillation.
Bin Lin, Tao Peng, Chen Zhang +10 authors
DistKV-LLM, a distributed LLM serving system using DistAttention, improves cloud-based LLM performance and context length handling through optimized resource management.
Quan Wang, Yiling Huang, Guanlong Zhao +3 authors
DiarizationLM framework enhances post-processed speaker diarization outputs by utilizing large language models, improving readability and reducing word diarization error rates.
Linyuan Gong, Mostafa Elhoushi, Alvin Cheung
AST-T5, a pretraining paradigm leveraging Abstract Syntax Trees, enhances code generation, transpilation, and understanding, particularly in code-to-code tasks, outperforming similar models like CodeT5.
Zhaowei Li, Qi Xu, Dong Zhang +9 authors
LEGO is a multi-modal grounding model that improves fine-grained understanding and localization of specific regions in images or moments in videos.
Gal Yona, Roee Aharoni, Mor Geva
GRANOLA QA evaluates model answers against multi-granularity responses, revealing that a new decoding method improves accuracy and highlights limitations in standard evaluation methods.
Chaoyang Wang, Peiye Zhuang, Aliaksandr Siarohin +4 authors
A finetuned RGB-D diffusion model combined with dynamic and static Neural Radiance Fields (NeRF) enhances dynamic novel view synthesis by achieving geometric consistency and robust hallucination of unseen video regions.
Hui Wu, Yi Gan, Feng Yuan +7 authors
The proposed LLM inference solution reduces latency and increases throughput by simplifying decoder layers, using a segment KV cache policy, and optimizing the Scaled-Dot-Product-Attention kernel.
Kevin Cai, Chonghua Liu, David M. Chan
Anim-400K is a large-scale dataset for various video-related tasks, including automated dubbing, with over 425K aligned segments in Japanese and English.
Thiemo Alldieck, Nikos Kolotouros, Cristian Sminchisescu
The paper addresses issues in the Score Distillation Sampling loss function by training a shallow network to remove noisy gradients introduced by high text guidance, demonstrating improvements in various applications like image synthesis, network training, and 3D synthesis.
Alex Gu, Baptiste Rozière, Hugh Leather +3 authors
CRUXEval, a benchmark of Python functions with input-output pairs, evaluates code models and reveals performance differences compared to HumanEval, with GPT-4 showing better results but still falling short.
Kumara Kahatapitiya, Adil Karjauv, Davide Abati +3 authors
Object-Centric Diffusion reduces memory and computational cost in diffusion-based video editing by focusing on salient regions and merging redundant tokens in the background, achieving up to a 10x latency reduction with comparable quality.
Dejia Xu, Ye Yuan, Morteza Mardani +4 authors
An Amortized Generative 3D Gaussian framework (AGG) efficiently generates 3D Gaussians from a single image, offering competitive performance and speed advantages over existing methods.
Prince Aboagye, Yan Zheng, Junpeng Wang +8 authors
A new evaluation method for pretrained models uses the consistency between entity representations and meta features across various domains.
Zhuoyuan Mao, Yen Yu
AlignInstruct enhances machine translation quality by introducing cross-lingual supervision using statistical word alignments, outperforming generative methods and improving zero-shot translation capabilities.
Brian Thompson, Mehak Preet Dhaliwal, Peter Frisch +2 authors
Multi-way parallel machine-generated translations dominate web content in lower resource languages, often stemming from low-quality English content, raising concerns for model training.
Yang Liu, Li Wan, Yun Li +5 authors
DI-AEC and FADI-AEC are diffusion-based methods for AEC that enhance processing efficiency and accuracy using far-end signals.
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