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
Wang Qun, Liu Yang, Lin Qingquan +1 authors
Xmodel-1.5, a 1-billion-parameter multilingual model pretrained on 2 trillion tokens, achieves strong performance across multiple languages and contributes a Thai evaluation dataset.
Yongliang Wu, Xinting Hu, Yuyang Sun +5 authors
Number-Prompt (NumPro) enhances Video Large Language Models (Vid-LLMs) for Video Temporal Grounding (VTG) by adding numerical identifiers to video frames, improving performance significantly without increased computational costs.
Javier Ferrando, Oscar Obeso, Senthooran Rajamanoharan +1 authors
Sparse autoencoders reveal that large language models can recognize known entities and have internal self-knowledge about their capabilities, influencing their tendency to hallucinate.
Boyuan Jiang, Xiaobin Hu, Donghao Luo +7 authors
FitDiT, a novel garment perception enhancement technique using Diffusion Transformers, addresses texture-aware maintenance and size-aware fitting in high-fidelity virtual try-on.
Huizhuo Yuan, Yifeng Liu, Shuang Wu +2 authors
MARS, a unified optimization framework combining preconditioned gradient methods with variance reduction, improves the training efficiency of large models like GPT-2 compared to AdamW.
Ciara Rowles
StyleCodes is an open-source style encoder enabling users to generate style-reference codes for image generation with minimal quality loss.
M. Arda Aydın, Efe Mert Çırpar, Elvin Abdinli +2 authors
ITACLIP enhances CLIP's semantic segmentation through architectural modifications, data augmentation, and language model-generated definitions, outperforming existing benchmarks.
Thang M. Pham, Phat T. Nguyen, Seunghyun Yoon +3 authors
SlimLM, a series of small language models optimized for document assistance tasks on mobile devices, demonstrates efficient performance and enhanced capabilities within size and time constraints, providing a benchmark for on-device language models.
Ruiyuan Gao, Kai Chen, Bo Xiao +3 authors
MagicDriveDiT enhances video synthesis for autonomous driving by integrating flow matching, progressive training, and spatial-temporal conditional encoding, improving both resolution and control.
Huzaifa Sidhpurwala, Garth Mollett, Emily Fox +2 authors
This paper explores the rapidly evolving ecosystem of publicly available AI models, and their potential implications on the security and safety landscape. As AI models become increasingly prevalent, understanding their potential risks and vulnerabilities is crucial. We review the current security and safety scenarios while highlighting challenges such as tracking issues, remediation, and the apparent absence of AI model lifecycle and ownership processes. Comprehensive strategies to enhance security and safety for both model developers and end-users are proposed. This paper aims to provide some of the foundational pieces for more standardized security, safety, and transparency in the development and operation of AI models and the larger open ecosystems and communities forming around them.
Shehzaad Dhuliawala, Ilia Kulikov, Ping Yu +4 authors
Adaptive Decoding with Latent Preference Optimization dynamically adjusts sampling temperature during model inference, improving performance across tasks requiring varied temperature settings.
Chang Liu, Rui Li, Kaidong Zhang +2 authors
StableV2V is a shape-consistent video editing method that decomposes the editing process into sequential steps to align motions with user prompts, outperforming existing methods in visual consistency and efficiency.
Jinqiang Long, Yanqi Dai, Guoxing Yang +4 authors
Awaker2.5-VL uses a Mixture of Experts architecture with low-rank adaptation to improve performance in various multimodal tasks.
Yuanhao Cai, He Zhang, Kai Zhang +10 authors
A novel single-stage 3D diffusion model, DiffusionGS, is proposed for generating 3D point clouds from a single view, offering better view consistency, quality, and speed compared to state-of-the-art methods.
Vipula Rawte, Sarthak Jain, Aarush Sinha +9 authors
ViBe is a large-scale Text-to-Video Benchmark for evaluating hallucination in video generation with annotations for five types of hallucination, using ensemble classifiers to improve detection and mitigation.
Yunchao Yao, Uksang Yoo, Jean Oh +2 authors
SWIFT, a system for learning dynamic pen spinning tasks with a soft robotic hand, achieves high success rates using real-world data without prior object knowledge, highlighting the potential for soft robotics in dynamic manipulation.
Jan Pfister, Julia Wunderle, Andreas Hotho
Two German-only decoder models, LL\"aMmlein 120M and 1B, were developed from scratch and achieved competitive performance on the SuperGLEBer benchmark, providing insights into the scalability and resource allocation for such models.
Joseph Liu, Joshua Geddes, Ziyu Guo +2 authors
SmoothCache accelerates DiT inference by caching key features, achieving significant speed-ups while maintaining quality across image, video, and audio tasks.
Zewen Chen, Juan Wang, Wen Wang +8 authors
SEAGULL, a vision-language model-based network, assesses quality for Regions of Interest using masks and a mask-based feature extractor, achieving high performance on fine-grained ROI quality assessment.
Bethel Melesse Tessema, Akhil Kedia, Tae-Sun Chung
UnifiedCrawl efficiently collects data from Common Crawl to fine-tune multilingual LLMs using adapter methods, enhancing performance on low-resource languages.
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