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。
Andrés Marafioti, Orr Zohar, Miquel Farré +14 authors
SmolVLM, a series of compact multimodal models, achieves high performance with minimal GPU memory usage, making efficient deployment on mobile and edge devices possible.
50 篇论文 · 按点赞排序
Yiying Yang, Wei Cheng, Sijin Chen +6 authors
OmniSVG, using pre-trained Vision-Language Models and discrete tokenization, generates high-quality SVGs efficiently with a large, multimodal dataset.
Kimi Team, Angang Du, Bohong Yin +89 authors
Kimi-VL, an efficient Mixture-of-Experts vision-language model, excels in multimodal reasoning, long-context understanding, and diverse vision-language tasks, achieving competitive performance with reduced computational cost.
Gleb Rodionov, Roman Garipov, Alina Shutova +5 authors
A parallel inference engine for LLMs allows concurrent task collaboration through a shared attention cache, improving efficiency without additional fine-tuning.
Karan Dalal, Daniel Koceja, Gashon Hussein +12 authors
Test-Time Training (TTT) layers enable pre-trained Transformers to generate coherent one-minute videos from text storyboards, outperforming alternatives like Mamba~2 and Gated DeltaNet.
Yi Peng, Chris, Xiaokun Wang +12 authors
Skywork R1V extends large language models to multimodal reasoning with efficient transfer, enhanced visual-text alignment, and dynamic reasoning chain optimization, achieving competitive performance in various benchmarks.
Sara Vera Marjanović, Arkil Patel, Vaibhav Adlakha +14 authors
DeepSeek-R1 generates detailed reasoning chains, offering insights into model behavior and cognitive phenomena, with findings highlighting performance nuances and safety concerns.
Jiacheng Liu, Taylor Blanton, Yanai Elazar +28 authors
OLMoTrace traces language model outputs to their training data in real-time, providing insights into fact checking, hallucination, and creativity through verbatim matches and infini-gram technology.
Shuai Wang, Zhi Tian, Weilin Huang +1 authors
A decoupled diffusion transformer improves performance and training speed in image generation by separating semantic extraction and high-frequency decoding.
Sixiang Chen, Jinbin Bai, Zhuoran Zhao +16 authors
An empirical study of GPT-4o's image generation capabilities across multiple tasks reveals its strengths and limitations compared to other models, highlighting the importance of architectural design and data scaling in unified generative frameworks.
Zhongyang Li, Ziyue Li, Tianyi Zhou
A novel test-time optimization method, C3PO, improves the accuracy and efficiency of Mixture-of-Experts Large Language Models by re-weighting experts based on surrogate objectives from neighboring samples.
Zhong-Yu Li, Ruoyi Du, Juncheng Yan +5 authors
VisualCloze framework enhances universal image generation through visual in-context learning, Graph200K dataset, and leveraging pre-trained infilling models.
Daoguang Zan, Zhirong Huang, Wei Liu +16 authors
A multilingual benchmark, Multi-SWE-bench, is introduced to evaluate Large Language Models across various programming languages and is complemented by a RL dataset, Multi-SWE-RL, to advance issue-resolving tasks.
Yukun Qi, Yiming Zhao, Yu Zeng +7 authors
VCR-Bench is a new benchmark designed to evaluate video chain-of-thought reasoning in large vision-language models, assessing both perception and reasoning capabilities.
M-A-P Team, Siwei Wu, Jincheng Ren +29 authors
A new Chinese preference dataset and reward model are introduced to scale human preference alignment using LLMs, enhancing performance and cost-effectiveness.
Minki Kang, Jongwon Jeong, Jaewoong Cho
Tool-integrated self-verification enhances small language models' performance in self-verification tasks and outperforms larger models on benchmarks.
Chenrui Fan, Ming Li, Lichao Sun +1 authors
Reasoning LLMs generate redundant responses to ill-posed questions, a phenomenon named MiP-Overthinking, which is not observed in non-reasoning LLMs and spreads through model distillation.
Essential AI, Darsh J Shah, Peter Rushton +25 authors
Models exhibit self-correcting ability during pre-training by recognizing and addressing errors in their reasoning, a skill that improves over time.
Shengyuan Ding, Shenxi Wu, Xiangyu Zhao +7 authors
MM-IFEngine generates high-quality image-instruction pairs for training Multi-modal Large Language Models, leading to improved performance in instruction-following tasks.
Mengchao Wang, Qiang Wang, Fan Jiang +5 authors
A novel framework uses pretrained video diffusion transformer models to generate realistic, high-fidelity talking portraits with controlled motion dynamics, achieving better realism and identity preservation.
Shaojin Wu, Mengqi Huang, Wenxu Wu +3 authors
A new data synthesis pipeline uses diffusion transformers and UNO, featuring progressive cross-modal alignment and universal rotary position embedding, to generate high-consistency multi-subject data while maintaining controllability.
Sangbeom Lim, Junwan Kim, Heeji Yoon +2 authors
The URECA dataset and model address multi-granularity region captioning by ensuring unique and contextually grounded captions through a stage-wise pipeline and dynamic mask modeling.
Fan Zhou, Zengzhi Wang, Nikhil Ranjan +5 authors
MegaMath is an open dataset designed for math-centric LLM pre-training, combining high-quality web data, math-related code, and synthetic content to enhance diversity and quality.
Mustafa Shukor, Enrico Fini, Victor Guilherme Turrisi da Costa +3 authors
Early-fusion architectures, which do not rely on image encoders, outperform late-fusion models in native multimodal architectures, especially when enhanced with Mixture of Experts (MoEs).
Ruikang Liu, Yuxuan Sun, Manyi Zhang +5 authors
A systematic study on quantized reasoning models evaluates the impact of quantization on performance and identifies critical factors affecting accuracy and inference cost.
Jifang Wang, Xue Yang, Longyue Wang +7 authors
CIGEval, a unified framework leveraging large multimodal models, provides a comprehensive and reliable evaluation mechanism for conditional image generation tasks, achieving high correlation with human assessments.
Yunhan Yang, Yuan-Chen Guo, Yukun Huang +5 authors
HoloPart, a diffusion-based model with local and global attention, addresses 3D part amodal segmentation by completing occluded parts and improving global shape consistency.
YuYue, Yufeng Yuan, Qiying Yu +23 authors
VAPO, a value-based augmented proximal policy optimization framework, achieves state-of-the-art performance on reasoning tasks using a value-based reinforcement learning approach.
Shuofei Qiao, Zhisong Qiu, Baochang Ren +8 authors
KnowSelf, a data-centric approach, enables LLM-based agents to autonomously regulate knowledge utilization and achieve optimal planning with minimal costs by switching between situations.
Mengchen Zhang, Tong Wu, Jing Tan +3 authors
A novel auto-regressive Transformer model, GenDoP, generates high-quality, text-guided, and context-aware camera trajectories using a large-scale multi-modal dataset, improving controllability and motion stability in cinematography.
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