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
Aether Team, Haoyi Zhu, Yifan Wang +8 authors
Aether is a unified framework that combines 4D dynamic reconstruction, action-conditioned video prediction, and goal-conditioned visual planning, achieving zero-shot generalization and superior reconstruction without real-world data.
Jinwei Qi, Chaonan Ji, Sheng Xu +3 authors
The novel framework for real-time portrait video generation uses hierarchical motion diffusion models and explicit hand control to produce expressive video chats with synchronized head and body movements.
Jianchuan Chen, Jingchuan Hu, Gaige Wang +4 authors
TaoAvatar is a high-fidelity, lightweight 3DGS-based avatar that achieves real-time performance across devices while maintaining high rendering quality, leveraging distillation techniques for non-rigid deformation.
Shitian Zhao, Qilong Wu, Xinyue Li +10 authors
A suite called LeX-Art for high-quality text-image synthesis includes data-centric pipeline, prompt enrichment, and text-to-image models, achieving state-of-the-art performance with a new benchmark and metric.
Edoardo Debenedetti, Ilia Shumailov, Tianqi Fan +7 authors
CaMeL defends LLM agents from prompt injection attacks by segregating control and data flows and preventing unauthorized data exfiltration, achieving 67% secure task completion in AgentDojo.
Yihe Deng, Hritik Bansal, Fan Yin +3 authors
Iterative supervised fine-tuning and reinforcement learning improve reasoning capabilities in large vision-language models, enhancing performance on multimodal reasoning tasks like MathVista, MathVerse, and MathVision.
Wenqi Zhang, Mengna Wang, Gangao Liu +10 authors
Embodied Reasoner, a deep learning model for interactive embodied tasks, outperforms existing visual reasoning models by leveraging observation-thought-action trajectories and a training pipeline that includes imitation learning, exploration, and reflection.
Weichen Fan, Amber Yijia Zheng, Raymond A. Yeh +1 authors
CFG-Zero* enhances classifier-free guidance in flow matching models by correcting velocity inaccuracies and zeroing early ODE solver steps, improving fidelity and controllability in text-to-image and text-to-video tasks.
Shu Pu, Yaochen Wang, Dongping Chen +10 authors
MLLMs evaluate cross-modal understanding and generation tasks across various modalities but face challenges with MMG tasks, leading to the proposal of OmniArena for more comprehensive evaluation.
Yiran Qin, Li Kang, Xiufeng Song +5 authors
Compositional constraints and tailored interfaces are introduced to automate data collection and improve imitation learning in embodied multi-agent systems, enabling the development of safe and efficient manipulation tasks.
Qi Qin, Le Zhuo, Yi Xin +20 authors
Lumina-Image 2.0, a text-to-image framework, achieves significant advancements through a unified architecture and efficient training, enabling high-quality image generation with reduced parameters.
Prin Phunyaphibarn, Phillip Y. Lee, Jaihoon Kim +1 authors
Replacing unconditional noise predictions in Classifier-Free Guidance with those from a pre-trained model improves conditional generation quality for diffusion models.
Samuel Schmidgall, Michael Moor
AgentRxiv facilitates collaboration among LLM agent laboratories, enabling iterative improvements and higher performance across various benchmarks.
Yujie Liu, Zonglin Yang, Tong Xie +7 authors
LLMs demonstrate proficiency in inspiration retrieval using a newly developed benchmark, showing potential in automated scientific hypothesis generation.
Siwei Han, Peng Xia, Ruiyi Zhang +4 authors
MDocAgent, a multi-modal multi-agent framework, enhances document understanding by integrating both textual and visual cues, improving question answering accuracy through collaborative agent insights.
Akhiad Bercovich, Mohammad Dabbah, Omri Puny +15 authors
FFN Fusion reduces sequential computation in large language models by parallelizing FFN layers, achieving significant speedup and cost reduction while maintaining performance.
Yufei Zhan, Yousong Zhu, Shurong Zheng +4 authors
Vision-R1 is a vision-guided reinforcement learning algorithm that enhances Large Vision-Language Models without requiring human-annotated preference data, achieving significant performance improvements.
Yupeng Cao, Haohang Li, Yangyang Yu +10 authors
FinAudio is introduced as the first benchmark to evaluate AudioLLMs for financial audio tasks, including short and long ASR, and summarization, highlighting their limitations.
Siwei Wen, Junyan Ye, Peilin Feng +7 authors
FakeVLM, a specialized multimodal model, enhances synthetic image and DeepFake detection with human-interpretable explanations.
Mingyang Chen, Tianpeng Li, Haoze Sun +9 authors
ReSearch is a framework that trains LLMs to integrate reasoning with search using reinforcement learning, enhancing capabilities for complex multi-hop questions.
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