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
Shengjia Zhang, Junjie Wu, Jiawei Chen +6 authors
OThink-R1 is introduced to reduce reasoning redundancy in complex problem-solving by classifying reasoning steps as essential or redundant and dynamically switching thinking modes based on task complexity.
Junyu Luo, Zhizhuo Kou, Liming Yang +10 authors
FinMME is a comprehensive multimodal dataset for financial research and FinScore is an evaluation system that highlights the challenges faced by even advanced models like GPT-4o in the finance domain.
Jiwan Chung, Junhyeok Kim, Siyeol Kim +3 authors
v1 enhances Multimodal Large Language Models by enabling selective and dynamic visual region retrieval during inference, improving performance on multimodal reasoning tasks.
Yuhao Wu, Yushi Bai, Zhiqiang Hu +2 authors
SuperWriter-Agent enhances long-form text generation by integrating structured planning and refinement, achieving top performance with a 7B model and hierarchical Direct Preference Optimization.
Wei Fu, Jiaxuan Gao, Xujie Shen +10 authors
AReaL, a fully asynchronous reinforcement learning system, decouples generation and training to achieve higher GPU utilization and up to 2.57x training speedup for large language models on reasoning tasks.
Junliang Ye, Zhengyi Wang, Ruowen Zhao +2 authors
A native 3D large language model, ShapeLLM-Omni, is proposed to understand and generate 3D assets and text, trained using a 3D vector-quantized variational autoencoder and a new 3D-Alpaca dataset.
Yi Yang, Jiaxuan Sun, Siqi Kou +2 authors
A unified vision language action framework, LoHoVLA, combines a large pretrained vision language model with hierarchical closed-loop control to improve performance on long-horizon embodied tasks.
Adrian Łańcucki, Konrad Staniszewski, Piotr Nawrot +1 authors
Inference-time hyper-scaling with Dynamic Memory Sparsification in Transformer LLMs allows for increased token generation within the same compute budget by compressing the key-value cache, thereby enhancing accuracy.
Kejian Zhu, Zhuoran Jin, Hongbang Yuan +6 authors
A new benchmark, MMR-V, is proposed to challenge multimodal large language models with long-range, multi-frame reasoning and hidden information processing in videos, revealing their limitations and inspiring further research.
Tarun Suresh, Debangshu Banerjee, Shubham Ugare +2 authors
DINGO, a dynamic programming-based decoding strategy, enhances diffusion language models by enforcing structured output constraints, significantly improving performance on symbolic math and JSON generation tasks.
Cailin Zhuang, Ailin Huang, Wei Cheng +12 authors
ViStoryBench is a comprehensive evaluation benchmark for story visualization frameworks, featuring diverse datasets and metrics to assess model performance across narrative and visual dimensions.
Dongyoung Kim, Sumin Park, Huiwon Jang +3 authors
Robot-R1, a reinforcement learning framework, enhances embodied reasoning for robotics by predicting keypoint states, outperforming supervised fine-tuning methods and even surpassing GPT-4o in low-level action control tasks.
Ruihan Yang, Yikai Zhang, Aili Chen +5 authors
ARIA aggregates rewards in an intention space to mitigate reward sparsity and improve policy optimization in language-based reinforcement learning tasks.
Kejian Zhu, Shangqing Tu, Zhuoran Jin +3 authors
A method called shortcut neuron patching identifies and suppresses shortcut neurons in language models to mitigate data contamination issues in trustworthy evaluations.
Yipeng Du, Tiehan Fan, Kepan Nan +6 authors
MotionSight, a zero-shot method using object-centric visual spotlight and motion blur as prompts, enhances fine-grained video motion understanding and achieves state-of-the-art performance on MotionVid-QA, a large-scale dataset with hierarchical annotations.
Chenbin Pan, Wenbin He, Zhengzhong Tu +1 authors
DINO-R1 incorporates reinforcement learning to enhance visual in-context reasoning capabilities in vision foundation models, achieving better performance than supervised fine-tuning across various visual prompting scenarios.
Yizhuo Li, Yuying Ge, Yixiao Ge +2 authors
A novel framework using flow-based generative models aligns learnable latent spaces to target distributions, reducing computational expense and improving log-likelihood maximization.
Tianyu Huang, Wangguandong Zheng, Tengfei Wang +8 authors
Voyager is a video diffusion framework that generates world-consistent 3D point-cloud sequences from a single image, enabling long-range, consistent 3D scene exploration with user-defined camera paths.
Yuansheng Ni, Ping Nie, Kai Zou +2 authors
VisCode-200K, a large-scale dataset for visualization, improves plot generation performance by integrating execution-grounded supervision and iterative code correction, outperforming open-source models and rivaling proprietary ones.
Xiangdong Zhang, Jiaqi Liao, Shaofeng Zhang +4 authors
VideoREPA enhances text-to-video synthesis by aligning token-level relations and distilling physics understanding from foundation models into T2V models.
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