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
Lingteng Qiu, Xiaodong Gu, Peihao Li +8 authors
LHM, a large-scale animatable human reconstruction model using 3D Gaussian splatting and a multimodal transformer, achieves high-fidelity avatar generation with detailed face identity preservation and multi-scale feature encoding.
Zhaowei Liu, Xin Guo, Fangqi Lou +13 authors
Fin-R1, a large language model tailored for finance, achieves state-of-the-art performance in financial reasoning tasks using supervised fine-tuning and reinforcement learning.
Xinyu Ma, Ziyang Ding, Zhicong Luo +5 authors
DeepPerception enhances Multimodal Large Language Models with cognitive visual perception capabilities, improving accuracy and cross-domain generalization through a novel visual grounding task and a two-stage training framework.
Foundation AI Team, Kiran Bhat, Nishchaie Khanna +42 authors
A foundation model for 3D intelligence tokenizes 3D shapes and integrates with LLMs for tasks like text-to-shape, shape-to-text, and text-to-scene generation, enabling scene analysis and reasoning in 3D development.
Xinyu Lian, Zichao Yu, Ruiming Liang +9 authors
A novel method, Infinite Mobility, synthesizes high-fidelity articulated objects using procedural generation, excelling state-of-the-art methods in quality and usability for training generative models.
Jingyi Zhang, Jiaxing Huang, Huanjin Yao +4 authors
A reinforcement learning framework enhances MLLMs' reasoning by rewarding step-wise accurate and valid reasoning paths, leading to improved performance on reasoning benchmarks.
Lan Chen, Qi Mao, Yuchao Gu +1 authors
Edit Transfer, a visual relation in-context learning method, learns from a single example to apply transformations to new images, outperforming text-only and appearance-centric methods with few-shot learning.
Keda Tao, Haoxuan You, Yang Sui +2 authors
VidKV, a new KV cache quantization method for VideoLLMs, compresses the KV cache to below 2 bits with minimal impact on performance by employing mixed-precision and channel-wise quantization techniques.
Zeyue Tian, Yizhu Jin, Zhaoyang Liu +5 authors
AudioX, a unified Diffusion Transformer, generates high-quality audio and music with natural language control and supports various modalities through masked multi-modal training.
Kanzhi Cheng, Wenpo Song, Jiaxin Fan +7 authors
CapArena evaluates current VLMs on image captioning, finding that models like GPT-4o surpass human performance, while automated metrics show biases; CapArena-Auto provides an accurate and efficient automated benchmark.
Yaowei Li, Lingen Li, Zhaoyang Zhang +6 authors
BlobCtrl is a framework that uses a probabilistic blob-based representation for element-level generation and editing, offering precision and flexibility through a dual-branch diffusion architecture, self-supervised training, and controllable dropout strategies.
Paul Engstler, Aleksandar Shtedritski, Iro Laina +2 authors
SynCity combines 3D and 2D generative models to create large, high-quality, and immersive 3D worlds from textual descriptions.
Xingtai Lv, Youbang Sun, Kaiyan Zhang +8 authors
This survey provides a comprehensive overview of State Space Models (SSMs), their mathematical foundations, and their applications, emphasizing their efficiency and effectiveness compared to transformer-based models.
Tao Yu, Yi-Fan Zhang, Chaoyou Fu +14 authors
A comprehensive review of alignment algorithms for multimodal large language models, covering their application scenarios, dataset construction, evaluation benchmarks, and future directions.
Qizhi Pei, Lijun Wu, Zhuoshi Pan +6 authors
A new framework, MathFusion, enhances mathematical reasoning by synthesizing cross-problem instructions, leading to significant improvements in accuracy with fewer additional synthetic instructions.
Jing Bi, Junjia Guo, Susan Liang +8 authors
VERIFY is a benchmark designed to evaluate the visual reasoning capabilities of MLLMs by minimizing textual context and assessing reasoning paths and fidelity.
Honglin Lin, Zhuoshi Pan, Yu Li +5 authors
MetaLadder enhances LLMs' mathematical reasoning by prompting them to recall analogous problems and restate the target question, significantly improving accuracy over standard methods.
Jian Zhang, Bifan Wei, Shihao Qi +3 authors
A unified framework for constructing generalized knowledge graphs improves performance across different knowledge graph types by leveraging a three-stage curriculum learning fine-tuning approach.
Harold Haodong Chen, Haojian Huang, Xianfeng Wu +5 authors
FluxFlow, a temporal augmentation strategy, enhances the temporal coherence and diversity in video generation without modifying model architectures.
Defa Zhu, Hongzhi Huang, Jundong Zhou +5 authors
Frac-Connections, a technique that divides hidden states, outperform traditional residual connections in deep learning by balancing memory consumption and performance, validated through experiments on large language 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