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
Bo-Yu Cheng, Wei-Chen Chiu, Yu-Lun Liu
Proposed algorithm uses decomposed low-rank tensor and convolutional Gaussian filters for joint optimization of camera pose and scene geometry, improving performance in novel view synthesis and convergence speed.
Delong Chen, Samuel Cahyawijaya, Jianfeng Liu +2 authors
Subobject-level tokenization in vision models improves image description learning by using semantically meaningful segments instead of fixed-size patches.
Shitao Tang, Jiacheng Chen, Dilin Wang +6 authors
MVDiffusion++, a neural architecture for 3D object reconstruction, synthesizes high-resolution views from limited images using a pose-free self-attention mechanism and a view dropout strategy, achieving superior performance in novel view synthesis and text-to-3D applications.
Jun Zhao, Can Zu, Hao Xu +6 authors
LongAgent, a multi-agent collaboration method, scales large language models to handle up to 128K tokens with improved accuracy and performance in long-text processing tasks compared to GPT-4.
Zekun Li, Zhiyu Zoey Chen, Mike Ross +7 authors
A novel function calling approach for dialogue state tracking using large language models enhances zero-shot and fine-tuned performance in task-oriented dialogues.
Letian Fu, Gaurav Datta, Huang Huang +7 authors
The study presents a new dataset and multimodal model that integrates touch, vision, and language, showing improvements in touch-vision-language alignment and visual-tactile understanding.
Run-Ze Fan, Xuefeng Li, Haoyang Zou +5 authors
ReAlign, a simple approach that reformats instruction data responses, enhances the quality and alignment of large language models without additional data or advanced techniques.
Lanqing Guo, Yingqing He, Haoxin Chen +9 authors
A self-cascade diffusion model that uses multi-scale upsampling and a pivot-guided noise re-schedule strategy efficiently adapts high-resolution image and video generation from a low-resolution diffusion model.
Chen Yang, Sikuang Li, Jiemin Fang +5 authors
GaussianObject framework uses Gaussian splatting and diffusion models to reconstruct and render 3D objects from sparse views, achieving high-quality results with only 4 images.
Junfei Xiao, Zheng Xu, Alan Yuille +2 authors
A progressively aligned language model improves vision-language alignment, achieving faster convergence and higher performance with fewer parameters than state-of-the-art models.
Cheng Chi, Zhenjia Xu, Chuer Pan +5 authors
UMI is a framework for transferring human manipulation skills to robots using hardware-agnostic policies that generalize well to new tasks and environments.
Xingyou Song, Oscar Li, Chansoo Lee +5 authors
OmniPred, a framework training language models as universal end-to-end regressors, outperforms traditional regression models using textual representations of parameters and values from diverse experimental data.
Kangle Deng, Timothy Omernick, Alexander Weiss +4 authors
A LightControlNet-based pipeline automatically generates high-quality, relightable textures for 3D meshes from text prompts using a two-stage process.
Liyan Tang, Igor Shalyminov, Amy Wing-mei Wong +11 authors
Existing large language models generate factually inconsistent topic-focused dialogue summaries and perform poorly as binary factual evaluators, whereas specialized non-LLM metrics capture diverse hallucination types more effectively.
Yusu Qian, Haotian Zhang, Yinfei Yang +1 authors
MAD-Bench is a benchmark designed to evaluate the susceptibility of Multimodal Large Language Models (MLLMs) to deceptive prompts, with GPT-4V performing significantly better than other models but still facing limitations.
Dawei Gao, Zitao Li, Weirui Kuang +11 authors
AgentScope is a developer-centric platform for multi-agent applications, facilitating message exchange, fault tolerance, and multi-modal data handling with an actor-based framework.
Jonas Geiping, Alex Stein, Manli Shu +3 authors
Adversarial attacks on large language models can cause a variety of unintended behaviors, including misdirection, control, denial-of-service, and data extraction, often due to coding capabilities and unsecured tokens in pre-training.
Zizheng Pan, Bohan Zhuang, De-An Huang +5 authors
T-Stitch, a technique for efficient sampling from diffusion probabilistic models, leverages smaller models early in the process to reduce computational cost without degrading image quality.
Zheng-Xin Yong, Cristina Menghini, Stephen H. Bach
A method for generating task-specific data in low-resource languages using bilingual lexicons and outperforming existing translation methods in sentiment analysis and topic classification.
Alex Havrilla, Sharath Raparthy, Christoforus Nalmpantis +4 authors
Stepwise ORMs and combinations of global and local refinement models improve reasoning accuracy in state-of-the-art language models by leveraging synthetic data and ORMs for better error detection and correction.
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