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
Graham Todd, Alexander Padula, Matthew Stephenson +3 authors
A novel approach using large language models and evolutionary computation generates new board games in the Ludii language, exploring previously uncovered rule spaces.
Qi Sun, Marc Pickett, Aakash Kumar Nain +1 authors
Empirical studies on frozen transformers reveal differences between lower/final and middle layers and suggest that these models can maintain robustness and trade accuracy for latency by skipping or rearranging layers.
Tu Vu, Kalpesh Krishna, Salaheddin Alzubi +3 authors
FLAMe, a family of Foundational Large Autorater Models trained on publicly licensed data, outperforms proprietary LLMs across multiple quality assessment tasks and demonstrates reduced bias.
Hongjin Su, Howard Yen, Mengzhou Xia +12 authors
BRIGHT is a new text retrieval benchmark requiring deep reasoning, showing that existing models struggle and performs robustly even with document leakage during training.
Yang Zhao, Di Huang, Chongxiao Li +12 authors
CodeV, an open-source instruction-tuned Verilog generation model, improves Verilog code generation by focusing on summarization, outperforming previous models and GPT-4.
Fei Shen, Xin Jiang, Xin He +5 authors
IMAGDressing-v1 enhances virtual try-on by generating editable human images with fixed garments using a garment UNet, hybrid attention module, and CLIP and VAE features, achieving state-of-the-art performance.
Sherwin Bahmani, Ivan Skorokhodov, Aliaksandr Siarohin +9 authors
A method for controlling camera movement in transformer-based video diffusion models using spatiotemporal embeddings achieves state-of-the-art performance.
Longwen Zhang, Ziyu Wang, Qixuan Zhang +6 authors
CLAY, a 3D geometry and material generator, uses a combination of VAE, DiT, and neural fields to produce high-quality 3D models and textures from various inputs, requiring minimal user expertise.
Jiwook Kim, Seonho Lee, Jaeyo Shin +2 authors
DreamCatalyst improves text-driven 3D editing by interpreting score distillation sampling as a diffusion reverse process, offering faster and higher-quality NeRF scene editing.
Sahil Jain, Avik Kuthiala, Prabhdeep Singh Sethi +1 authors
StyleSplat efficiently applies localized style transfer to multiple 3D objects within scenes using 3D Gaussian splatting and spherical harmonic coefficients alignment.
Damien Sileo
Large language models struggle with suggesting missing elements from long lists due to attention overflow, affecting their performance and novelty derivation.
Mengzhao Chen, Wenqi Shao, Peng Xu +5 authors
Efficient quantization-aware training compresses large language models with reduced memory and minimal accuracy loss through block-wise parameter training and end-to-end quantization parameter optimization.
Yanqin Jiang, Chaohui Yu, Chenjie Cao +3 authors
Animate3D uses a novel multi-view video diffusion model trained on multi-view videos to enhance the animation of static 3D models, improving spatiotemporal consistency and appearance fidelity.
Garrett Tanzer, Biao Zhang
YouTube-SL-25 is a large, multilingual sign language dataset that improves sign-to-text tasks through multilingual transfer learning.
Yongyuan Liang, Tingqiang Xu, Kaizhe Hu +3 authors
Make-An-Agent uses conditional diffusion models to generate policies with a single behavioral demonstration, demonstrating versatility and generalization across tasks and robot platforms.
Shraman Pramanick, Rama Chellappa, Subhashini Venugopalan
SPIQA addresses the limitations of existing QA datasets by introducing a large-scale dataset for interpreting complex figures and tables in scientific papers, using multimodal large language models and a Chain-of-Thought evaluation strategy.
Xiangkun Hu, Tong He, David Wipf
Current DPO methods face limitations in interpolating between pre-trained models and human preferences, so an alternative DPO loss is introduced to address these issues.
Siddharth Jha, Lutfi Eren Erdogan, Sehoon Kim +2 authors
The study evaluates various prompt compression methods, finding extractive compression to outperform others, enabling significant context reduction with minimal accuracy loss.
Pengxiang Li, Zhi Gao, Bofei Zhang +6 authors
FIRE, a feedback-refinement dataset, enhances VLMs' performance through user feedback, leading to a FIRE-LLaVA model that outperforms its untrained counterparts by 50% on a comprehensive benchmark.
Yifan Gong, Zheng Zhan, Yanyu Li +6 authors
A weight generator using diffusion models and GANs accelerates training for image translation tasks by generating denoised weights for initialization.
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