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
Siqi Zheng, Luyao Cheng, Yafeng Chen +2 authors
A large-scale speech corpus, 3D-Speaker, facilitates research on disentangling speech representations by varying speakers, recording devices, distances, and dialects, enabling the evaluation of universal speech models and out-of-domain learning.
Kalyani Marathe, Mahtab Bigverdi, Nishat Khan +4 authors
A dataset-curation method for pixelwise dense prediction tasks, using unlabeled video datasets and synthetic environments, produces larger datasets that outperform annotated ones and enhance downstream task performance.
Dhruv Shah, Ajay Sridhar, Nitish Dashora +4 authors
ViNT, a Transformer-based foundation model, excels in vision-based robotic navigation by learning from diverse datasets and adapting to various tasks through techniques like diffusion and prompt-tuning.
Michael J. Black, Priyanka Patel, Joachim Tesch +1 authors
Synthetic data with realistic simulation of clothing dynamics and diverse scenarios achieves top accuracy in 3D human pose and shape estimation from real images.
Hao-Lun Sun, Lei Hsiung, Nandhini Chandramoorthy +2 authors
NeuralFuse is a model-agnostic add-on for deep neural networks that mitigates low-voltage-induced bit errors in SRAM, reducing memory access energy and improving accuracy without retraining.
Yixuan Wang, Yunzhu Li, Katherine Driggs-Campbell +2 authors
A dynamic-resolution particle representation using graph neural networks achieves better performance in object pile manipulation tasks compared to fixed-resolution models.
Zeyi Liu, Arpit Bahety, Shuran Song
A framework called REFLECT uses LLMs to generate informative failure explanations from multi-sensory data, aiding in correction planning for robots.
Naama Pearl, Yaron Brodsky, Dana Berman +4 authors
SVNR, a novel denoising diffusion model, effectively handles realistic, spatially-variant noise by adapting the diffusion process with pixel-wise time embeddings and conditioning.
Esin Durmus, Karina Nyugen, Thomas I. Liao +15 authors
A quantitative framework evaluates the similarities between large language model responses and human perspectives on global issues, highlighting biases and limitations.
Xiaochuang Han, Daniel Simig, Todor Mihaylov +3 authors
Investigating the pretraining data reveals that long-tail tokens and challenging examples with low long-range context contribute to the in-context learning ability of language models.
Fuxiao Liu, Kevin Lin, Linjie Li +3 authors
A new dataset and evaluation method are introduced to reduce hallucinations in large multi-modal models by incorporating both positive and negative instructions, improving model robustness and performance.
Haochen Shi, Huazhe Xu, Samuel Clarke +2 authors
A robotic system, RoboCook, utilizes point cloud scene representations and Graph Neural Networks to learn and adapt to complex soft object manipulation tasks with flexible tools, outperforming existing methods.
Jingyuan Zhu, Huimin Ma, Jiansheng Chen +1 authors
DomainStudio enhances DDPMs to generate diverse and high-quality images in target domains using limited data, improving overfitting and quality compared to GANs.
Guillaume Couairon, Marlène Careil, Matthieu Cord +2 authors
ZestGuide enhances pre-trained text-to-image diffusion models with zero-shot segmentation guidance using implicit segmentation maps from cross-attention layers, improving image quality and alignment accuracy.
Zhengrong Xue, Han Zhang, Jingwen Cheng +5 authors
ArrayBot uses reinforcement learning to discover control policies for distributed manipulation of tabletop objects through tactile feedback, achieving generalization to unseen objects in both simulated and real-world settings.
Nicholas Carlini, Milad Nasr, Christopher A. Choquette-Choo +8 authors
Adversarial attacks on multimodal models can induce unaligned text generation, challenging the robustness of aligned large language models.
Jinkyu Koo, John Yang, Le An +2 authors
Swin-Free introduces size-varying windows to reduce memory copy operations, improving inference speed and accuracy over Swin Transformer in computer vision tasks.
Yilun Xu, Mingyang Deng, Xiang Cheng +3 authors
A new Restart sampling algorithm enhances both speed and quality for generative processes involving differential equations, outperforming previous ODE and SDE samplers in various tasks.
Haihao Shen, Hengyu Meng, Bo Dong +9 authors
A sparse inference software stack for Transformer-based language models significantly improves performance through efficient SpMM operations on CPUs, offering substantial speedups over existing frameworks and solutions.
Samy Jelassi, Stéphane d'Ascoli, Carles Domingo-Enrich +3 authors
Transformers use relative position embeddings for length generalization in simple tasks like addition and propose train set priming to achieve generalization in complex tasks like multiplication.
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