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。
48 篇论文 · 按点赞排序
Jesse Farebrother, Jordi Orbay, Quan Vuong +9 authors
Using categorical cross-entropy for value function training in deep reinforcement learning improves scalability and performance across various tasks compared to traditional regression methods.
Hossein Aboutalebi, Hwanjun Song, Yusheng Xie +7 authors
MAGID, a framework augmenting text-only dialogues with high-quality images using a diffusion model and feedback loops, outperforms existing multimodal dialogue generation methods.
Jiaxiang Cheng, Pan Xie, Xin Xia +7 authors
The Resolution Adapter (ResAdapter) enhances diffusion models to generate images in arbitrary resolutions and aspect ratios without extensive post-processing, preserving style consistency.
Yanjie Ze, Gu Zhang, Kangning Zhang +3 authors
3D Diffusion Policy (DP3) utilizes 3D visual representations for efficient imitation learning, achieving robust and generalized skills with fewer demonstrations compared to existing methods.
Yair Schiff, Chia-Hsiang Kao, Aaron Gokaslan +3 authors
A new family of DNA language models, Caduceus, supports bi-directionality and reverse complementarity, achieving superior performance on long-range genomic sequence modeling tasks.
János Kramár, Tom Lieberum, Rohin Shah +1 authors
Attribution Patching (AtP) is a scalable method for approximating causal attributions in large language models, outperforming other techniques including its improved variant AtP*, which also provides a method to bound false negatives.
Hanlin Tang, Yifu Sun, Decheng Wu +3 authors
EasyQuant is a data-independent, training-free quantization method for large language models that maintains performance while reducing computational and memory requirements.
Rose E. Wang, Pawan Wirawarn, Omar Khattab +2 authors
The backtracing task aims to retrieve specific text segments causing user queries, enhancing content creation by identifying confusion, curiosity, and emotional triggers, outperforming traditional information retrieval methods.
Imad Eddine Toubal, Aditya Avinash, Neil Gordon Alldrin +10 authors
A framework using language-vision models and conversations automates data labeling and produces efficient classifiers for subjective visual concepts, outperforming traditional methods and zero-shot models.
Gen Luo, Yiyi Zhou, Yuxin Zhang +3 authors
A novel method called Mixture-of-Resolution Adaptation (MRA) enhances multimodal large language models by integrating low- and high-resolution visual features, improving visual recognition and efficiency.
Priya Sundaresan, Quan Vuong, Jiayuan Gu +10 authors
RT-Sketch, a goal-conditioned policy using hand-drawn sketches for manipulation, demonstrates robustness and adaptability in varied settings over language and image-conditioned agents.
Weijie Li, Litong Gong, Yiran Zhu +4 authors
The proposed method enhances the fidelity of image-to-video generation by adding noise to image latent information and rectifying noise during the denoising process in diffusion models.
Suneel Belkhale, Tianli Ding, Ted Xiao +6 authors
A method RT-H uses language-motion hierarchies to improve robot imitation learning by predicting low-level actions from high-level language tasks, enhancing flexibility and robustness.
Lukas Höllein, Aljaž Božič, Norman Müller +5 authors
The paper introduces a method for generating photorealistic 3D assets using pretrained text-to-image models integrated with 3D volume-rendering and cross-frame-attention layers, achieving high-quality results in real-world settings.
Amir Barda, Vladimir G. Kim, Noam Aigerman +2 authors
A hybrid method combining mesh and SDF representations enables consistent sculpting and text-guided editing of 3D shapes with high resolution rendering and adaptive triangle allocation.
Toru Lin, Zhao-Heng Yin, Haozhi Qi +2 authors
Deep reinforcement learning policies trained in simulation can effectively transfer to real-world bimanual object manipulation tasks, such as twisting bottle lids, with robust generalization to unseen objects.
Yuanhao Cai, Yixun Liang, Jiahao Wang +5 authors
A 3D Gaussian splatting-based framework, X-Gaussian, is proposed for efficient X-ray novel view synthesis with faster training and inference compared to NeRF.
Jiakai Sun, Han Jiao, Guangyuan Li +3 authors
3DGStream efficiently streams photo-realistic Free-Viewpoint Videos in real-time using 3D Gaussians and a Neural Transformation Cache, competitive in speed, quality, training time, and storage.
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