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。
Suriya Gunasekar, Yi Zhang, Jyoti Aneja +16 authors
A new compact Transformer-based large language model for code, phi-1, achieves high accuracy on coding benchmarks despite having fewer parameters than competing models.
37 篇论文 · 按点赞排序
Paul K. Rubenstein, Chulayuth Asawaroengchai, Duc Dung Nguyen +27 authors
A unified multimodal language model combining text and speech capabilities outperforms existing systems in speech translation and zero-shot speech-to-text translation.
Kai Zhang, Lingbo Mo, Wenhu Chen +2 authors
MagicBrush, a large-scale manually annotated dataset for text-guided real image editing, enhances InstructPix2Pix model performance and highlights challenges in current baselines.
Xu Zhao, Wenchao Ding, Yongqi An +5 authors
A speed-up method using a CNN detector with an instance segmentation branch achieves comparable performance to SAM with 50 times faster runtime by training on a small subset of SAM's dataset.
Kai Lv, Yuqing Yang, Tengxiao Liu +3 authors
A new optimizer reduces memory usage for full parameter fine-tuning of large language models, enabling 65B model training on a single machine with limited GPUs.
Lionel Wong, Gabriel Grand, Alexander K. Lew +4 authors
A framework combining large language models and probabilistic programs translates language into a symbolic language of thought for rational, commonsense reasoning across multiple cognitive domains.
Haocheng Xi, Changhao Li, Jianfei Chen +1 authors
A novel method for training transformers with 4-bit quantization achieves competitive accuracy and accelerates training on current GPUs.
Theo X. Olausson, Jeevana Priya Inala, Chenglong Wang +2 authors
GPT-4 demonstrates superior self-repair capabilities on APPS dataset compared to GPT-3.5, with performance boosted when feedback is provided by GPT-4 or human programmers.
Yaqi Zhang, Di Huang, Bin Liu +7 authors
MotionGPT generates realistic human motion using multimodal control signals by adapting a small portion of LLM parameters to handle unified input instructions.
Sriram Yenamandra, Arun Ramachandran, Karmesh Yadav +15 authors
The HomeRobot OVMM benchmark evaluates robots' ability to manipulate unseen objects in new environments using a combination of simulated and real-world testing.
Matthias Minderer, Alexey Gritsenko, Neil Houlsby
OWL-ST self-training enhances OWLv2 model's open-vocabulary object detection performance by scaling up to over 1B examples, significantly improving rare class detection.
Alessandro Sordoni, Xingdi Yuan, Marc-Alexandre Côté +6 authors
A Deep Language Network (DLN) achieves high performance by stacking language layers and optimizing prompts through variational inference, outperforming single layers and sometimes matching few-shot GPT-4.
Yifei Zeng, Yuanxun Lu, Xinya Ji +3 authors
AvatarBooth generates high-quality 3D avatars from text prompts or images using dual fine-tuned diffusion models and a multi-resolution rendering strategy, achieving superior rendering and geometric quality.
Disha Shrivastava, Denis Kocetkov, Harm de Vries +2 authors
RepoFusion improves code completion by incorporating repository context, leading to better performance on unseen repositories compared to larger models.
Ilija Radosavovic, Baifeng Shi, Letian Fu +3 authors
A self-supervised sensorimotor pre-training method using a Transformer model on sequences of sensorimotor tokens improves robotic performance in tasks like block stacking.
Yelysei Bondarenko, Markus Nagel, Tijmen Blankevoort
Two modifications to the attention mechanism, clipped softmax and gated attention, reduce outliers in transformer models, enabling full INT8 quantization without additional effort.
Yukun Huang, Jianan Wang, Yukai Shi +3 authors
Proposing a redesign in timestep sampling for text-to-3D diffusion models to enhance quality and diversity by aligning with NeRF optimization.
Boxin Wang, Weixin Chen, Hengzhi Pei +16 authors
A comprehensive trustworthiness evaluation of GPT models, focusing on vulnerabilities related to toxicity, bias, adversarial robustness, privacy, ethics, and fairness, reveals that GPT-4, while generally more trustworthy, can be misled by misleading instructions.
Yoni Kasten, Ohad Rahamim, Gal Chechik
SDS-Complete uses a text-to-image diffusion model to reconstruct incomplete point clouds, achieving reduced Chamfer loss on out-of-distribution objects.
Mahan Fathi, Jonathan Pilault, Pierre-Luc Bacon +3 authors
A hybrid model combining state space models and block-wise attention outperforms Transformer-based architectures in language modeling, offering faster processing speeds.
Ian R. McKenzie, Alexander Lyzhov, Michael Pieler +24 authors
Empirical evidence suggests that large language models may show inverse scaling in performance on certain tasks, leading to the importance of careful consideration of training data and objectives.
Fahad Shamshad, Muzammal Naseer, Karthik Nandakumar
A novel two-step method uses adversarial latent codes in a pretrained generative model to enhance facial privacy while maintaining identity in face recognition systems.
Ori Gordon, Omri Avrahami, Dani Lischinski
Blended-NeRF framework enables robust 3D scene editing using text prompts or image patches with a volumetric blending technique, supporting various editing tasks.
Konstantinos Bousmalis, Giulia Vezzani, Dushyant Rao +36 authors
RoboCat, a visual goal-conditioned decision transformer, leverages heterogeneous robotic experience to achieve zero-shot and few-shot adaptation across various tasks and robot embodiments.
Michiel de Jong, Yury Zemlyanskiy, Nicholas FitzGerald +3 authors
GLIMMER enhances memory-augmented language models by using a shallow reranker and multi-task training, improving retrieval quality and speed over LUMEN and FiD on knowledge-intensive tasks.
Zhiyuan Zhang, Zhitong Huang, Jing Liao
A novel framework using masked textual inversion and a training-free optimization method for layout editing of single images preserves visual properties and outperforms existing methods.
Ayush Tewari, Tianwei Yin, George Cazenavette +5 authors
A novel denoising diffusion model integrates a forward model to sample indirect signals, enabling end-to-end training for challenging vision tasks like inverse graphics.
Lakshya A Agrawal, Aditya Kanade, Navin Goyal +2 authors
Using static analysis monitors during code generation improves the accuracy and compilation rates of language models by providing context-aware guidance.
Shaolei Zhang, Qingkai Fang, Zhuocheng Zhang +8 authors
BayLing, an instruction-following LLM based on LLaMA, achieves performance comparable to GPT-3.5-turbo using automatic translation for instructing tuning across languages, with strong results in translation and knowledge assessment tasks.
Yinxuan Huang, Tonglin Chen, Zhimeng Shen +3 authors
OCTScenes, a comprehensive real-world dataset of tabletop scenes, addresses the challenge of applying object-centric representation learning methods to complex, real-world environments.
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