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
Jerry Wei, Le Hou, Andrew Lampinen +8 authors
Symbol tuning enhances language models by using arbitrary symbols instead of natural language labels, improving in-context learning, algorithmic reasoning, and handling flipped labels.
Luke Friedman, Sameer Ahuja, David Allen +10 authors
An end-to-end large-scale conversational recommender system using LLMs addresses challenges in understanding dialogue, retrieving information, and generating synthetic conversations, demonstrating improved personalization and functionality in video recommendations.
Rabeeh Karimi Mahabadi, Jaesung Tae, Hamish Ivison +4 authors
Text-to-text Self-conditioned Simplex Diffusion (TESS) achieves strong performance on natural language tasks by using a fully non-autoregressive approach that applies diffusion in logit simplex space.
Qiuqiang Kong, Ke Chen, Haohe Liu +4 authors
A proposed universal audio source separation framework uses weakly labeled data to achieve significant improvements in separating various sound classes, including music, speech, and sound events.
Peize Sun, Shoufa Chen, Chenchen Zhu +4 authors
A detector that predicts open-vocabulary objects and their parts using multi-granularity alignment outperforms baselines in cross-dataset and cross-category generalization.
David Uthus, Santiago Ontañón, Joshua Ainslie +1 authors
A multilingual text-to-text transformer called mLongT5, combining LongT5 with datasets from mT5 and UL2, outperforms existing models like mBART and M-BERT in summarization and question-answering tasks.
Zorik Gekhman, Jonathan Herzig, Roee Aharoni +2 authors
TrueTeacher, a method for generating synthetic data using a large language model, improves factual consistency evaluation of summaries compared to existing models and synthetic data techniques.
Ang Lv, Xu Tan, Peiling Lu +4 authors
GETMusic, a unified diffusion framework with a novel music representation, generates high-quality music tracks from scratch or conditioned on sources, outperforming previous methods in diverse combinations.
Chenshuo Wang, Shaoguang Mao, Tao Ge +5 authors
The paper introduces the Smart Word Suggestions (SWS) task and benchmark, focusing on end-to-end evaluation and realistic writing assistance scenarios through human-labeled and rules-generated datasets.
Markus Nußbaum-Thom, Lyan Verwimp, Youssef Oualil
The study explores novel feed-forward architectures for building a single, application-agnostic on-device language model to reduce memory usage while maintaining speed and accuracy.
Michal Yarom, Yonatan Bitton, Soravit Changpinyo +5 authors
SeeTRUE evaluates text-image alignment using question generation and visual question answering or end-to-end classification with fine-tuned multimodal models, surpassing prior methods in complex scenarios.
Guangxuan Xiao, Tianwei Yin, William T. Freeman +2 authors
FastComposer achieves efficient, personalized, and high-quality multi-subject text-to-image generation without fine-tuning by using subject embeddings and cross-attention localization supervision.
Shengyi Qian, David F. Fouhey
A transformer-based model predicts 3D object location, properties, and affordance, improving machine interaction with 3D scenes and objects.
Yuxian Gu, Li Dong, Furu Wei +1 authors
PICL enhances language models' in-context learning by pre-training on intrinsic tasks, improving performance and generalization across various text classification and task generation benchmarks.
Enrico Fini, Pietro Astolfi, Adriana Romero-Soriano +2 authors
Improved training techniques and augmentations can effectively boost CLIP's performance on multimodal tasks, surpassing models with additional non-contrastive losses.
Yuang Wang, Xingyi He, Sida Peng +3 authors
AutoRecon automates object reconstruction from multi-view images using self-supervised 2D vision transformers and dense supervision from decomposed point clouds.
Zhengxuan Wu, Atticus Geiger, Christopher Potts +1 authors
DAS, an enhanced gradient descent method, identifies interpretable causal structures in large language models like Alpaca by aligning neural representations with symbolic boolean variables.
Harsh Jhamtani, Hao Fang, Patrick Xia +3 authors
An approach using hierarchical natural language decomposition and a pre-trained language model allows a language-to-code model to handle complex utterances with minimal training data, outperforming standard few-shot prompting.
Le Xue, Ning Yu, Shu Zhang +7 authors
ULIP-2 generates language descriptions for 3D objects using large language models, enabling scalable multimodal 3D representation learning with improved accuracy and fewer parameters.
Zhen Guo, Peiqi Wang, Yanwei Wang +1 authors
Dr. LLaMA uses generative data augmentation from Large Language Models to improve Small Language Models in domain-specific question-answering tasks, specifically enhancing performance on medical datasets like PubMedQA.
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