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。
Jade Copet, Felix Kreuk, Itai Gat +5 authors
MusicGen, a single-stage transformer language model, generates high-quality music conditioned on text or melodic features using efficient token interleaving patterns, outperforming existing models on text-to-music benchmarks.
50 篇论文 · 按点赞排序
Subhabrata Mukherjee, Arindam Mitra, Ganesh Jawahar +3 authors
Orca, a 13-billion parameter model, enhances small models by imitating the reasoning process from large foundation models using rich signals and diverse data, surpassing existing models in complex reasoning benchmarks.
Guilherme Penedo, Quentin Malartic, Daniel Hesslow +6 authors
Filtered and deduplicated web data alone can produce powerful large language models that outperform those trained on The Pile.
Hang Zhang, Xin Li, Lidong Bing
Video-LLaMA is a multi-modal framework that integrates visual and auditory understanding into LLMs by extending image encoders to video and aligning audio signals, achieving effective video-language correspondence.
Bo Li, Yuanhan Zhang, Liangyu Chen +5 authors
MIMIC-IT, a multimodal instruction-response dataset, enhances VLMs' zero-shot performance on vision-language tasks through extensive and diverse data, leading to improved perception, reasoning, and planning.
Youcai Zhang, Xinyu Huang, Jinyu Ma +9 authors
The Recognize Anything Model (RAM) is a powerful image tagging model trained using large-scale image-text pairs, achieving superior zero-shot performance compared to supervised models and Google API.
Hyunwoong Ko, Kichang Yang, Minho Ryu +4 authors
Polyglot focuses on developing specialized Korean language models to improve non-English language performance in multilingual models by leveraging a large dataset.
Qianqian Wang, Yen-Yu Chang, Ruojin Cai +4 authors
OmniMotion, a globally consistent motion representation using a quasi-3D canonical volume, outperforms state-of-the-art methods in motion estimation by modeling camera and object motion through occlusions.
Lei Li, Yuwei Yin, Shicheng Li +9 authors
The M$^3$IT dataset and Ying-VLM model enhance vision-language models' alignment with human instructions across various languages and modalities.
Lei Ke, Mingqiao Ye, Martin Danelljan +4 authors
HQ-SAM enhances the segmentation capabilities of SAM by introducing a learnable output token that leverages additional Vit features, improving mask quality with minimal computational overhead.
Zhenhui Ye, Ziyue Jiang, Yi Ren +5 authors
Ada-TTA synthesizes realistic, identity-preserving, and synchronized talking avatars using a generic zero-shot multi-speaker TTS model and advances in neural rendering.
Yuseung Lee, Kunho Kim, Hyunjin Kim +1 authors
SyncDiffusion synchronizes multiple diffusions via perceptual similarity loss to generate coherent panoramic images without visible seams and maintains fidelity and compatibility with input prompts.
Muhammad Maaz, Hanoona Rasheed, Salman Khan +1 authors
Video-ChatGPT is a multimodal model combining a video-adapted visual encoder with a Large Language Model to understand and generate human-like conversations about videos, trained on a scalable dataset of video-instruction pairs.
Yidong Wang, Zhuohao Yu, Zhengran Zeng +10 authors
PandaLM, a judge LLM, evaluates and distinguishes between LLMs based on both objective correctness and subjective factors, providing a fairer and more cost-effective evaluation method.
Jiachen Li, Jitesh Jain, Humphrey Shi
The Matting Anything Model (MAM) is a single, efficient image matting framework that leverages SAM and a lightweight M2M module for flexible user guidance and performs comparably to specialized models with fewer parameters.
John Kirchenbauer, Jonas Geiping, Yuxin Wen +7 authors
Watermarking LLM-generated text is reliable and effective, especially in scenarios involving human paraphrasing and multiple text sources, with evidence compounding as more samples are observed.
Luming Tang, Menglin Jia, Qianqian Wang +2 authors
Image diffusion models extract implicit features (DIFT) that outperform supervised and weakly-supervised methods in establishing semantic, geometric, and temporal correspondences between images.
Zhan Ling, Yunhao Fang, Xuanlin Li +4 authors
A natural language-based format decomposes reasoning verification into subprocesses, enhancing the rigor and trustworthiness of language model reasoning steps and improving accuracy on complex tasks.
Dongfu Jiang, Xiang Ren, Bill Yuchen Lin
An ensembling framework combines multiple large language models by ranking and merging outputs to achieve superior performance across various metrics.
Yew Ken Chia, Pengfei Hong, Lidong Bing +1 authors
INSTRUCTEVAL provides a comprehensive evaluation suite for instruction-tuned large language models, focusing on problem-solving, writing ability, and alignment to human values.
Yizhong Wang, Hamish Ivison, Pradeep Dasigi +8 authors
Various instruction-tuned language models are evaluated across multiple datasets, revealing differences in skill enhancement and suggesting ongoing need for better base models and data.
Chandra Shekhara Kaushik Valmeekam, Krishna Narayanan, Dileep Kalathil +2 authors
A new estimate of English entropy using LLaMA-7B for next-token prediction leads to a compression algorithm that outperforms current text compression methods.
Ziyue Jiang, Yi Ren, Zhenhui Ye +9 authors
Mega-TTS, a zero-shot TTS system, uses attribute-specific models trained on large datasets to generate high-quality speech with superior naturalness, robustness, and speaker similarity in zero-shot tasks.
Lichang Chen, Jiuhai Chen, Tom Goldstein +2 authors
InstructZero optimizes a soft prompt to generate instructions for black-box LLMs, improving zero-shot performance using Bayesian optimization.
Zezhou Cheng, Carlos Esteves, Varun Jampani +3 authors
LU-NeRF jointly estimates camera poses and neural radiance fields without restrictive assumptions, improving unposed NeRF performance and complementing feature-based SfM pipelines.
Ganesh Jawahar, Haichuan Yang, Yunyang Xiong +10 authors
Mixture-of-supernets with an architecture-based routing mechanism improves NAS efficiency and performance in NLP tasks by minimizing retraining time.
Chun-Han Yao, Amit Raj, Wei-Chih Hung +4 authors
ARTIC3D reconstructs 3D articulated shapes from complex monocular images using a skeleton-based model, 2D diffusion priors, and diffusion-guided optimization, achieving higher fidelity and more realistic animations.
Chiyu Max Jiang, Andre Cornman, Cheolho Park +3 authors
MotionDiffuser, a diffusion model for multi-agent trajectory prediction, achieves state-of-the-art results by learning a multimodal distribution with a simple predictor and permutation-invariant joint learning, using PCA for trajectory compression and differentiable cost functions for constrained sampling.
Xiao Han, Yukang Cao, Kai Han +5 authors
HeadSculpt introduces a coarse-to-fine pipeline for generating and editing 3D head avatars with high fidelity and identity preservation using a diffusion model enhanced with 3D awareness and a novel editing score distillation strategy.
Qianqian Dong, Zhiying Huang, Chen Xu +14 authors
PolyVoice, a language model-based framework for speech-to-speech translation, uses discretized speech units and VALL-E X for high-quality voice and style preservation in translation.
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