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。
Albert Q. Jiang, Alexandre Sablayrolles, Antoine Roux +23 authors
Mixtral 8x7B, a Sparse Mixture of Experts language model, achieves superior performance across benchmarks by using a selective architecture that leverages fewer active parameters.
50 篇论文 · 按点赞排序
Maciej Pióro, Kamil Ciebiera, Krystian Król +2 authors
A combination of Mixture of Experts (MoE) with State Space Models (SSMs) improves model efficiency and performance compared to both standalone SSMs and MoE-enhanced Transformers.
Lichao Sun, Yue Huang, Haoran Wang +64 authors
This study assesses the trustworthiness of large language models across various dimensions, including truthfulness, safety, fairness, robustness, privacy, and machine ethics, finding a positive correlation with utility and highlighting differences between proprietary and open-source models.
Damai Dai, Chengqi Deng, Chenggang Zhao +14 authors
The DeepSeekMoE architecture improves expert specialization in Mixture-of-Experts models by segmenting experts and isolating shared ones, achieving better performance and computational efficiency compared to GShard and other models.
DeepSeek-AI, Xiao Bi, Deli Chen +84 authors
DeepSeek LLM, an open-source language model project, develops a large dataset and employs SFT and DPO to achieve performance surpassing LLaMA-2 70B and GPT-3.5 in various benchmarks and open-ended evaluations.
Xiaoding Lu, Adian Liusie, Vyas Raina +2 authors
A combination of smaller chat AI models, through a method called blending, can achieve performance comparable to or better than large single models with lower computational resources.
Moab Arar, Andrey Voynov, Amir Hertz +5 authors
Prompt-aligned personalization improves text alignment and enhances personalized image creation with complex prompts by using score distillation sampling.
Weimin Wang, Jiawei Liu, Zhijie Lin +9 authors
MagicVideo-V2 generates high-fidelity and smooth videos from text using an integrated pipeline that includes text-to-image, video motion generation, and frame interpolation modules, outperforming existing systems in user evaluations.
Junsong Chen, Yue Wu, Simian Luo +5 authors
PIXART-{\delta} integrates Latent Consistency Model (LCM) and ControlNet into PIXART-{\alpha} for accelerated, high-quality text-to-image synthesis with efficient training and inference.
Alon Ziv, Itai Gat, Gael Le Lan +6 authors
MAGNeT is a single-stage, non-autoregressive transformer for audio generation that uses masked spans during training, a rescoring method, and a hybrid approach to achieve fast generation.
Matanel Oren, Michael Hassid, Yossi Adi +1 authors
Decoder-only transformers can be conceptualized as finite multi-state RNNs, and a new cache compression technique, TOVA, significantly reduces their computational cost while maintaining high performance.
Dongsheng Wang, Zhiqiang Ma, Armineh Nourbakhsh +2 authors
A new framework, DocGraphLM, combines pre-trained language models with graph semantics to improve information extraction and question answering on documents with complex layouts.
Jiawei Yang, Katie Z Luo, Jiefeng Li +3 authors
A novel noise model is proposed to eliminate grid-like artifacts in Vision Transformers by decomposing outputs into noise-free and artifact-related components, enhancing performance in downstream tasks.
Evan Hubinger, Carson Denison, Jesse Mu +36 authors
Larger language models can exhibit and maintain deceptive behavior even when trained with standard safety techniques, which can sometimes make the behavior harder to detect.
Peitian Zhang, Zheng Liu, Shitao Xiao +3 authors
Activation Beacon enhances large language models' ability to process long contexts using compact activation forms, maintaining original capabilities and improving efficiency.
Mohamad Shahbazi, Liesbeth Claessens, Michael Niemeyer +4 authors
InseRF inserts new 3D objects into scenes based on textual descriptions and 2D bounding boxes, using diffusion models for text-to-image generation and monocular depth estimation.
Zhen Qin, Weigao Sun, Dong Li +3 authors
Lightning Attention-2 efficiently processes sequences of unlimited length using a tiling technique, achieving consistent speed and outperforming conventional attention mechanisms.
Binghai Wang, Rui Zheng, Lu Chen +24 authors
Methods are introduced to enhance reward models in Reinforcement Learning from Human Feedback by addressing data quality and generalization issues using voting mechanisms, contrastive learning, and meta-learning.
Yatharth Gupta, Vishnu V. Jaddipal, Harish Prabhala +2 authors
Two scaled-down variants of Stable Diffusion XL (SDXL), achieved through layer-level loss-driven pruning, maintain competitive image quality with reduced computational demands.
Zhaoxi Chen, Gyeongsik Moon, Kaiwen Guo +20 authors
URHand is a universal relightable hand model that generalizes across different identities, viewpoints, and illuminations, using a spatially varying linear lighting model and joint learning of a physically based model for high-quality, real-time rendering.
Zilong Wang, Hao Zhang, Chun-Liang Li +8 authors
The Chain-of-Table framework integrates tabular data into the reasoning chain of large language models (LLMs) to improve performance in table-based question answering and fact verification tasks.
Linus Franke, Darius Rückert, Laura Fink +1 authors
TRIPS, combining Gaussian Splatting and ADOP techniques, achieves high-quality rendering of detailed scenes with real-time performance using a screen-space image pyramid and a differentiable render pipeline.
Seung Hyun Lee, Yinxiao Li, Junjie Ke +11 authors
Parrot, a multi-reward RL framework, enhances text-to-image generation by automatically balancing multiple quality rewards and generating quality-aware text prompts without forgetting the original user input.
Haobo Yuan, Xiangtai Li, Chong Zhou +3 authors
A unified framework combining SAM and CLIP through knowledge transfer improves performance in both segmentation and recognition tasks.
Tong Wu, Guandao Yang, Zhibing Li +5 authors
An automatic evaluation metric for text-to-3D generative models uses GPT-4V to generate prompts and perform pairwise comparisons for assigning Elo ratings, aligning with human preferences.
Pratyush Maini, Zhili Feng, Avi Schwarzschild +2 authors
TOFU, a benchmark for evaluating unlearning in large language models, uses synthetic data to measure how effectively unlearning methods can remove specific information from the model.
Asma Ghandeharioun, Avi Caciularu, Adam Pearce +2 authors
Patchscopes is a framework that uses LLMs to explain their internal representations, unifying and improving upon prior interpretability methods and enabling new applications like self-correction.
Xiaojuan Wang, Taesung Park, Yang Zhou +2 authors
A framework for smoothing jump cuts in talking head videos using DensePose keypoints, face landmarks, and cross-modal attention for seamless transitions.
Tao Tu, Anil Palepu, Mike Schaekermann +22 authors
An AI system named AMIE, designed with a large language model optimized for diagnostic dialogue, demonstrated better performance than primary care physicians in several clinical axes through a simulated environment and textual consultations with patient actors.
Yue Zhao, Long Zhao, Xingyi Zhou +9 authors
A video-language model, fine-tuned from an image-language baseline with synthesized data, outperforms existing methods on various benchmarks by leveraging auto-generated captions for better textual supervision.
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