Exponentially Faster Language Modelling
Peter Belcak, Roger Wattenhofer
FastBERT achieves significant inference speedup over BERT by selectively engaging a small fraction of neurons using fast feedforward networks.
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。
本周最热253 赞
GAIA: a benchmark for General AI Assistants
Grégoire Mialon, Clémentine Fourrier, Craig Swift +3 authors
GAIA benchmarks general AI assistants using real-world questions that challenge both reasoning and multi-modality handling, showcasing a significant gap between human and AI performance.
44 篇论文 · 按点赞排序
Orca 2: Teaching Small Language Models How to Reason
Arindam Mitra, Luciano Del Corro, Shweti Mahajan +12 authors
Orca 2 enhances smaller language models' reasoning abilities by teaching them diverse solution strategies, outperforming larger models on complex reasoning tasks.
Make Pixels Dance: High-Dynamic Video Generation
Yan Zeng, Guoqiang Wei, Jiani Zheng +4 authors
PixelDance, a diffusion model-based approach, generates high-dynamic videos by incorporating image instructions for first and last frames alongside text instructions, surpassing current text-to-video methods in complexity and motion.
FusionFrames: Efficient Architectural Aspects for Text-to-Video Generation Pipeline
Vladimir Arkhipkin, Zein Shaheen, Viacheslav Vasilev +3 authors
A new two-stage latent diffusion model generates videos from text, achieving high quality and efficiency in keyframe synthesis and interpolation.
LucidDreamer: Domain-free Generation of 3D Gaussian Splatting Scenes
Jaeyoung Chung, Suyoung Lee, Hyeongjin Nam +2 authors
LucidDreamer generates domain-free 3D scenes using diffusion-based generative models and Gaussian splats, producing highly detailed results.
Diffusion Model Alignment Using Direct Preference Optimization
Bram Wallace, Meihua Dang, Rafael Rafailov +7 authors
A method called Diffusion-DPO aligns text-to-image diffusion models to human preferences using direct optimization on comparison data, improving visual appeal and prompt alignment.
ZipLoRA: Any Subject in Any Style by Effectively Merging LoRAs
Viraj Shah, Nataniel Ruiz, Forrester Cole +4 authors
ZipLoRA effectively combines independently trained style and subject LoRAs to enhance both subject and style fidelity in generative models.
System 2 Attention (is something you might need too)
Jason Weston, Sainbayar Sukhbaatar
System 2 Attention in Transformer-based Large Language Models improves factual accuracy and reduces bias by refining input context.
GPQA: A Graduate-Level Google-Proof Q&A Benchmark
David Rein, Betty Li Hou, Asa Cooper Stickland +5 authors
A dataset of extremely difficult multiple-choice questions challenges both experts and AI systems, facilitating the development of scalable oversight methods for AI-generated knowledge.
MultiLoRA: Democratizing LoRA for Better Multi-Task Learning
Yiming Wang, Yu Lin, Xiaodong Zeng +1 authors
MultiLoRA improves multi-task adaptation for large language models by reducing the dominance of top singular vectors in LoRA parameter updates through horizontal scaling and modified initialization.
MagicDance: Realistic Human Dance Video Generation with Motions & Facial Expressions Transfer
Di Chang, Yichun Shi, Quankai Gao +6 authors
MagicDance, a diffusion-based model, transfers human motion and facial expression across identities in dance videos, achieving high generalization and robust appearance control.
SuGaR: Surface-Aligned Gaussian Splatting for Efficient 3D Mesh Reconstruction and High-Quality Mesh Rendering
Antoine Guédon, Vincent Lepetit
A method for fast and precise mesh extraction from 3D Gaussian Splatting using a regularization term and Poisson reconstruction, enabling real-time editing and rendering quality.
HierSpeech++: Bridging the Gap between Semantic and Acoustic Representation of Speech by Hierarchical Variational Inference for Zero-shot Speech Synthesis
Sang-Hoon Lee, Ha-Yeong Choi, Seung-Bin Kim +1 authors
HierSpeech++, an enhanced hierarchical variational autoencoder, achieves high-quality zero-shot speech synthesis by combining self-supervised representations and an efficient super-resolution framework.
Using Human Feedback to Fine-tune Diffusion Models without Any Reward Model
Kai Yang, Jian Tao, Jiafei Lyu +6 authors
D3PO method directly fine-tunes diffusion models using human feedback without requiring a separate reward model, reducing computational cost and improving image quality.
Adapters: A Unified Library for Parameter-Efficient and Modular Transfer Learning
Clifton Poth, Hannah Sterz, Indraneil Paul +7 authors
Adapters is an open-source library designed for parameter-efficient and modular transfer learning in large language models, offering a unified interface for various adapter methods.
Video-LLaVA: Learning United Visual Representation by Alignment Before Projection
Bin Lin, Bin Zhu, Yang Ye +3 authors
Video-LLaVA is a unified large vision-language model that enhances performance across various image and video benchmarks by integrating visual representations into the language feature space.
Text-to-Sticker: Style Tailoring Latent Diffusion Models for Human Expression
Animesh Sinha, Bo Sun, Anmol Kalia +14 authors
Style Tailoring is a finetuning method for Latent Diffusion Models that enhances the generation of sticker images, improving visual quality, prompt alignment, and scene diversity.
NeuroPrompts: An Adaptive Framework to Optimize Prompts for Text-to-Image Generation
Shachar Rosenman, Vasudev Lal, Phillip Howard
An adaptive framework automatically enhances text-to-image prompts to improve generation quality using a pre-trained language model and constrained text decoding.
Rethinking Attention: Exploring Shallow Feed-Forward Neural Networks as an Alternative to Attention Layers in Transformers
Vukasin Bozic, Danilo Dordervic, Daniele Coppola +1 authors
Shallow feed-forward networks can emulate the performance of the attention mechanism in Transformers, as demonstrated by their competitive results on sequence-to-sequence tasks using knowledge distillation.
Emu Video: Factorizing Text-to-Video Generation by Explicit Image Conditioning
Rohit Girdhar, Mannat Singh, Andrew Brown +7 authors
A text-to-video generation model that factors video creation into two steps—image generation followed by video generation—achieves superior quality and resolution without deep model cascades.
Concept Sliders: LoRA Adaptors for Precise Control in Diffusion Models
Rohit Gandikota, Joanna Materzynska, Tingrui Zhou +2 authors
Concept sliders enable precise control over image generation attributes using diffusion models by identifying low-rank parameter directions with minimal interference.
PhysGaussian: Physics-Integrated 3D Gaussians for Generative Dynamics
Tianyi Xie, Zeshun Zong, Yuxin Qiu +4 authors
PhysGaussian integrates Newtonian dynamics within 3D Gaussians for high-quality motion synthesis using a Material Point Method (MPM) that aligns with continuum mechanics principles, eliminating the need for traditional meshing techniques.
LucidDreamer: Towards High-Fidelity Text-to-3D Generation via Interval Score Matching
Yixun Liang, Xin Yang, Jiantao Lin +3 authors
Interval Score Matching improves text-to-3D generation by addressing Score Distillation Sampling's over-smoothing issue and using 3D Gaussian Splatting, achieving superior quality and efficiency.
PF-LRM: Pose-Free Large Reconstruction Model for Joint Pose and Shape Prediction
Peng Wang, Hao Tan, Sai Bi +6 authors
A Pose-Free Large Reconstruction Model using self-attention for 3D reconstruction from few unposed images, estimating camera poses, and generalizing across datasets.
Camels in a Changing Climate: Enhancing LM Adaptation with Tulu 2
Hamish Ivison, Yizhong Wang, Valentina Pyatkin +8 authors
T\"ULU 2, an advanced suite of language models, achieves state-of-the-art performance through improved datasets, fine-tuning techniques, and direct preference optimization, surpassing even GPT-3.5-turbo-0301 on several benchmarks.
Visual In-Context Prompting
Feng Li, Qing Jiang, Hao Zhang +9 authors
A universal visual in-context prompting framework enhances zero-shot capabilities for both referring and generic vision tasks by using a versatile prompt encoder and reference image segments as context.
PG-Video-LLaVA: Pixel Grounding Large Video-Language Models
Shehan Munasinghe, Rusiru Thushara, Muhammad Maaz +4 authors
Video-LLaVA integrates audio transcriptions and uses a novel grounding module to enhance pixel-level grounding in large multimodal video models, leading to improved performance in generative and question-answering tasks.
Memory Augmented Language Models through Mixture of Word Experts
Cicero Nogueira dos Santos, James Lee-Thorp, Isaac Noble +2 authors
A Mixture of Word Experts (MoWE) model demonstrates superior performance in NLP tasks compared to T5 models and traditional MoE models, using memory-augmented techniques and knowledge-rich routing functions.
SelfEval: Leveraging the discriminative nature of generative models for evaluation
Sai Saketh Rambhatla, Ishan Misra
SelfEval automates the assessment of text-image understanding in generative models by computing image likelihoods given text prompts, showing competitive performance to discriminative models on fine-grained tasks.