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。
Ohad Rubin, Jonathan Berant
The Retrieval-Pretrained Transformer jointly trains a retrieval-augmented language model to model long texts by fusing retrieved information for improved prediction.
50 篇论文 · 按点赞排序
Shouyuan Chen, Sherman Wong, Liangjian Chen +1 authors
Position Interpolation extends the context window of RoPE-based LLMs up to 32768 tokens with minimal fine-tuning, maintaining task performance and stability compared to extrapolation methods.
Hugo Laurençon, Lucile Saulnier, Léo Tronchon +9 authors
The OBELICS dataset, containing interleaved image-text documents, is introduced and used to train competitive multimodal models.
Minghua Liu, Chao Xu, Haian Jin +4 authors
A novel feed-forward method generates high-quality 360-degree 3D meshes from single images using a view-conditioned 2D diffusion model and an SDF-based neural surface reconstruction module.
Rishabh Agarwal, Nino Vieillard, Yongchao Zhou +4 authors
Generalized Knowledge Distillation addresses distribution mismatch in auto-regressive sequence models by training the student on its own output sequences and integrates with RL fine-tuning for effective distillation in tasks like summarization and instruction-tuning.
Zhiliang Peng, Wenhui Wang, Li Dong +4 authors
Kosmos-2, a Multimodal Large Language Model, combines text-to-visual grounding with existing MLLM capabilities, enabling tasks like referring expression comprehension and phrase grounding.
Yunpeng Bai, Xintao Wang, Yanpei Cao +3 authors
DreamDiffusion generates high-quality images from EEG signals using pre-trained text-to-image models and CLIP image encoder for improved alignment and robust representations, overcoming noise and data limitations.
William Berrios, Gautam Mittal, Tristan Thrush +2 authors
LENS uses language models to reason over outputs from vision modules, achieving competitive performance in vision and vision-language tasks without multimodal training.
Biao Jiang, Xin Chen, Wen Liu +3 authors
MotionGPT, a unified motion-language model using discrete vector quantization, achieves top performance in various motion-related tasks by treating motion as a language similar to text.
Yuheng Li, Haotian Liu, Yangming Wen +1 authors
A text-to-image diffusion model is enhanced with adapter layers and regionally-guided sampling to achieve controlled generation of personalized objects with high fidelity.
Tung Phung, Victor-Alexandru Pădurean, José Cambronero +5 authors
Systematic evaluation of ChatGPT and GPT-4 in educational scenarios demonstrates GPT-4's superior performance over ChatGPT and identifies areas needing improvement.
Yujun Shi, Chuhui Xue, Jiachun Pan +3 authors
DragDiffusion extends interactive point-based image editing to diffusion models, offering precise spatial control and high-quality editing across diverse scenarios.
Eric Nguyen, Michael Poli, Marjan Faizi +10 authors
HyenaDNA, a genomic foundation model based on implicit convolutions, achieves state-of-the-art performance on various benchmarks with longer context lengths, single nucleotide resolution, and fewer parameters.
Kaiyu Yang, Aidan M. Swope, Alex Gu +6 authors
LeanDojo provides open-source tools, data, and benchmarks to facilitate research on machine learning methods for theorem proving, introducing ReProver, an LLM-based prover that uses retrieval for premise selection.
Gregor Bachmann, Sotiris Anagnostidis, Thomas Hofmann
MLPs achieve competitive performance on vision tasks with large-scale pre-training, challenging the narrative that inductive bias is necessary for high accuracy.
Chaoning Zhang, Dongshen Han, Yu Qiao +4 authors
MobileSAM enhances the Segment Anything Model (SAM) for mobile devices by decoupling the image encoder and mask decoder to create a lightweight version that maintains performance while significantly reducing size and improving speed.
Neel Jain, Khalid Saifullah, Yuxin Wen +6 authors
A self-supervised evaluation framework quantifies large language model behavior without human-curated labels by analyzing text transformations.
Zhenyu Zhang, Ying Sheng, Tianyi Zhou +9 authors
A novel KV cache management technique, Heavy Hitter Oracle (H2O), optimizes memory usage in Large Language Models by prioritizing significant tokens, enhancing throughput and reducing latency.
Xianhang Li, Zeyu Wang, Cihang Xie
CLIPA-v2 extends CLIPA by applying an inverse scaling law during fine-tuning, achieving high zero-shot ImageNet accuracy with reduced computational costs.
Zibo Zhao, Wen Liu, Xin Chen +7 authors
A novel alignment-before-generation approach using SITA-VAE and ASLDM for generating high-quality 3D shapes from 2D images or texts leverages a shape-image-text-aligned space.
Yanzhe Zhang, Ruiyi Zhang, Jiuxiang Gu +4 authors
LLaVAR enhances visual instruction-tuned models by incorporating text recognition from images, resulting in improved performance on text-based VQA datasets and better interaction capabilities.
Alycia Lee, Brando Miranda, Sanmi Koyejo
Task2Vec diversity coefficient is used to assess and validate data diversity in pre-training datasets for Large Language Models (LLMs).
Hariharan Manikandan, Yiding Jiang, J Zico Kolter
Prompt-based large language models can effectively serve as weak learners in boosting algorithms, enhancing performance on tabular data with limited samples compared to traditional methods.
Weizhe Yuan, Kyunghyun Cho, Jason Weston
A framework is proposed to utilize system-level natural language feedback to improve model design and performance through metric design and prompt refinement, with case studies showing its effectiveness in search and dialog generation.
Ayça Takmaz, Elisabetta Fedele, Robert W. Sumner +3 authors
OpenMask3D is a zero-shot approach for open-vocabulary 3D instance segmentation using class-agnostic instance masks and multi-view fusion of CLIP-based image embeddings.
Nadine Chang, Francesco Ferroni, Michael J. Tarr +2 authors
A novel framework generates labeling instructions for datasets using a pre-trained vision and language model, improving the quality and accessibility of dataset annotations.
Jianyuan Wang, Christian Rupprecht, David Novotny
Camera pose estimation using a diffusion framework improves upon traditional SfM methods by incorporating geometric constraints and generalizing across datasets.
Jonathan N. Lee, Annie Xie, Aldo Pacchiano +4 authors
Transformers pretrained with Decision-Pretrained Transformer (DPT) exhibit strong in-context learning capabilities in reinforcement learning, achieving exploration, generalization, and faster learning compared to algorithms used for pretraining.
Jingyu Zhuang, Chen Wang, Lingjie Liu +2 authors
DreamEditor enables controlled editing of neural fields using text prompts through mesh-based representations and score distillation sampling, resulting in realistic and consistent scene modifications.
Joshua Maynez, Priyanka Agrawal, Sebastian Gehrmann
Researchers adapt application-specific generation benchmarks to pre-trained large language models (PLMs) to evaluate their limitations and capabilities across different dimensions like scale, architecture, and language.
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