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
Pierre Harvey Richemond, Yunhao Tang, Daniel Guo +15 authors
Direct Reward Optimisation (DRO) framework uses single-trajectory datasets for large language model alignment, offering a simpler and effective alternative to preference-based methods.
Shizun Wang, Xingyi Yang, Qiuhong Shen +2 authors
GFlow reconstructs 4D dynamic scenes and相机 poses from a single monocular video using Gaussian splatting and 2D priors, enabling novel view rendering and object tracking.
Feng Liang, Akio Kodaira, Chenfeng Xu +3 authors
StreamV2V, a diffusion model for real-time video-to-video translation with user prompts, uses a streaming approach and a feature bank to fuse past frame information, achieving high frame rates and temporal consistency.
Huy V. Vo, Vasil Khalidov, Timothée Darcet +10 authors
A clustering-based approach for automatic curation of high-quality datasets for self-supervised pre-training yields features that outperform those from uncurated data and match manually curated data.
Zijin Gu, Tatiana Likhomanenko, He Bai +3 authors
Denoising language models (DLMs) trained with synthetic data achieve state-of-the-art ASR performance on Librispeech without external audio data, surpassing conventional language models and some self-supervised methods.
Minghan Li, Xilun Chen, Ari Holtzman +4 authors
Nearest Neighbor Speculative Decoding enhances language model generation quality, attribution, and speed through token-level retrieval and speculative decoding.
Roy Miles, Pradyumna Reddy, Ismail Elezi +1 authors
The proposed algorithm reduces memory usage during the training and fine-tuning of large language models by compressing intermediate activations without affecting performance.
Yikai Wang, Xinzhou Wang, Zilong Chen +3 authors
Vidu4D, a novel reconstruction model using Dynamic Gaussian Surfels, accurately reconstructs 4D representations from single generated videos, addressing non-rigidity and frame distortion for high-fidelity virtual contents.
Ce Ge, Zhijian Ma, Daoyuan Chen +2 authors
Research proposes BiMix scalability law to model bivariate data scaling behaviors in language models, leading to efficient data curation and improved performance.
Chen Wang, Jiatao Gu, Xiaoxiao Long +2 authors
GECO, a two-stage generative model using score distillation, achieves efficient high-quality 3D generation from images by addressing view inconsistency.
Vicky Zayats, Peter Chen, Melissa Merrari +1 authors
Zipper, a multi-tower decoder using cross-attention, integrates unimodal generative models for effective cross-modal tasks with limited aligned data while maintaining unimodal performance.
Sirui Xie, Zhisheng Xiao, Diederik P Kingma +6 authors
EM Distillation stabilizes the distillation of diffusion models into one-step generators, enhancing perceptual quality and outperforming existing methods in FID scores and text-to-image tasks.
Anran Liu, Cheng Lin, Yuan Liu +5 authors
Part123 framework uses diffusion models and contrastive learning within neural rendering to produce part-aware 3D reconstructions from single-view images with high-quality segmentation.
Zhengfei Kuang, Shengqu Cai, Hao He +4 authors
Collaborative Video Diffusion introduces a cross-video synchronization module that enhances consistency in multi-video generation from different camera trajectories.
Zachary Novack, Julian McAuley, Taylor Berg-Kirkpatrick +1 authors
A new method, DITTO-2, accelerates diffusion model inference for real-time music generation while enhancing control and quality, including text adherence.
Muyao Niu, Xiaodong Cun, Xintao Wang +3 authors
MOFA-Video generates videos from images using controllable signals, combining domain-aware motion field adapters to improve motion consistency and stability in diffusion-based video generation.
Jiaqi Xu, Xinyi Zou, Kunzhe Huang +5 authors
EasyAnimate leverages transformer architecture to generate 3D videos with consistent motion and varying frame rates and resolutions using a motion module and slice VAE.
Xinyu Zhou, Boris Knyazev, Alexia Jolicoeur-Martineau +1 authors
LoGAH, a low-rank parameter decoder Graph HyperNetwork, efficiently initializes large neural networks with better performance compared to random or existing hypernetworks.
Jia Li, Ge Li, Yunfei Zhao +15 authors
How to evaluate the coding abilities of Large Language Models (LLMs) remains an open question. We find that existing benchmarks are poorly aligned with real-world code repositories and are insufficient to evaluate the coding abilities of LLMs. To address the knowledge gap, we propose a new benchmark named DevEval, which has three advances. (1) DevEval aligns with real-world repositories in multiple dimensions, e.g., code distributions and dependency distributions. (2) DevEval is annotated by 13 developers and contains comprehensive annotations (e.g., requirements, original repositories, reference code, and reference dependencies). (3) DevEval comprises 1,874 testing samples from 117 repositories, covering 10 popular domains (e.g., Internet, Database). Based on DevEval, we propose repository-level code generation and evaluate 8 popular LLMs on DevEval (e.g., gpt-4, gpt-3.5, StarCoder 2, DeepSeek Coder, CodeLLaMa). Our experiments reveal these LLMs' coding abilities in real-world code repositories. For example, in our experiments, the highest Pass@1 of gpt-4-turbo is only 53.04%. We also analyze LLMs' failed cases and summarize their shortcomings. We hope DevEval can facilitate the development of LLMs in real code repositories. DevEval, prompts, and LLMs' predictions have been released.
Qiaowei Miao, Yawei Luo, Yi Yang
PLA4D advances text-to-4D synthesis by improving texture, geometry, and motion alignment through novel pixel-level techniques.
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