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。
Wenqi Zhang, Hang Zhang, Xin Li +6 authors
A new high-quality instructional video-based corpus improves VLM pretraining through more coherent context and richer knowledge, enhancing performance in knowledge-intensive tasks.
48 篇论文 · 按点赞排序
Junying Chen, Zhenyang Cai, Ke Ji +5 authors
The proposed approach enhances medical reasoning in large language models using verifiable problems, guided search, and reinforcement learning to achieve superior performance in medical problem-solving.
Qiushi Sun, Kanzhi Cheng, Zichen Ding +12 authors
OS-Genesis is a novel GUI data synthesis pipeline that enhances the training of GUI agents by reversing the trajectory collection process to improve data quality and diversity.
Chenglin Yang, Celong Liu, Xueqing Deng +4 authors
1.58-bit quantization of the FLUX.1-dev text-to-image model achieves comparable generation quality with reduced storage and inference overhead using self-supervised methods.
Yang Shen, Xiu-Shen Wei, Yifan Sun +6 authors
An auto-regressive vision-language model trained with explanatory instructions achieves instruction-level zero-shot capabilities and strong zero-shot generalization in computer vision tasks.
Liang Chen, Zekun Wang, Shuhuai Ren +24 authors
A comprehensive taxonomy unifies multimodal understanding and generation through Next Token Prediction, covering aspects like tokenization, model architectures, task representation, datasets, and open challenges.
Yuanpeng Tu, Hao Luo, Xi Chen +3 authors
VideoAnydoor is a zero-shot framework for object insertion in videos that preserves detail and controls motion using a text-to-video model, ID extractor, box sequence, pixel warper, and diffusion U-Net.
Yoav HaCohen, Nisan Chiprut, Benny Brazowski +13 authors
LTX-Video, a transformer-based latent diffusion model, integrates Video-VAE and denoising transformer for efficient high-resolution video generation with temporal consistency.
Shanghaoran Quan, Jiaxi Yang, Bowen Yu +14 authors
CodeElo is a standardized code generation benchmark based on CodeForces contests, providing comprehensive evaluation of LLMs' coding abilities with reliable Elo ratings.
Yuqian Yuan, Hang Zhang, Wentong Li +9 authors
The VideoRefer Suite enhances Video LLMs for finer spatial-temporal understanding through a comprehensive dataset, model, and benchmark.
Jingfeng Yao, Xinggang Wang
VA-VAE aligns latent spaces with pre-trained vision foundation models, enhancing DiT's training efficiency and performance on high-fidelity image generation.
Zhenyang Cai, Junying Chen, Rongsheng Wang +6 authors
Compositional generalization aids multimodal large language models in understanding and generalizing to unseen medical images across various domains and tasks.
Ohad Rahamim, Ori Malca, Dvir Samuel +1 authors
A method for animating 3D objects using text prompts to guide 4D generation, employing Neural Radiance Fields and diffusion models, achieves high identity preservation and visual quality.
Xingyu Chen, Jiahao Xu, Tian Liang +11 authors
The paper addresses overthinking in models like OpenAI o1 by introducing efficiency metrics and strategies to reduce computational overhead without sacrificing accuracy.
Yichao Fu, Junda Chen, Siqi Zhu +4 authors
Dynasor optimizes compute resources for large language model inference by dynamically adjusting allocation based on query difficulty and reasoning progress.
Zhenting Wang, Shuming Hu, Shiyu Zhao +12 authors
A method using Multimodal Large Language Models for zero-shot detection of unsafe images improves upon human-labeled fine-tuning by objectifying safety rules and using cascaded reasoning for complex judgments.
Wang Qun, Liu Yang, Lin Qingquan +2 authors
Xmodel-2, a 1.2-billion-parameter large language model, utilizes a unified set of hyperparameters across different scales and the WSD learning rate scheduler to achieve state-of-the-art performance in complex reasoning tasks with efficient training.
Xiaoshuai Song, Yanan Wu, Weixun Wang +3 authors
Program-driven Self-Correction (ProgCo) improves large language models' self-verification and refinement by using self-generated and self-executing verification pseudo-programs, particularly in complex reasoning tasks.
Jiayi Pan, Xingyao Wang, Graham Neubig +4 authors
An environment for training software engineering agents using real-world Python tasks and language models achieves state-of-the-art performance on SWE-Bench.
Mahir Labib Dihan, Md Tanvir Hassan, Md Tanvir Parvez +5 authors
MapEval assesses foundation models' geo-spatial reasoning and map-based query handling, revealing performance gaps and suggesting improvements in spatial understanding.
Yuxiang Chai, Hanhao Li, Jiayu Zhang +5 authors
Android Agent Arena (A3) provides an evaluation platform for mobile GUI agents with real-world tasks, flexible action spaces, and automated LLM-based testing.
Chia-Yu Hung, Navonil Majumder, Zhifeng Kong +4 authors
TangoFlux is an efficient Text-to-Audio (TTA) generative model that uses CLAP-Ranked Preference Optimization (CRPO) to enhance audio alignment and achieve state-of-the-art performance.
Zehan Wang, Ziang Zhang, Tianyu Pang +3 authors
Orient Anything model estimates object orientation from single images using 3D-world knowledge and achieves state-of-the-art accuracy with zero-shot abilities across various applications.
Jihoon Chung, Tyler Zhu, Max Gonzalez Saez-Diez +3 authors
MERV, a Multi-Encoder Representation of Videos, enhances VideoLLMs by using multiple visual encoders for better visual information processing, accuracy, and faster training.
Qingyan Bai, Hao Ouyang, Yinghao Xu +5 authors
Edicho, a training-free diffusion-based editing method, uses explicit image correspondence and a classifier-free guidance denoising strategy for consistent editing across diverse images.
Sangyun Chung, Youngjoon Yu, Youngchae Chee +3 authors
A benchmark and optimization method are proposed to enhance(VLMs) ability to reason with multi-vision sensor data, improving their performance on tasks requiring sensor-specific understanding.
Hyunsoo Cha, Inhee Lee, Hanbyul Joo
PERSE generates animatable personalized avatars from portraits using 3D Gaussian Splatting to enable continuous and disentangled facial attribute editing.
Ziang Yan, Zhilin Li, Yinan He +9 authors
Task Preference Optimization (TPO) enhances multimodal large language models with fine-grained visual tasks by using learnable task tokens and multi-task co-training, leading to significant performance improvements and robust zero-shot capabilities.
Mikhail Tikhomirov, Daniil Chernyshev
Learned Embedding Propagation (LEP) enables cost-effective adaptation of large language models for new languages by directly implanting knowledge without comprehensive instruction-tuning.
Yujie Luo, Xiangyuan Ru, Kangwei Liu +10 authors
OneKE is a dockerized schema-guided knowledge extraction system that extracts knowledge from the Web and PDFs using multiple agents and a configurable knowledge base, effective across various domains.
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