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
Junhao Zhuang, Lingen Li, Xuan Ju +3 authors
Cobra, a method utilizing a Causal Sparse DiT architecture, achieves accurate, efficient, and consistent line art colorization using extensive reference images and color hints.
Tao Zhang, Xiangtai Li, Zilong Huang +6 authors
Pixel-SAIL is a single transformer model that performs pixel-level understanding tasks without additional components, achieving comparable results through a simplified pipeline with three technical improvements.
Ahmed Masry, Mohammed Saidul Islam, Mahir Ahmed +11 authors
ChartQAPro introduces a diverse benchmark for evaluating the performance of large vision-language models in chart reasoning, revealing significant performance gaps compared to existing benchmarks.
Tianhui Song, Weixin Feng, Shuai Wang +4 authors
The paper introduces score distillation based model merging for text-to-image generation, enabling a single versatile model to produce images in different styles based on style vectors.
Xingjian Leng, Jaskirat Singh, Yunzhong Hou +3 authors
End-to-end training of latent diffusion models and VAE tokenizers using representation-alignment loss enhances diffusion model training speed and VAE performance.
Wenyuan Zhang, Shuaiyi Nie, Xinghua Zhang +2 authors
S1-Bench evaluates the efficiency of Large Reasoning Models in simple tasks requiring intuitive thinking, revealing significant inefficiencies and a tendency for unnecessary deliberation.
Jang Hyun Cho, Andrea Madotto, Effrosyni Mavroudi +26 authors
A fully transparent Perception Language Model (PLM) for image and video understanding is developed without relying on proprietary models, using large-scale synthetic and human-labeled data and introducing PLM-VideoBench for evaluation.
Sicheng Feng, Gongfan Fang, Xinyin Ma +1 authors
The survey discusses methods to accelerate reasoning models by compressing Chain-of-Thoughts, developing compact models, and designing efficient decoding strategies.
Yutaro Yamada, Robert Tjarko Lange, Cong Lu +5 authors
The AI Scientist-v2 autonomously proposes hypotheses, performs experiments, analyzes data, and writes peer-reviewed scientific papers, marking the first fully AI-generated paper accepted by a conference.
Haojian Huang, Haodong Chen, Shengqiong Wu +5 authors
VistaDPO, a new framework, enhances text-video preference alignment through hierarchical spatial-temporal optimization, addressing misalignment and hallucination in large video models.
Ian Magnusson, Nguyen Tai, Ben Bogin +10 authors
Small-scale experiments accurately predict the best large language models when using continuous likelihood metrics as proxies.
Wei Xiong, Jiarui Yao, Yuhui Xu +8 authors
Reinforce-Rej, a minimal extension of policy gradient, outperforms GRPO and PPO in fine-tuning large language models by filtering out both entirely incorrect and correct samples, improving efficiency and stability.
Yanrui Bin, Wenbo Hu, Haoyuan Wang +2 authors
A novel method, NormalCrafter, enhances temporal coherence in video-based surface normal estimation by utilizing video diffusion models with Semantic Feature Regularization and two-stage training in latent and pixel spaces.
Chenghao Xiao, Isaac Chung, Imene Kerboua +7 authors
The Massive Image Embedding Benchmark (MIEB) evaluates image and image-text embedding models across various tasks, revealing hidden capabilities and performance correlations with multimodal large language models.
Xiangyan Liu, Jinjie Ni, Zijian Wu +5 authors
NoisyRollout, an RL approach that introduces targeted diversity through noise in image trajectories, enhances VLM policy exploration without additional training cost, achieving state-of-the-art performance on out-of-domain benchmarks.
Yangshen Deng, Zhengxin You, Long Xiang +13 authors
AlayaDB is a vector database system that decouples KV cache and attention computation from LLMs, optimizing inference for higher quality and reduced hardware usage.
Zhenting Wang, Guofeng Cui, Kun Wan +1 authors
A distribution-level curriculum learning framework using UCB for RL-based LLM post-training enhances convergence speed and performance by dynamically adjusting training schedules across diverse data distributions.
Yunxiang Zhang, Muhammad Khalifa, Shitanshu Bhushan +6 authors
MLRC-Bench evaluates large language model agents in tackling novel machine learning research competitions using rigorous protocols and objective metrics, highlighting significant challenges and misalignments compared to previous benchmarks.
Yongsheng Yu, Haitian Zheng, Zhifei Zhang +7 authors
ZipIR, a novel framework, uses a compressed latent space and Latent Pyramid VAE to enhance efficiency and scalability of diffusion models for high-resolution image restoration.
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