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
Yuchuan Tian, Yuchen Liang, Jiacheng Sun +10 authors
Adapting autoregressive models to block-wise diffusion enables parallel generation and retains pretrained knowledge, achieving state-of-the-art performance among 7B-class diffusion language models.
Mingzhi Chen, Taiming Lu, Jiachen Zhu +2 authors
Derf, a novel point-wise normalization function, outperforms existing alternatives across various domains, enhancing generalization without increased fitting capacity.
Chuhan Zhang, Guillaume Le Moing, Skanda Koppula +11 authors
D4RT, a unified transformer-based model, efficiently reconstructs 4D scenes from videos by querying 3D positions in space-time, outperforming previous methods.
Thao Nguyen, Sicheng Mo, Krishna Kumar Singh +6 authors
Vision-Language models fine-tuned on anonymized image captions can capture relational similarity between images, a capability lacking in current visual similarity metrics.
Yuan Gao, Chen Chen, Tianrong Chen +1 authors
FAE, a framework using a feature auto-encoder and dual decoders, adapts pre-trained visual representations for generative models, achieving high performance in image generation tasks.
Meituan LongCat Team, Hanghang Ma, Haoxian Tan +10 authors
LongCat-Image is a bilingual open-source foundation model for image generation that addresses multilingual text rendering, photorealism, and deployment efficiency through rigorous data curation, compact design, and comprehensive open-source support.
Elisabetta Fedele, Francis Engelmann, Ian Huang +3 authors
SpaceControl enables explicit spatial control of 3D generation using various geometric inputs, outperforming existing methods in geometric faithfulness while maintaining visual quality.
Bowen Ping, Chengyou Jia, Minnan Luo +4 authors
PaCo-RL combines reinforcement learning with a specialized consistency reward model and an efficient optimization strategy to improve consistent image generation.
Hongyuan Tao, Bencheng Liao, Shaoyu Chen +4 authors
InfiniteVL, a linear-complexity VLM architecture combining sliding window attention and Gated DeltaNet, achieves competitive performance with less data and faster inference than leading Transformer-based models.
Wenxi Yang, Yuzhong Zhao, Fang Wan +1 authors
sCoT, a language-only CoT paradigm with self-calling subagents, enhances visual reasoning performance and efficiency through group-relative policy optimization.
Meng Wei, Chenyang Wan, Jiaqi Peng +8 authors
DualVLN integrates high-level reasoning and low-level action execution to improve vision-language navigation in dynamic environments, achieving robust real-time control and long-horizon planning.
Wenhao Yan, Sheng Ye, Zhuoyi Yang +6 authors
SCAIL framework improves character animation by using a novel 3D pose representation and a diffusion-transformer architecture with full-context pose injection, achieving studio-grade quality and realism.
Long Lian, Sida Wang, Felix Juefei-Xu +7 authors
ThreadWeaver, a framework for adaptive parallel reasoning, achieves accuracy comparable to sequential models while reducing inference latency through parallel trajectory generation, trie-based training-inference co-design, and parallelization-aware reinforcement learning.
Gyeongjin Kang, Seungkwon Yang, Seungtae Nam +3 authors
MVP, a scalable multi-view transformer architecture, efficiently reconstructs large 3D scenes from multiple images using dual hierarchies and achieves state-of-the-art quality.
Inna Wanyin Lin, Yushi Hu, Shuyue Stella Li +5 authors
A framework for self-training a Vision-Language Model (VLM) judge using self-synthesized data improves judge accuracy on VL-RewardBench, surpassing larger models in several dimensions.
Gemini Robotics Team, Coline Devin, Yilun Du +18 authors
A generative evaluation system using a frontier video model (Veo) enables comprehensive policy evaluation in robotics, including nominal performance, out-of-distribution generalization, and safety checks.
Yiwu Zhong, Zi-Yuan Hu, Yin Li +1 authors
Efficient video reasoning can be achieved using concise chains of thought and reduced visual tokens without manual annotations or supervised fine-tuning.
Xiaojun Jia, Jie Liao, Qi Guo +11 authors
OmniSafeBench-MM is a comprehensive tool for evaluating multi-modal jailbreak attacks and defenses, covering various attack methods, defense strategies, and risk domains.
Yexin Liu, Manyuan Zhang, Yueze Wang +8 authors
OpenSubject, a large-scale video-derived dataset, enhances subject-driven image generation and manipulation through a four-stage pipeline that maintains identity fidelity and handles complex scenes.
Monishwaran Maheswaran, Rishabh Tiwari, Yuezhou Hu +8 authors
Arbitrage is a dynamic routing framework for speculative decoding that improves efficiency in large language model inference by predicting when the target model will provide a better step, outperforming traditional step-level methods.
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