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。
42 篇论文 · 按点赞排序
Yan Ma, Weicong Liang, Yiduo Hao +4 authors
Thorough evaluation reveals that synthetic pre-training datasets derived from recent captioning and generative models significantly improve the performance of DETR-based object detection approaches.
Bang An, Sicheng Zhu, Michael-Andrei Panaitescu-Liess +2 authors
A two-step zero-shot classification method, named PerceptionCLIP, improves CLIP's performance by inferring contextual attributes from images before classification.
Yutong Lin, Yuhui Yuan, Zheng Zhang +3 authors
A plain DETR detector with global cross-attention and box-to-pixel relative position bias achieves competitive performance using masked image modeling and a single-scale feature map.
Rohith Kuditipudi, John Thickstun, Tatsunori Hashimoto +1 authors
The paper presents a method to embed watermarks in text generated by autoregressive language models that can withstand perturbations while preserving the text distribution, validated across multiple models and various types of attacks.
Zexin Chen, Eric Zhou, Kenneth Eaton +2 authors
Large language models generate stories used to guide a robotic agent in imaginary play, which is evaluated in a simulated text adventure game.
Stephen Kelly, Daniel S. Park, Xingyou Song +8 authors
AutoRobotics-Zero utilizes AutoML-Zero to develop adaptable control policies for robots that can modify their algorithms on-the-fly, surpassing neural network baselines in robustness to sudden environmental changes.
Jun Yan, Vikas Yadav, Shiyang Li +6 authors
Virtual Prompt Injection allows attackers to subtly manipulate the behavior of Large Language Models in response to specific instructions, highlighting the need for robust defenses against data poisoning.
Bohao Li, Rui Wang, Guangzhi Wang +3 authors
A benchmark named SEED-Bench evaluates generative comprehension in Multimodal Large Language Models (MLLMs) across multiple dimensions using 19K human-annotated multiple-choice questions.
Shihao Liang, Kunlun Zhu, Runchu Tian +6 authors
Unified Instruction Tuning improves instruction tuning performance by addressing format inconsistencies across datasets using automatic format transfer and denoising techniques.
Xiangyu Wang, Jingsen Zhu, Qi Ye +4 authors
Seal-3D enables interactive and free-form pixel-level editing of NeRF models with instant preview using a proxy function and teacher-student training strategy.
Eric Tabellion, Nikhil Karnad, Noa Glaser +3 authors
A smartphone app automatically captures and composites foreground or background blur effects using computational burst photography techniques.
Qi Yang, Joel Jung, Timon Deschamps +2 authors
The Tencent-Dynamic Colored Mesh Database (TDMD) examines the impact of various distortions on DCM perception using objective metrics and subjective experiments.
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