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
Xiao Pu, Michael Saxon, Wenyue Hua +1 authors
THOUGHTTERMINATOR, a training-free decoding technique, improves the calibration of reasoning models by optimizing token usage across easy and difficult problems.
Ziwen Xu, Shuxun Wang, Kewei Xu +7 authors
EasyEdit2 is a user-friendly framework for adjusting Large Language Model behaviors through test-time steering vectors without parameter modifications.
Chun-Hsiao Yeh, Chenyu Wang, Shengbang Tong +7 authors
All-Angles Bench evaluates the multi-view reasoning capabilities of MLLMs using diverse human-annotated datasets, revealing significant gaps in cross-view correspondence and camera pose estimation.
Lutao Jiang, Jiantao Lin, Kanghao Chen +6 authors
DiMeR, a disentangled dual-stream feed-forward model, improves sparse-view 3D mesh reconstruction using normal maps for geometry and RGB images for texture, outperforming previous methods by over 30% in Chamfer Distance.
Chong Mou, Yanze Wu, Wenxu Wu +12 authors
DreamO, an image customization framework using diffusion transformers, supports various tasks and integrates multiple conditions through a unified approach.
Xu Ma, Peize Sun, Haoyu Ma +22 authors
A Token-Shuffle method enhances high-resolution AR image generation by reducing token numbers in Transformers, enabling efficient generation and outperforming diffusion models.
Fengze Liu, Weidong Zhou, Binbin Liu +7 authors
A unified data selection framework called QuaDMix optimizes the distribution of training data for large language models by balancing quality and diversity, leading to improved performance.
Cailin Zhuang, Yaoqi Hu, Xuanyang Zhang +7 authors
StyleMe3D enhances 3D Gaussian Splatting for stylized scene reconstruction by integrating multi-modal style conditioning, semantic alignment, and perceptual quality enhancement.
Siyu Zhou, Tianyi Zhou, Yijun Yang +4 authors
World alignment and model-based agent WALL-E 2.0 enhance LLMs' performance as world models in new environments by integrating symbolic knowledge and efficient planning.
Thomas Schmied, Jörg Bornschein, Jordi Grau-Moya +2 authors
Reinforcement learning fine-tuning with Chain-of-Thought rationales enhances Large Language Models' decision-making abilities by addressing greediness, frequency bias, and the knowing-doing gap.
Yunhui Xia, Wei Shen, Yan Wang +5 authors
LeetCodeDataset provides a benchmark for evaluating and training code-generation models with reasoning-focused coding tasks and supports efficient supervised fine-tuning.
Ali Behrouz, Meisam Razaviyayn, Peilin Zhong +1 authors
A novel framework, Miras, introduces associative memory and new attentional bias mechanisms to design deep learning architectures capable of exceptional performance in various tasks.
Chenjie Cao, Jingkai Zhou, Shikai Li +5 authors
Uni3C, a unified 3D-enhanced framework, improves camera and human motion control in video generation by leveraging unprojected point clouds and SMPL-X characters for robust and precise control.
Khiem Vuong, Anurag Ghosh, Deva Ramanan +2 authors
A scalable hybrid dataset combining pseudo-synthetic renderings and real ground-level images improves performance in geometric reconstruction and novel-view synthesis across aerial-ground viewpoints.
Muhammad Khalifa, Rishabh Agarwal, Lajanugen Logeswaran +5 authors
ThinkPRM, a long chain-of-thought verifier, uses minimal supervision to outperform discriminative PRMs and LLM-as-a-Judge across various benchmarks and token budgets.
Min Wei, Chaohui Yu, Jingkai Zhou +1 authors
A diffusion-based framework generates high-fidelity and temporally consistent video try-on by using animatable textured 3D meshes as frame-level guidance.
Chenghao Xiao, Hou Pong Chan, Hao Zhang +4 authors
LLMs' knowledge boundary perception is encoded in specific layers and exhibits linear differences across languages; a training-free alignment method and bilingual fine-tuning improve cross-lingual knowledge boundary recognition.
Shijie Xia, Yiwei Qin, Xuefeng Li +11 authors
Transition from knowledge-retrieval to thought-construction in generative AI through test-time scaling techniques marks a new era of human-AI interaction.
David Ma, Yuanxing Zhang, Jincheng Ren +17 authors
IV-Bench evaluates Image-Grounded Video Perception and Reasoning in MLLMs, revealing substantial underperformance across multiple tasks and factors influencing model accuracy.
Junshu Pan, Wei Shen, Shulin Huang +2 authors
Pre-DPO enhances preference optimization in RLHF for LLMs by using a guiding reference model to improve data utilization and performance.
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