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
Pranav Pawar, Kavish Shah, Akshat Bhalani +6 authors
A novel framework evaluates Vision-Language Models on 2D physics understanding, revealing a correlation between model scale and reasoning ability, particularly highlighting challenges in abstract spatial reasoning.
Yuming Jiang, Siteng Huang, Shengke Xue +10 authors
RynnVLA-001, a vision-language-action model, uses a two-stage pretraining approach and ActionVAE to achieve superior performance on robotics tasks.
Zhaokai Wang, Penghao Yin, Xiangyu Zhao +5 authors
GenExam is a benchmark for evaluating text-to-image generation in exam-style settings across multiple disciplines, highlighting the challenges in integrating knowledge, reasoning, and generation.
Zixin Yin, Xili Dai, Duomin Wang +4 authors
LazyDrag, a drag-based image editing method for Multi-Modal Diffusion Transformers, eliminates implicit point matching, enabling precise geometric control and text guidance without test-time optimization.
Zhaoyang Chu, Yao Wan, Zhikun Zhang +7 authors
CodeEraser effectively and efficiently removes sensitive memorized information from Code Language Models using machine unlearning techniques without full retraining.
Iman Barati, Mostafa Amiri, Heshaam Faili
SearchInstruct enhances supervised fine-tuning datasets for large language models by expanding domain-specific questions and retrieving accurate answers, improving model performance and enabling efficient model editing.
Gang Cheng, Xin Gao, Li Hu +23 authors
Wan-Animate is a unified framework for character animation and replacement, using spatially-aligned skeleton signals and implicit facial features to generate high-fidelity character videos with seamless environmental integration.
Siyan Zhao, Mengchen Liu, Jing Huang +8 authors
IGPO, an RL framework utilizing inpainting in masked diffusion large language models, enhances sample efficiency and achieves state-of-the-art results in mathematical benchmarks.
Qikai Chang, Zhenrong Zhang, Pengfei Hu +6 authors
THOR, a tool-integrated hierarchical optimization framework using RL, enhances mathematical reasoning and code generation by constructing high-quality datasets, optimizing reasoning paths, and correcting errors during inference.
Fei Xiong, Xiang Zhang, Aosong Feng +2 authors
QuantAgent, a multi-agent LLM framework, excels in high-frequency trading by leveraging specialized agents for technical indicators, chart patterns, trends, and risk, outperforming existing neural and rule-based systems.
Zikang Guo, Benfeng Xu, Chiwei Zhu +3 authors
MCP-AgentBench is a benchmark designed to evaluate language agents in MCP-mediated tool interactions, providing a standardized framework for assessing real-world performance.
Mingsong Li, Lin Liu, Hongjun Wang +7 authors
MultiEdit, a comprehensive dataset with over 107K high-quality image editing samples, improves performance on sophisticated editing tasks using a novel pipeline with multi-modal large language models.
Moritz Reuss, Hongyi Zhou, Marcel Rühle +3 authors
FLOWER, a 950 M-parameter VLA policy, achieves competitive performance with reduced computational costs through intermediate-modality fusion and action-specific Global-AdaLN conditioning.
Zhiyuan Huang, Jiahao Chen, Bing Su +1 authors
Foundation model fine-tuning addresses long-tailed semi-supervised learning challenges by reducing hypothesis complexity and improving robustness through parameter-efficient training techniques.
An-Chieh Cheng, Yang Fu, Yukang Chen +10 authors
A Spatial Region 3D (SR-3D) vision-language model unifies 2D and 3D representations by enriching 2D features with 3D positional embeddings, enabling flexible region prompting and accurate spatial reasoning across frames.
Yining Lu, Zilong Wang, Shiyang Li +6 authors
Dynamic reward weighting in multi-objective reinforcement learning adaptively adjusts weights during training to explore Pareto fronts effectively, outperforming fixed-weight scalarization methods.
Klemen Kotar, Wanhee Lee, Rahul Venkatesh +13 authors
Probabilistic Structure Integration (PSI) learns richly controllable world models from data through probabilistic prediction, structure extraction, and integration, enhancing video prediction and understanding.
Artem Lukoianov, Chenyang Yuan, Justin Solomon +1 authors
Research shows that locality in deep diffusion models is a statistical property of image datasets rather than an inductive bias of convolutional neural networks, leading to the development of a more accurate analytical denoiser.
Nicholas Pipitone, Ghita Houir Alami, Advaith Avadhanam +2 authors
A novel training methodology named zELO optimizes retrieval performance by treating ranking tasks as equivalent to a Thurstone model, resulting in state-of-the-art open-weight reranker models that outperform proprietary models across various domains.
Sung-Lin Tsai, Bo-Lun Huang, Yu Ting Shen +5 authors
A training-free framework uses a large language model to disambiguate color terms and refine text embeddings for improved color accuracy in text-to-image generation.
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