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
Petr Anokhin, Roman Khalikov, Stefan Rebrikov +3 authors
HeroBench evaluates long-horizon planning and structured reasoning in complex virtual worlds, revealing performance disparities in state-of-the-art LLMs.
Haokun Lin, Haobo Xu, Yichen Wu +6 authors
A systematic study on quantizing diffusion large language models identifies challenges and evaluates state-of-the-art methods across various configurations to improve deployment on edge devices.
Zhuoqun Li, Xuanang Chen, Hongyu Lin +3 authors
PaperRegister enhances paper search by using hierarchical indexing and adaptive retrieval, supporting flexible and fine-grained queries beyond traditional abstract-based systems.
Shilong Li, Xingyuan Bu, Wenjie Wang +21 authors
A new benchmark, MM-BrowseComp, evaluates AI agents' multimodal retrieval and reasoning capabilities, revealing limitations in current models' handling of images and videos in web browsing tasks.
Ruoxi Chen, Dongping Chen, Siyuan Wu +6 authors
Experiments with multiple image-text models and agentic frameworks show that even state-of-the-art systems struggle with generating images from multiple visual references, highlighting the need for more flexible creative tools.
Pengsong Zhang, Xiang Hu, Guowei Huang +20 authors
aiXiv is an open-access platform that facilitates the submission, review, and refinement of scientific proposals and papers by both human and AI scientists, enhancing the quality and dissemination of AI-generated research.
Ronghao Dang, Yuqian Yuan, Yunxuan Mao +6 authors
RynnEC, a video multimodal large language model with a region-centric approach, achieves state-of-the-art performance in object property understanding, segmentation, and spatial reasoning, using an egocentric video pipeline and a region-centered benchmark.
Shiwen Ni, Guhong Chen, Shuaimin Li +11 authors
A systematic review of large language model benchmarks identifies issues such as data contamination, cultural biases, and lack of process credibility, and proposes a design paradigm for future improvements.
Yanxu Meng, Haoning Wu, Ya Zhang +1 authors
SceneGen generates multiple 3D assets from a single scene image using a novel framework that integrates local and global scene information, enabling efficient and robust 3D content creation.
Yifu Yuan, Haiqin Cui, Yaoting Huang +6 authors
A pointing-centric representation and reinforced fine-tuning approach improve embodied AI generalization across various tasks and visual disturbances.
Jinhyung Park, Javier Romero, Shunsuke Saito +7 authors
ATLAS is a high-fidelity body model that decouples shape and skeleton bases, improving pose accuracy and shape expressivity using 600k high-resolution scans.
Greta Tuckute, Klemen Kotar, Evelina Fedorenko +1 authors
AuriStream, a biologically inspired two-stage model, encodes speech using cochlear tokens and an autoregressive sequence model, achieving state-of-the-art performance on speech tasks and generating interpretable audio continuations.
Kai Li, Guo Chen, Wendi Sang +8 authors
A survey of DNN-based speech separation techniques, covering learning paradigms, separation scenarios, and architectural components, with a focus on current advancements and promising future directions.
Zixin Yin, Xili Dai, Ling-Hao Chen +7 authors
ColorCtrl, a training-free method using Multi-Modal Diffusion Transformers, achieves precise and consistent color editing in images and videos with word-level control and superior performance compared to existing approaches.
Yibo Zhang, Li Zhang, Rui Ma +1 authors
TexVerse is a large-scale 3D dataset with high-resolution textures, including PBR materials, rigged models, and animated models, suitable for texture synthesis, PBR material development, and 3D vision tasks.
Francesco Fabbri, Gustavo Penha, Edoardo D'Amico +7 authors
A novel framework uses Large Language Models to evaluate podcast recommendations by constructing user profiles and providing context for the LLM to make judgments, improving efficiency and interpretability.
Yutong Feng, Linlin Zhang, Hengyuan Cao +5 authors
OmniTry extends Virtual Try-ON to various wearable objects using a two-stage pipeline that combines unpaired and paired image training to improve object localization and appearance consistency.
Bishanka Seal, Rahul Seetharaman, Aman Bansal +1 authors
LLMs predict misery scores from text using various prompting strategies, with few-shot approaches outperforming zero-shot, and a gamified framework assessing their adaptability in emotional reasoning tasks.
Jianshu Zeng, Yuxuan Liu, Yutong Feng +6 authors
Lumen is an end-to-end video relighting framework that uses a large-scale dataset of realistic and synthetic videos to achieve consistent lighting and foreground preservation in edited videos.
Jinyi Han, Tingyun Li, Shisong Chen +8 authors
FineCE is a novel confidence estimation method for large language models that provides accurate, fine-grained confidence scores during text generation, outperforming existing 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