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
Jihai Zhang, Xiaoye Qu, Tong Zhu +1 authors
Diversified Multiplet Upcycling (DMU) enhances CLIP models by fine-tuning them to capture diverse feature spaces, transforming them into CLIP-MoEs, and demonstrating superior performance in various zero-shot tasks and MLLM benchmarks.
Junlin Han, Jianyuan Wang, Andrea Vedaldi +2 authors
Flex3D generates high-quality 3D content from diverse input views using a two-stage framework with image and video diffusion models, followed by a transformer-based flexible reconstruction model.
Sathya Krishnan Suresh, Wu Mengjun, Tushar Pranav +1 authors
DiaSynth generates high-quality, contextually rich, and domain-specific synthetic dialogues using a Large Language Model with Chain of Thought reasoning, demonstrating competitive performance in comparison to in-domain datasets.
Amirhossein Kazemnejad, Milad Aghajohari, Eva Portelance +4 authors
VinePPO, a Monte Carlo-based method, improves reinforcement learning fine-tuning of large language models by accurately assigning credit to reasoning steps, outperforming PPO and RL-free baselines.
Zechen Bai, Tong He, Haiyang Mei +6 authors
VideoLISA, a video-based multimodal large language model, uses reasoning and a Sparse Dense Sampling strategy to perform temporally consistent segmentation and object tracking in videos based on language instructions.
Pengxin Guo, Shuang Zeng, Yanran Wang +3 authors
FedSA-LoRA integrates low-rank adaptation with federated learning by asymmetrically sharing only the general knowledge matrices, enhancing efficiency and effectiveness in natural language tasks.
Cheolhun Jang
Modulated Intervention Preference Optimization (MIPO) adapts the degree of intervention from a reference model based on data alignment, improving performance over traditional methods like DPO in preference optimization tasks.
Marco Gaido, Sara Papi, Luisa Bentivogli +6 authors
Researchers address the lack of compliance with open-source principles in speech foundation models by releasing a multilingual dataset and transcripts to facilitate the creation of EU language-specific open-source models.
Rinon Gal, Adi Haviv, Yuval Alaluf +3 authors
Automatic generation of text-to-image workflows based on user prompts enhances image quality compared to monolithic models or generic workflows.
Gabriel Mongaras, Trevor Dohm, Eric C. Larson
Cottention, a linear memory complexity attention mechanism based on cosine similarity, reduces memory requirements without sacrificing performance in transformer-based models.
Sang Hyun Park, Jun Young Koh, Junha Lee +5 authors
Illustrious, a text-to-image anime model, achieves high-resolution, dynamic color output, and strong restoration capabilities through controlled token activations, higher training resolutions, and refined multi-level captions, outperforming existing models.
Qiaojun Yu, Siyuan Huang, Xibin Yuan +9 authors
UniAff integrates 3D object-centric manipulation and task understanding using MLLMs to improve robotic manipulation generalization for tools and articulated objects.
Zhenting Qi, Hongyin Luo, Xuliang Huang +5 authors
Scylla, a dynamic evaluation framework, measures LLM generalization by assessing performance on in-distribution and out-of-distribution data, revealing a non-monotonic relationship between task complexity and generalization.
Lirui Wang, Xinlei Chen, Jialiang Zhao +1 authors
Heterogeneous Pre-trained Transformers enhance general robot model capabilities by learning a shared representation across diverse robotic tasks and embodiments.
Wenhao Wang, Yifan Sun, Zhentao Tan +1 authors
PDF-Embedding is a novel deep embedding method for detecting content replication in images generated by diffusion models, achieving superior performance on a new D-Rep dataset compared to existing methods.
Ulyana Piterbarg, Lerrel Pinto, Rob Fergus
LintSeq generates synthetic code editing data to improve the diversity and performance of LLMs in code synthesis, outperforming models trained on baseline datasets.
Leheng Li, Weichao Qiu, Yingjie Cai +4 authors
SyntheOcc, a diffusion model using 3D semantic multi-plane images, generates photorealistic and aligned 3D occupancy datasets for autonomous driving perception models.
Zhen Han, Zeyinzi Jiang, Yulin Pan +5 authors
ACE, a unified Transformer-based diffusion model, achieves comparable performance to expert models across various visual generation tasks by using a Long-context Condition Unit and addressing data scarcity with a novel data collection method.
Shayekh Bin Islam, Md Asib Rahman, K S M Tozammel Hossain +3 authors
Open-RAG enhances open-source LLMs by transforming them into sparse mixture-of-experts models, improving reasoning capabilities and factual accuracy in knowledge-intensive tasks.
Zecheng Tang, Keyan Zhou, Juntao Li +3 authors
L-CiteEval, a benchmark for long-context understanding with citations, evaluates the faithfulness and understanding capability of LCMs across diverse tasks, revealing differences between open-source and closed-source models and the impact of attention mechanisms.
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