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。
Matt Deitke, Christopher Clark, Sangho Lee +48 authors
Molmo, a new family of open-weight VLMs, excels through a high-quality human-annotated image caption dataset and a diverse fine-tuning dataset, outperforming both proprietary and open models in benchmark and human evaluations.
50 篇论文 · 按点赞排序
Zecheng He, Bo Sun, Felix Juefei-Xu +14 authors
Imagine yourself is a tuning-free diffusion model for personalized image generation that enhances identity preservation, text alignment, and visual quality through synthetic data generation, parallel attention architecture, and multi-stage fine-tuning.
Fan Zhou, Zengzhi Wang, Qian Liu +2 authors
ProX, a novel framework treating data refinement as a programming task, enhances large language model pre-training by generating fine-grained operations for each example, outperforming traditional human-crafted rule-based methods significantly across various benchmarks and domains.
Abhilash Nandy, Yash Agarwal, Ashish Patwa +5 authors
Current Vision-Language models struggle with satirical image detection, understanding, and completion tasks on a newly released high-quality dataset called YesBut.
Johannes Schmude, Sujit Roy, Will Trojak +26 authors
Prithvi WxC, a 2.3 billion parameter foundation model using an encoder-decoder architecture with transformer concepts, addresses weather forecasting, downscaling, and extreme events estimation with a mixed masked reconstruction and forecasting objective.
Gongfan Fang, Hongxu Yin, Saurav Muralidharan +5 authors
MaskLLM introduces a learnable pruning method using Gumbel Softmax sampling to achieve N:M sparsity in LLMs, improving inference efficiency and transferability across tasks.
Haoran Que, Feiyu Duan, Liqun He +11 authors
A comprehensive benchmark, HelloBench, and evaluation method, HelloEval, are introduced to assess the long text generation capabilities of Large Language Models, revealing limitations and providing human-aligned evaluations.
Kai Chen, Yunhao Gou, Runhui Huang +27 authors
EMOVA integrates speech, vision, and language capabilities, achieving top performance in benchmarks and enabling emotive spoken dialogue using a disentangled speech tokenizer and style module.
Yunfei Xie, Juncheng Wu, Haoqin Tu +6 authors
OpenAI's o1 demonstrates strong performance in medical tasks but shows weaknesses in reasoning, multilinguality, and evaluation metrics compared to GPT-4.
Chenming Zhu, Tai Wang, Wenwei Zhang +2 authors
LLaVA-3D enhances 2D models with 3D scene understanding by integrating 3D Patch representations and joint 2D/3D instruction tuning, achieving fast convergence and superior performance across tasks.
Yifang Men, Yuan Yao, Miaomiao Cui +1 authors
A novel framework called MIMO synthesizes character videos with controllable attributes using 2D encodings and 3D depth, achieving scalability, generality, and applicability to interactive scenes.
Yinpei Dai, Jayjun Lee, Nima Fazeli +1 authors
RACER, a supervisor-actor framework enhanced with failure recovery data and rich language descriptions, improves robotic manipulation by outperforming RVT in various tasks and environments.
Jing He, Haodong Li, Wei Yin +6 authors
A diffusion-based visual foundation model, Lotus, is introduced for dense prediction tasks, achieving state-of-the-art performance in zero-shot depth and normal estimation while improving efficiency.
Yizhi Li, Ge Zhang, Yinghao Ma +17 authors
OmniBench is a benchmark for evaluating multimodal large language models' ability to simultaneously process and reason across visual, acoustic, and textual inputs, revealing significant limitations in current models.
Pedro Henrique Martins, Patrick Fernandes, João Alves +12 authors
The EuroLLM project develops multilingual language models for European Union languages, detailing data collection, tokenizer creation, and model performance.
Chaofan Li, MingHao Qin, Shitao Xiao +5 authors
A novel model leveraging in-context learning and few-shot examples in large language models improves text embedding generation, achieving state-of-the-art performance on MTEB and AIR-Bench benchmarks.
Byung-Kwan Lee, Sangyun Chung, Chae Won Kim +2 authors
Phantom, a new efficient LLVM family with smaller model sizes, enhances learning capabilities by temporarily increasing latent hidden dimensions during MHSA and employing Phantom Optimization for improved performance over larger models.
John Hewitt, Nelson F. Liu, Percy Liang +1 authors
Adaptations without explicit instruction tuning, such as training on responses or narrow-domain instruction-response pairs, can still lead to broad instruction-following behavior in language models.
Zhenmei Shi, Yifei Ming, Xuan-Phi Nguyen +2 authors
GemFilter algorithm accelerates Large Language Model inference and reduces GPU memory usage by filtering input tokens using early model layers.
Satyapriya Krishna, Kalpesh Krishna, Anhad Mohananey +4 authors
A new evaluation dataset called FRAMES assesses LLMs' factuality, retrieval, and reasoning in retrieval-augmented generation (RAG) tasks, demonstrating the need for improved multi-step retrieval pipelines.
Weifeng Lin, Xinyu Wei, Renrui Zhang +7 authors
PixWizard, a versatile visual assistant, uses Diffusion Transformers to perform diverse image-to-image tasks with flexible resolution handling and instruction-based guidance.
Clément Christophe, Tathagata Raha, Svetlana Maslenkova +4 authors
Four techniques—continuous pretraining, instruct fine-tuning, NEFTune, and prompt engineering—improve Large Language Models for clinical applications, with NEFTune and advanced prompt engineering yielding significant performance enhancements.
Yaoxun Xu, Hangting Chen, Jianwei Yu +5 authors
MuCodec, using MuEncoder, RVQ, flow-matching, MEL-VAE, and HiFi-GAN, achieves high-fidelity music reconstruction at ultra-low bitrates.
Christina Zhang, Simran Motwani, Matthew Yu +6 authors
Adding pixel-space supervision to latent diffusion models improves high-frequency detail preservation during post-training without compromising text alignment quality.
Pablo Bernabeu-Perez, Enrique Lopez-Cuena, Dario Garcia-Gasulla
Synthetic image detectors must generalize widely and resist alterations in a rapidly evolving field, where the improvement of generators drives advancements in detectors and vice versa.
Jordi Bayarri-Planas, Ashwin Kumar Gururajan, Dario Garcia-Gasulla
The study enhances the factuality and reliability of large language models in healthcare through optimized context retrieval methods, achieving performance comparable to private solutions in benchmarks and proposing OpenMedPrompt for realistic open-ended answers.
Mark Hamilton, Simon Stent, Vasha DuTell +4 authors
A study investigating the detection of face-like structures in non-face images by both humans and face detectors, revealing differences in pareidolia and proposing a statistical model to explain these detections.
Chuyang Zhao, Yuxing Song, Wenhao Wang +5 authors
A shared transformer backbone for both autoregressive text generation and diffusion-based visual generation achieves comparable performance to state-of-the-art methods.
Mark Weber, Lijun Yu, Qihang Yu +4 authors
Modernized VQGAN and embedding-free bit token based Transformer models achieve state-of-the-art results in class-conditional image generation.
Xiaoming Shi, Shiyu Wang, Yuqi Nie +4 authors
Time-MoE, a scalable and efficient architecture using a sparse mixture-of-experts design, achieves state-of-the-art time series forecasting by pre-training large models on a massive dataset without increasing inference costs.
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