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
Joya Chen, Zhaoyang Lv, Shiwei Wu +7 authors
The LIVE framework enhances Large Language Models for real-time video streaming dialogue through a specialized training objective, data generation scheme, and optimized inference pipeline.
Quanfeng Lu, Wenqi Shao, Zitao Liu +7 authors
GUI Odyssey is a dataset for training and evaluating cross-app navigation agents, and OdysseyAgent, a multimodal agent based on Qwen-VL, demonstrates superior accuracy compared to existing models.
Ihor Stepanov, Mykhailo Shtopko
A small encoder model, GLiNER, achieves state-of-the-art performance on zero-shot and various information extraction tasks, combining size efficiency with strong generalization.
Junjie Wang, Yin Zhang, Yatai Ji +13 authors
A new dataset format called PIN is introduced to enhance LMMs by improving perceptual and reasoning capabilities through diverse, high-quality multimodal training data.
Sreyan Ghosh, Sonal Kumar, Ashish Seth +6 authors
GAMA, a General-purpose Large Audio-Language Model, excels in diverse audio understanding and complex reasoning tasks through advanced audio processing and specialized instruction tuning.
Alessio Devoto, Yu Zhao, Simone Scardapane +1 authors
By compressing the KV cache based on the $L_2$ norm of key embeddings, the sizes required for transformer models can be significantly reduced without sacrificing accuracy.
Bingqi Ma, Zhuofan Zong, Guanglu Song +2 authors
A novel framework and LLM-Infused Diffusion Transformer (LI-DiT) improve text-to-image generation by effectively integrating state-of-the-art large language models and overcoming prompt encoding limitations.
Zeyu Liu, Weicong Liang, Yiming Zhao +3 authors
Glyph-ByT5-v2 and Glyph-SDXL-v2 enhance multilingual visual text rendering accuracy and aesthetic quality using a large multilingual dataset and an advanced preference learning approach.
Hitesh Wadhwa, Rahul Seetharaman, Somyaa Aggarwal +6 authors
The study investigates the RAG framework and reveals that language models prefer using external context over their internal parametric memory, as demonstrated through causal mediation analysis and attention contributions.
Anas Awadalla, Le Xue, Oscar Lo +11 authors
MINT-1T, the largest open-source multimodal interleaved dataset featuring a trillion text tokens and three billion images, enhances the performance of large multimodal models.
Jie Wu, Yufeng Zhu, Lei Shen +1 authors
GEB-1.3B, a lightweight LLM trained on multilingual data, employs novel techniques to achieve efficient CPU inference and outperforms other lightweight models on benchmarks.
Hao Bai, Yifei Zhou, Mert Cemri +4 authors
A novel autonomous RL approach, DigiRL, fine-tunes a pre-trained VLM for in-the-wild device control, achieving superior performance over existing methods.
Siyuan Qi, Bangcheng Yang, Kailin Jiang +5 authors
Consistent In-Context Editing leverages in-context learning to robustly fine-tune language models, improving knowledge editing by maintaining accuracy, locality, generalization, and linguistic quality.
Zhihan Zhang, Zhenwen Liang, Wenhao Yu +4 authors
Reflective augmentation enhances problem-solving abilities of language models in complex scenarios by embedding reflection into training instances.
Andrea Amaduzzi, Pierluigi Zama Ramirez, Giuseppe Lisanti +2 authors
LLaNA, an MLLM integrated with NeRFs, effectively performs tasks such as NeRF captioning and Q&A by processing NeRF weights directly, outperforming 2D or 3D representation extraction methods.
Vasu Singla, Kaiyu Yue, Sukriti Paul +7 authors
Training large vision-language models requires extensive, high-quality image-text pairs. Existing web-scraped datasets, however, are noisy and lack detailed image descriptions. To bridge this gap, we introduce PixelProse, a comprehensive dataset of over 16M (million) synthetically generated captions, leveraging cutting-edge vision-language models for detailed and accurate descriptions. To ensure data integrity, we rigorously analyze our dataset for problematic content, including child sexual abuse material (CSAM), personally identifiable information (PII), and toxicity. We also provide valuable metadata such as watermark presence and aesthetic scores, aiding in further dataset filtering. We hope PixelProse will be a valuable resource for future vision-language research. PixelProse is available at https://huggingface.co/datasets/tomg-group-umd/pixelprose
Guanting Dong, Keming Lu, Chengpeng Li +4 authors
AutoIF is a scalable method for automatically generating high-quality instruction-following training data by using code verification and execution feedback-based rejection sampling, improving various training algorithms for LLMs.
Egor Bogomolov, Aleksandra Eliseeva, Timur Galimzyanov +8 authors
Long Code Arena introduces six benchmarks for code processing tasks requiring project-wide context, supported by open-source datasets and baseline solutions.
Zhen Huang, Zengzhi Wang, Shijie Xia +25 authors
OlympicArena, a benchmark for evaluating AI's cognitive reasoning across disciplines, demonstrates current models' limitations and provides resources to advance AI towards superintelligence.
Wenxuan Zhou, Ravi Agrawal, Shujian Zhang +5 authors
A novel weighted preference optimization method addresses distributional gaps in reinforcement learning from human feedback, improving performance on instruction-following benchmarks.
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