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
Maitreya Patel, Sangmin Jung, Chitta Baral +1 authors
The new model $\lambda$-ECLIPSE achieves resource-efficient, high-quality personalized text-to-image generation with fewer parameters and training hours, outperforming existing methods in composition and concept alignment.
Yiyuan Zhang, Yuhao Kang, Zhixin Zhang +3 authors
InteractiveVideo is a user-centric framework enabling dynamic interaction and fine-grained refinement of video generation through multimodal inputs.
Zach Nussbaum, John X. Morris, Brandon Duderstadt +1 authors
The report details the development and release of nomic-embed-text-v1, an open-source English text embedding model that surpasses OpenAI models in both short and long-context tasks.
Tianjun Zhang, Aman Madaan, Luyu Gao +5 authors
LEAP, a novel few-shot learning approach, enhances LLM performance by inducing and reflecting on mistakes from initial examples to learn task-specific principles, applied across various benchmarks.
Yadong Zhang, Shaoguang Mao, Tao Ge +4 authors
Research introduces "K-Level Reasoning" to enhance LLM performance in dynamic, interactive scenarios using game theory-based challenges and recursive k-level thinking.
Peng Gao, Renrui Zhang, Chris Liu +16 authors
SPHINX-X, an extensive Multimodality Large Language Model series, enhances training efficiency and performance by modifying the SPHINX framework and using comprehensive, curated datasets covering language, vision, and vision-language tasks.
Geoffrey Cideron, Sertan Girgin, Mauro Verzetti +11 authors
MusicRL, a pretrained MusicLM model refined using reinforcement learning from human feedback, produces high-quality text-to-music output preferred by human raters over a baseline model.
Bo-Kyeong Kim, Geonmin Kim, Tae-Ho Kim +4 authors
A depth pruning method for large language models matches width pruning in zero-shot task performance and improves inference speed under memory constraints.
Zhifeng Kong, Arushi Goel, Rohan Badlani +3 authors
Audio Flamingo, a novel audio language model, achieves strong audio understanding, in-context learning, and multi-turn dialogue capabilities through advanced training techniques and architecture design.
Jihan Yang, Runyu Ding, Ellis Brown +2 authors
V-IRL provides a virtual platform for AI to interact with the real world, enhancing realism for development and testing in perception, decision-making, and interaction.
Yang Jin, Zhicheng Sun, Kun Xu +10 authors
A framework for pre-training multimodal Large Language Models using video decomposition into keyframes and temporal motions achieves competitive performance across benchmarks.
Michael Zhang, Kush Bhatia, Hermann Kumbong +1 authors
Hedgehog, a learnable linear attention mechanism that mimics the properties of softmax attention, significantly improves the efficiency and performance of linear Transformers in various conversion settings.
Natasha Butt, Blazej Manczak, Auke Wiggers +4 authors
A Code Iteration method for language models achieves state-of-the-art performance on the Abstraction and Reasoning Corpus by combining program sampling, hindsight relabeling, and prioritized experience replay.
Xiangxiang Chu, Limeng Qiao, Xinyu Zhang +8 authors
MobileVLM V2 enhances vision language models through improved architecture, training, and dataset curation, achieving performance on par with or better than larger models.
Jianyuan Guo, Hanting Chen, Chengcheng Wang +3 authors
Weak-to-strong supervision using an adaptively adjustable loss function significantly enhances vision foundation models' performance across various learning scenarios.
Tu Anh Nguyen, Benjamin Muller, Bokai Yu +11 authors
A foundation multimodal language model combines text and speech using speech and text units, supporting semantic and expressive capabilities and few-shot learning across tasks.
Chengjian Feng, Yujie Zhong, Zequn Jie +2 authors
A diffusion model enhanced with an instance-level grounding head serves as a synthetic data generator to improve object detection performance, especially in open-vocabulary and data-sparse scenarios.
Yehui Tang, Fangcheng Liu, Yunsheng Ni +7 authors
Tiny language models optimized through tokenizer compression, architecture tweaking, and parameter inheritance achieve high performance on mobile devices with reduced computation and memory costs.
Dejiao Zhang, Wasi Ahmad, Ming Tan +5 authors
A two-stage pretraining scheme using a combination of masking language modeling and contrastive learning significantly improves code representation learning and performance on various downstream tasks.
Shijia Liao, Shiyi Lan, Arun George Zachariah
EVA-GAN improves HiFi audio generation by using a large dataset and scalable GANs with a context-aware module, enhancing spectral and high-frequency accuracy and out-of-domain robustness.
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