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
Ziteng Wang, Jianfei Chen, Jun Zhu
ReMoE, a differentiable MoE architecture using ReLU-based routing, outperforms traditional TopK routing in terms of scalability and performance across various models and expert counts.
Saehyung Lee, Seunghyun Yoon, Trung Bui +2 authors
A multiagent approach using LLM-MLLM collaboration improves the factual accuracy of detailed image captions, surpassing existing methods and highlighting the limitations of VQA benchmarks.
Zhen Zheng, Xiaonan Song, Chuanjie Liu
MixLLM, a mixed-precision quantization technique, enhances the accuracy and efficiency of large language models by optimizing output feature quantization and system-level dequantization processes.
Haofei Yu, Zhaochen Hong, Zirui Cheng +5 authors
ResearchTown, a multi-agent framework using TextGNN, simulates collaborative research activities in scientific domains with realistic interaction and interdisciplinary idea generation.
Yanheng He, Jiahe Jin, Shijie Xia +5 authors
The PC Agent system demonstrates digital work capabilities through human cognition transfer using a PC Tracker, cognition completion pipeline, and multi-agent system, showing data efficiency in training digital agents.
Viorica Pătrăucean, Xu Owen He, Joseph Heyward +10 authors
A new TRecViT architecture combines gated LRUs, self-attention, and MLPs for efficient video modeling, outperforming ViViT-L with reduced parameters, memory, and computational cost.
Shayne Longpre, Nikhil Singh, Manuel Cherep +40 authors
Progress in AI is driven largely by the scale and quality of training data. Despite this, there is a deficit of empirical analysis examining the attributes of well-established datasets beyond text. In this work we conduct the largest and first-of-its-kind longitudinal audit across modalities--popular text, speech, and video datasets--from their detailed sourcing trends and use restrictions to their geographical and linguistic representation. Our manual analysis covers nearly 4000 public datasets between 1990-2024, spanning 608 languages, 798 sources, 659 organizations, and 67 countries. We find that multimodal machine learning applications have overwhelmingly turned to web-crawled, synthetic, and social media platforms, such as YouTube, for their training sets, eclipsing all other sources since 2019. Secondly, tracing the chain of dataset derivations we find that while less than 33% of datasets are restrictively licensed, over 80% of the source content in widely-used text, speech, and video datasets, carry non-commercial restrictions. Finally, counter to the rising number of languages and geographies represented in public AI training datasets, our audit demonstrates measures of relative geographical and multilingual representation have failed to significantly improve their coverage since 2013. We believe the breadth of our audit enables us to empirically examine trends in data sourcing, restrictions, and Western-centricity at an ecosystem-level, and that visibility into these questions are essential to progress in responsible AI. As a contribution to ongoing improvements in dataset transparency and responsible use, we release our entire multimodal audit, allowing practitioners to trace data provenance across text, speech, and video.
Zhexin Zhang, Shiyao Cui, Yida Lu +4 authors
Agent-SafetyBench evaluates LLM agents in interactive environments, revealing significant safety challenges and emphasizing the need for advanced strategies to improve agent reliability.
Govind Mittal, Sarthak Gupta, Shruti Wagle +5 authors
WavePulse, a real-time framework for radio content analysis, effectively tracks political science questions during livestreams of news radio stations.
Aakash Mahalingam, Vinesh Kumar Gande, Aman Chadha +2 authors
SKETCH enhances RAG systems by integrating semantic text retrieval with knowledge graphs, improving retrieval performance and context integrity across various datasets.
Hyun-kyu Ko, Dongheok Park, Youngin Park +3 authors
Video super-resolution models enhance 3D super-resolution by ensuring spatial consistency across low-resolution multi-view images without the need for complex alignment techniques.
Yuchi Wang, Junliang Guo, Xinyi Xie +3 authors
VidTwin, a novel video autoencoder, decouples video into structure and dynamics latent spaces using a Q-Former and downsampling, achieving high compression and reconstruction quality.
Sara Papi, Peter Polak, Ondřej Bojar +1 authors
Research on simultaneous speech-to-text translation focuses on pre-segmented speech, lacking standard terminology and addressing challenges for real-world applications.
Yuxiang Zhang, Yuqi Yang, Jiangming Shu +3 authors
OpenAI's recent introduction of Reinforcement Fine-Tuning (RFT) showcases the potential of reasoning foundation model and offers a new paradigm for fine-tuning beyond simple pattern imitation. This technical report presents OpenRFT, our attempt to fine-tune generalist reasoning models for domain-specific tasks under the same settings as RFT. OpenRFT addresses two key challenges of lacking reasoning step data and the limited quantity of training samples, by leveraging the domain-specific samples in three ways: question augmentation, synthesizing reasoning-process data, and few-shot ICL. The evaluation is conducted on SciKnowEval, where OpenRFT achieves notable performance gains with only 100 domain-specific samples for each task. More experimental results will be updated continuously in later versions. Source codes, datasets, and models are disclosed at: https://github.com/ADaM-BJTU/OpenRFT
Yueqian Wang, Xiaojun Meng, Yuxuan Wang +3 authors
A Friends-MMC dataset facilitates research on multi-modal multi-party conversations by enabling studies on speaker identification and response prediction with video context.
Minda Hu, Qiyuan Zhang, Yufei Wang +7 authors
The NILE framework optimizes Instruction Fine-Tuning datasets by aligning them with a pre-trained LLM's internal knowledge, significantly enhancing LLM performance across various evaluations.
Shijie Wang, Samaneh Azadi, Rohit Girdhar +3 authors
MotiF improves text-guided image-to-video generation by focusing on motion regions, enhancing text alignment and motion quality using a motion heatmap and a newly proposed benchmark dataset.
Jiangnan Fang, Cheng-Tse Liu, Jieun Kim +9 authors
A framework for multi-LLM summarization using both centralized and decentralized strategies improves performance over single-LLM models.
Yiyu Zhuang, Jiaxi Lv, Hao Wen +7 authors
A scalable transformer model using a large dataset of high-quality human images achieves fast and high-fidelity 3D avatar reconstruction from a single image with support for animation and editing.
Sophia Tang, Yinuo Zhang, Pranam Chatterjee
PepTune, a multi-objective discrete diffusion model based on MDLM, uses Monte Carlo Tree Search to generate diverse, optimized therapeutic peptides with multiple desirable properties.
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