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。
Mukul Singh, José Cambronero, Sumit Gulwani +3 authors
CodeFusion, a diffusion model for code generation, outperforms auto-regressive models in natural language to code tasks by iteratively refining the entire program.
47 篇论文 · 按点赞排序
Sanchit Gandhi, Patrick von Platen, Alexander M. Rush
Distil-Whisper, a smaller and faster variant of the Whisper model, achieves nearly the same performance with fewer resources and is optimized for low-latency environments.
Wei-Ge Chen, Irina Spiridonova, Jianwei Yang +2 authors
LLaVA-Interactive is a cost-efficient multimodal dialogue system that integrates visual chat, image segmentation, and generation capabilities to enhance human-AI interaction.
Ke Hong, Guohao Dai, Jiaming Xu +6 authors
FlashDecoding++ addresses challenges in LLM inference through asynchronized softmax, flat GEMM optimization, and heuristic dataflow adaptation, achieving significant speedups.
Houwen Peng, Kan Wu, Yixuan Wei +17 authors
A new FP8 automatic mixed-precision framework for training large language models reduces memory usage and increases speed compared to BF16 and Nvidia Transformer Engine.
Yufei Wang, Zhou Xian, Feng Chen +6 authors
RoboGen automates the learning of diverse robotic skills at scale using generative simulation to propose tasks and generate training environments and supervision.
Shengnan An, Zexiong Ma, Zeqi Lin +3 authors
LeMa, a learning-from-mistakes approach, enhances LLMs' mathematical reasoning by learning from inaccurate reasoning paths corrected by GPT-4, surpassing SOTA performance on math problems.
Qiying Yu, Quan Sun, Xiaosong Zhang +4 authors
CapsFusion is an advanced framework that improves multimodal pretraining data by combining web-based image-text pairs and synthetic captions, leading to enhanced model performance, sample efficiency, and scalability.
Assaf Shocher, Amil Dravid, Yossi Gandelsman +3 authors
A new generative modeling approach uses an idempotent neural network to project data from a source distribution to a target distribution while maintaining consistency and allowing for refinement.
Mark Levy, Bruno Di Giorgi, Floris Weers +2 authors
Conditional generation from diffusion models is used for various music production tasks using sampling-time guidance, ensuring context matching and conforming to class distributions.
Tianyi Chen, Tianyu Ding, Badal Yadav +2 authors
LoRAShear is an efficient method for pruning and knowledge recovery in Large Language Models, reducing their size by 20% with minimal performance loss.
Chen Wei, Chenxi Liu, Siyuan Qiao +3 authors
De-Diffusion uses a text-to-image diffusion model to represent images as text, enhancing interpretability and enabling transferable prompts for various vision-language tasks.
Micah Goldblum, Hossein Souri, Renkun Ni +10 authors
Battle of the Backbones benchmarks diverse pretrained models across various computer vision tasks, highlighting the current strengths and weaknesses of convolutional neural networks, vision transformers, and self-supervised learning approaches.
Kevin Lin, Faisal Ahmed, Linjie Li +9 authors
MM-VID integrates GPT-4V with specialized tools to transcribe and understand long-form, multimodal videos, supporting advanced tasks like audio description and character identification.
Peter West, Ximing Lu, Nouha Dziri +11 authors
Generative AI models can produce outputs surpassing human experts but fall short in understanding, suggesting a divergence from human intelligence.
Ruizhe Shi, Yuyao Liu, Yanjie Ze +2 authors
The LaMo framework uses Decision Transformers initialized with pre-trained Language Models to improve offline RL, particularly in sparse-reward tasks and limited data scenarios.
Zhaoyang Liu, Zeqiang Lai, Zhangwei Gao +7 authors
ControlLLM enhances large language models by integrating multi-modal tools through task decomposition, optimal solution path finding, and efficient tool execution, achieving superior performance in diverse tasks.
Cameron Jones, Benjamin Bergen
GPT-4 performed better than GPT-3.5 and ELIZA in a public online Turing Test but failed to reach human-level performance, highlighting the importance of linguistic style and socio-emotional traits in human detection.
Yuan Gao, Nobuyuki Morioka, Yu Zhang +1 authors
E3 TTS is a diffusion-based text-to-speech model that generates high-fidelity audio directly from text without intermediate representations or additional conditioning, supporting zero-shot tasks.
Haoxin Chen, Menghan Xia, Yingqing He +9 authors
Two diffusion models, T2V and I2V, generate high-quality videos from text and image inputs respectively, with T2V producing realistic cinematic videos and I2V preserving the content of the reference image.
Zhixuan Chu, Huaiyu Guo, Xinyuan Zhou +9 authors
A data-centric approach using multitask prompt-based fine-tuning and abductive augmentation reasoning enables large language models to excel in financial analysis tasks with limited labeled data.
Ligeng Zhu, Lanxiang Hu, Ji Lin +4 authors
PockEngine is an efficient on-device learning engine that supports sparse backpropagation and graph optimizations for fine-tuning models on edge devices with varying hardware and constraints.
Simon Lermen, Charlie Rogers-Smith, Jeffrey Ladish
Researchers demonstrate that low-rank adaptation can subversively fine-tune large language models, such as Llama 2-Chat, reducing their refusal of harmful instructions with minimal resources.
Sergio Calvo-Ordonez, Jiahao Huang, Lipei Zhang +3 authors
A novel denoising network using continuous dynamical systems improves diffusion models by reducing parameters and FLOPs while enhancing convergence speed and noise robustness.
Jonas F. Lotz, Elizabeth Salesky, Phillip Rust +1 authors
Character bigram rendering improves performance in pixel-based language models, enabling more compact models while maintaining equivalent performance across sentence, token, and multilingual tasks.
Yanai Elazar, Akshita Bhagia, Ian Magnusson +10 authors
WIMBD is a platform for analyzing large text corpora, revealing duplicates, synthetic content, and contamination in datasets used for language models.
Yilong Zhao, Chien-Yu Lin, Kan Zhu +7 authors
Atom is a low-bit quantization method that enhances LLM serving throughput with minimal accuracy loss by utilizing low-bit operators and a mixed-precision process.
Zejun Wang, Jia Li, Ge Li +1 authors
ChatCoder enhances large language models' code generation by refining human requirements through interactive chat sessions, leading to better performance and precision.
Xinyuan Chen, Yaohui Wang, Lingjun Zhang +7 authors
A short-to-long video diffusion model, SEINE, generates smooth and creative transitions between scenes with text-based control, enabling the creation of coherent long videos.
Jackson Petty, Sjoerd van Steenkiste, Ishita Dasgupta +3 authors
Transformers with greater depth (more layers) exhibit enhanced compositional generalization out-of-distribution, albeit with diminishing returns.
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