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。
Opher Lieber, Barak Lenz, Hofit Bata +19 authors
Jamba, a hybrid Transformer-Mamba MoE model, achieves state-of-the-art performance with high throughput and minimal memory usage, supporting large context lengths.
50 篇论文 · 按点赞排序
David Raposo, Sam Ritter, Blake Richards +3 authors
Transformer-based models can dynamically allocate computational resources to specific positions in input sequences, leading to efficient performance with reduced FLOPs.
Zhengxuan Wu, Aryaman Arora, Zheng Wang +4 authors
Representation Finetuning (ReFT) methods, exemplified by Low-rank Linear Subspace ReFT (LoReFT), achieve high efficiency and performance by adapting representations in frozen base models, outperforming state-of-the-art Parameter-efficient Fine-tuning (PEFT) methods.
Keyu Tian, Yi Jiang, Zehuan Yuan +2 authors
Visual AutoRegressive (VAR) modeling improves image generation performance and speed over autoregressive and diffusion transformer models, achieving superior results in image quality, data efficiency, and scalability.
Wei Chen, Zhiyuan Li
A new on-device method improves the performance of a large language model, matching GPT-4's accuracy and latency while reducing context length and enhancing deployment on edge devices.
Hyungjoo Chae, Yeonghyeon Kim, Seungone Kim +8 authors
Think-and-Execute is a novel framework that enhances algorithmic reasoning in large language models by decomposing the reasoning process into task-level logic expressed as pseudocode and instance-specific execution.
Jinhyuk Lee, Zhuyun Dai, Xiaoqi Ren +17 authors
Gecko is a compact and versatile text embedding model that distills knowledge from large language models to achieve strong retrieval performance on benchmarks.
Lifan Yuan, Ganqu Cui, Hanbin Wang +12 authors
Eurus, a suite of reasoning-optimized large language models, achieves state-of-the-art performance on various benchmarks through UltraInteract, a large-scale, high-quality alignment dataset, and a novel reward modeling objective.
Tianle Li, Ge Zhang, Quy Duc Do +2 authors
LIConBench evaluates long-context LLMs on extreme-label classification tasks with sequences up to 50K tokens, highlighting performance dips beyond 20K tokens and favoring of recent labels.
Dongzhi Jiang, Guanglu Song, Xiaoshi Wu +5 authors
CoMat, an end-to-end diffusion model fine-tuning strategy with an image-to-text concept matching mechanism, improves text-to-image alignment in SDXL without additional data.
Luchang Li, Sheng Qian, Jie Lu +3 authors
Transformer-Lite optimizes LLM deployment on device GPUs through various techniques to achieve significant speedups in prefill and decoding over CPU and GPU-based methods.
Agneet Chatterjee, Gabriela Ben Melech Stan, Estelle Aflalo +8 authors
The paper introduces SPRIGHT, a dataset focused on spatial relationships, and fine-tuning methods that significantly improve spatial consistency in text-to-image models.
Taishi Nakamura, Mayank Mishra, Simone Tedeschi +42 authors
Aurora-M, a 15B parameter multilingual open-source model trained on multiple languages and code, addresses challenges of catastrophic forgetting and AI safety through continual pretraining and human-reviewed safety instructions.
Hanyu Lai, Xiao Liu, Iat Long Iong +8 authors
AutoWebGLM, an automated web navigation agent based on ChatGLM3-6B, addresses webpage complexity and action versatility through HTML simplification and reinforcement learning, outperforming GPT-4 in real-world web navigation tasks.
Kirolos Ataallah, Xiaoqian Shen, Eslam Abdelrahman +4 authors
MiniGPT4-Video, an extension of MiniGPT-v2, processes sequences of video frames and textual conversations, outperforming state-of-the-art models on multiple video-text benchmarks.
Musashi Hinck, Matthew L. Olson, David Cobbley +2 authors
LLaVA-Gemma models, trained with varying design features, achieve moderate performance relative to state-of-the-art models with similar sizes.
Kang Min Yoo, Jaegeun Han, Sookyo In +374 authors
HyperCLOVA X is a multilingual large language model trained in Korean, English, and code, demonstrating strong reasoning, knowledge, and cross-lingual capabilities.
Gabriela Ben Melech Stan, Raanan Yehezkel Rohekar, Yaniv Gurwicz +7 authors
An interactive application enhances interpretability of vision-language models by analyzing image patches and assessing language model's grounding in images, with a case study on LLaVA.
Hao He, Yinghao Xu, Yuwei Guo +4 authors
CameraCtrl enhances video generation by enabling precise camera pose control for text-to-video models, improving controllability and generalization with diverse datasets.
Ruowen Zhao, Zhengyi Wang, Yikai Wang +2 authors
FlexiDreamer is a novel end-to-end framework for single-image 3D generation using FlexiCubes and a multi-resolution hash grid encoding scheme to improve efficiency and reduce visual artifacts.
Brian Lester, Jaehoon Lee, Alex Alemi +4 authors
Equal-Info Windows compress text uniformly for effective LLM training, reducing sequence lengths and latency.
Joel Ruben Antony Moniz, Soundarya Krishnan, Melis Ozyildirim +5 authors
LLMs are effectively utilized to resolve various types of references, including non-conversational entities on screen, by converting reference resolution into a language modeling problem, showing significant improvements over existing systems.
Haofan Wang, Qixun Wang, Xu Bai +2 authors
InstantStyle addresses style consistency and detail retention in image generation by decoupling style and content and focusing style-specific feature injection, enhancing stylization without weight tuning.
Kangfu Mei, Zhengzhong Tu, Mauricio Delbracio +3 authors
Smaller latent diffusion models often outperform larger ones in terms of sampling efficiency and generate high-quality results within limited inference budgets.
Yifan Xu, Xiao Liu, Xinghan Liu +9 authors
A self-critique pipeline enhances LLMs' mathematical problem-solving and language abilities through Math-Critique, rejective fine-tuning, and direct preference optimization, outperforming larger models.
Takayuki Hara, Tatsuya Harada
A method for generating 3D scenes from multimodal conditions, including partial images, top-view layouts, and text prompts, employs text-to-image models and neural radiance fields to produce diverse 3D scenes with reduced dataset requirements.
Zhiwen Fan, Wenyan Cong, Kairun Wen +9 authors
InstantSplat integrates dense stereo models and point-based representations to efficiently reconstruct large-scale scenes from sparse, pose-free images with high SSIM and low ATE.
Hao Li, Yang Zou, Ying Wang +7 authors
Empirical study examines the scaling properties of diffusion-based text-to-image models, identifying efficient architecture and dataset practices to improve performance and reduce cost.
Adrian Mirza, Nawaf Alampara, Sreekanth Kunchapu +25 authors
ChemBench, an automated evaluation framework, assesses the chemical reasoning of large language models against human experts, revealing their proficiency and shortcomings in specific tasks.
Shikai Li, Jianglin Fu, Kaiyuan Liu +3 authors
CosmicMan generates high-fidelity human images using a specialized data production paradigm and Decomposed-Attention-Refocusing framework within a text-to-image diffusion model.
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