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。
Tsendsuren Munkhdalai, Manaal Faruqui, Siddharth Gopal
A novel Infini-attention mechanism allows Transformer-based LLMs to handle infinitely long inputs with limited memory and compute, enabling efficient long-context modeling and fast inference.
50 篇论文 · 按点赞排序
Zhenghao Lin, Zhibin Gou, Yeyun Gong +8 authors
Rho-1, a novel language model using Selective Language Modeling, improves efficiency and performance by selectively training on useful tokens rather than all tokens in the corpus.
Keen You, Haotian Zhang, Eldon Schoop +5 authors
Ferret-UI, a multimodal large language model tailored for mobile UI screens, enhances understanding and interaction through region annotations and a comprehensive dataset of UI tasks, outperforming existing models including GPT-4V.
Elizaveta Goncharova, Anton Razzhigaev, Matvey Mikhalchuk +6 authors
The OmniFusion model, leveraging pretrained LLMs and visual adapters, demonstrates superior performance across multiple visual-language benchmarks compared to open-source alternatives.
Parishad BehnamGhader, Vaibhav Adlakha, Marius Mosbach +3 authors
LLM2Vec transforms decoder-only LLMs into effective text encoders using bidirectional attention, masked next token prediction, and unsupervised contrastive learning, achieving state-of-the-art performance on text embedding tasks.
Corby Rosset, Ching-An Cheng, Arindam Mitra +3 authors
A Direct Nash Optimization (DNO) algorithm is introduced for post-training large language models using preference feedback, achieving state-of-the-art performance with polynomial parameters.
Tianbao Xie, Danyang Zhang, Jixuan Chen +14 authors
OSWorld provides a scalable, real computer environment for evaluating multimodal agents across various operating systems and tasks, highlighting significant gaps in current LLM/VLM capabilities.
Ming Li, Taojiannan Yang, Huafeng Kuang +4 authors
ControlNet++ enhances text-to-image generation controllability through pixel-level cycle consistency optimization using an efficient noise-disturbed reward strategy.
Aleksandar Botev, Soham De, Samuel L Smith +59 authors
RecurrentGemma, using Griffin architecture, achieves excellent language performance with efficient memory usage and fewer training tokens compared to Gemma-2B.
Cheng-Ping Hsieh, Simeng Sun, Samuel Kriman +4 authors
RULER is a comprehensive benchmark that extends beyond simple retrieval tasks to evaluate long-context language models' performance on diverse tasks and context lengths.
Bo Peng, Daniel Goldstein, Quentin Anthony +24 authors
Eagle and Finch, improvements over RWKV-4, use multi-headed matrix-valued states and dynamic recurrence for enhanced expressivity and efficiency, trained on a multilingual corpus.
Yikang Shen, Zhen Guo, Tianle Cai +1 authors
JetMoE-8B, a cost-effective large language model with 8 billion parameters, achieves impressive performance using a Sparsely-gated Mixture-of-Experts architecture, demonstrating efficient use of resources and computational savings.
Haotian Zhang, Haoxuan You, Philipp Dufter +8 authors
Ferret-v2 improves on Ferret by incorporating flexible high-resolution grounding, multi-granularity visual encoding with DINOv2, and a three-stage training paradigm, leading to enhanced visual understanding and performance.
Shenghai Yuan, Jinfa Huang, Yujun Shi +6 authors
MagicTime enhances Text-to-Video synthesis by integrating real-world physics knowledge through metamorphic time-lapse videos, using the MagicAdapter, Dynamic Frames Extraction, and Magic Text-Encoder techniques.
Ruibo Liu, Jerry Wei, Fangyu Liu +8 authors
The success of AI models relies on the availability of large, diverse, and high-quality datasets, which can be challenging to obtain due to data scarcity, privacy concerns, and high costs. Synthetic data has emerged as a promising solution by generating artificial data that mimics real-world patterns. This paper provides an overview of synthetic data research, discussing its applications, challenges, and future directions. We present empirical evidence from prior art to demonstrate its effectiveness and highlight the importance of ensuring its factuality, fidelity, and unbiasedness. We emphasize the need for responsible use of synthetic data to build more powerful, inclusive, and trustworthy language models.
Xiaoyi Dong, Pan Zhang, Yuhang Zang +21 authors
InternLM-XComposer2-4KHD enhances high-resolution understanding capabilities in LVLMs up to 4K, surpassing existing models with dynamic patch configuration and consistent performance improvements.
Kanishk Gandhi, Denise Lee, Gabriel Grand +4 authors
Language models can learn complex problem-solving strategies via search by pretraining on streams of search generated by heuristic solvers and further improving with policy-based methods, solving previously unsolvable problems.
Vishaal Udandarao, Ameya Prabhu, Adhiraj Ghosh +5 authors
Studies reveal that multimodal models require exponentially more pretraining data for linear improvements in zero-shot performance and perform poorly on long-tailed datasets.
Jaidev Shriram, Alex Trevithick, Lingjie Liu +1 authors
RealmDreamer generates 3D scenes from text descriptions using 3D Gaussian Splatting, image-conditional diffusion models, and depth diffusion models, achieving high-quality and diverse scene synthesis without requiring video or multi-view data.
Yuxi Ren, Jie Wu, Yanzuo Lu +11 authors
ByteEdit, a feedback learning framework, enhances generative image editing through image reward models, pixel-level coherence, and adversarial learning, significantly improving quality, consistency, and speed compared to existing products.
Jing Gu, Yilin Wang, Nanxuan Zhao +7 authors
SwapAnything is a framework for precise object swapping in images using latent feature maps and appearance adaptation, offering better context preservation and concept adaptation than existing methods.
Makesh Narsimhan Sreedhar, Traian Rebedea, Shaona Ghosh +1 authors
The CantTalkAboutThis dataset improves language models' ability to stay on topic during conversations by including distractor turns, making them more resilient and coherent compared to general-purpose instruction-tuned models.
Shengding Hu, Yuge Tu, Xu Han +22 authors
MiniCPM, a family of small language models, demonstrates performance comparable to large models while introducing efficient scaling techniques and learning rate management.
Yuxi Xiao, Qianqian Wang, Shangzhan Zhang +4 authors
SpatialTracker estimates 3D point trajectories using monocular depth estimation and transformer updates, achieving superior video tracking performance with ARAP constraints and rigidity embeddings.
Jiacheng Zhang, Jie Wu, Yuxi Ren +9 authors
UniFL enhances diffusion models via feedback learning, improving visual quality, aesthetic appeal, and inference speed across various applications.
Michael Lutz, Arth Bohra, Manvel Saroyan +2 authors
Wilbur uses a differentiable ranking model and instruction synthesis to fine-tune a large language model for web agent tasks, achieving state-of-the-art results with text-only inputs and superior performance on the WebVoyager benchmark.
Sijun Tan, Xiuyu Li, Shishir Patil +5 authors
LLoCO method enhances large language models' ability to handle long contexts efficiently by combining context compression, retrieval, and parameter-efficient finetuning.
Gwanghyun Kim, Hayeon Kim, Hoigi Seo +2 authors
BeyondScene generates high-resolution human-centric scenes with detailed text descriptions and naturalness using pretrained diffusion models through a hierarchical enlargement process.
Bo He, Hengduo Li, Young Kyun Jang +5 authors
A memory bank is integrated into large multimodal models for efficient long-term video understanding, achieving state-of-the-art performance across various video tasks.
Zhen Qin, Songlin Yang, Weixuan Sun +4 authors
The introduction of an outer-product-based state expansion mechanism in hierarchical gated linear RNN (HGRN) enhances its performance and expressiveness without additional parameters, outperforming architectures like Mamba and LLaMa in various tasks.
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