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。
Zhengzhong Liu, Aurick Qiao, Willie Neiswanger +25 authors
LLM360 initiative promotes full transparency and reproducibility in LLM training by open-sourcing training code, data, model checkpoints, and intermediate results.
50 篇论文 · 按点赞排序
Julian D. Parker, Janne Spijkervet, Katerina Kosta +6 authors
A transformer-based, non-autoregressive model generates musically coherent audio by responding to context, achieving audio quality comparable to text-conditioned models.
Bingbin Liu, Sebastien Bubeck, Ronen Eldan +5 authors
A synthetic dataset of math problems paired with solutions enables small language models to achieve high accuracy in mathematical reasoning, outperforming larger models.
Róbert Csordás, Piotr Piękos, Kazuki Irie +1 authors
SwitchHead reduces memory and compute for Transformers by using Mixture-of-Experts layers, matching performance with significant speedup.
Luka Ribar, Ivan Chelombiev, Luke Hudlass-Galley +3 authors
SparQ Attention reduces memory bandwidth requirements in LLM attention blocks, enhancing inference throughput without accuracy loss.
Mengyang Feng, Jinlin Liu, Kai Yu +11 authors
DreaMoving is a diffusion-based framework that generates high-quality, customized human dance videos by controlling motion and preserving identity.
Wenyi Hong, Weihan Wang, Qingsong Lv +8 authors
CogAgent, a visual language model with strong GUI understanding and navigation capabilities, outperforms LLM-based methods in both PC and Android GUI tasks using only screenshots.
Avi Singh, John D. Co-Reyes, Rishabh Agarwal +35 authors
Self-training with scalar feedback using expectation-maximization improves language model performance on tasks with verifiable answers, reducing reliance on human data.
Alexander Borzunov, Max Ryabinin, Artem Chumachenko +5 authors
A decentralized system pools idle compute resources to enable efficient inference and fine-tuning of large language models over a consumer-grade network, showcasing significant performance improvements.
Tianxing Wu, Chenyang Si, Yuming Jiang +2 authors
FreeInit addresses the temporal consistency and unnatural dynamics issues in diffusion-based video generation by refining spatial-temporal low-frequency components during inference, improving subject appearance and consistency.
Fangfu Liu, Diankun Wu, Yi Wei +2 authors
Sherpa3D leverages 2D diffusion models and coarse 3D priors to generate high-fidelity, geometrically consistent, and diverse 3D content from text prompts.
Agrim Gupta, Lijun Yu, Kihyuk Sohn +6 authors
A transformer-based diffusion model using causal encoding and window attention generates high-resolution, photorealistic videos, achieving state-of-the-art performance without classifier-free guidance and includes a cascade for text-to-video generation.
Xiang Wang, Shiwei Zhang, Han Zhang +4 authors
VideoLCM framework leverages consistency models for efficient video synthesis with minimal sampling steps, achieving high fidelity and temporal consistency.
Kaiwen Zhang, Yifan Zhou, Xudong Xu +2 authors
DiffMorpher enhances smooth image interpolation using diffusion models by capturing image semantics with LoRA and applying attention interpolation techniques.
Manish Bhatt, Sahana Chennabasappa, Cyrus Nikolaidis +18 authors
CyberSecEval is a benchmark that evaluates the cybersecurity of large language models, focusing on insecure code generation and compliance with cyberattacks, providing insights for enhancing model security.
Ji Lin, Hongxu Yin, Wei Ping +7 authors
Enhanced pre-training methods improve visual language models by balancing LLM freezing, interleaved data use, and text-only instruction re-blending, leading to superior performance across benchmarks.
Haoran Wei, Lingyu Kong, Jinyue Chen +6 authors
Vary enhances Vision-Language Models with a new scalable and fine-grained vision vocabulary through autoregressive generation and integration, improving performance in tasks like document OCR and chart understanding.
Maitreya Patel, Changhoon Kim, Sheng Cheng +2 authors
ECLIPSE is a parameter and data-efficient contrastive learning method for Text-to-Image models that outperforms larger models with fewer resources and maintains high preference scores.
Jack Urbanek, Florian Bordes, Pietro Astolfi +3 authors
A human-annotated dataset with dense captions significantly improves the evaluation and fine-tuning of vision-language models beyond standard benchmarks.
Shuyang Sun, Runjia Li, Philip Torr +2 authors
A recurrent framework improves open-vocabulary image segmentation by filtering irrelevant texts and enhancing mask quality without fine-tuning.
Sicheng Mo, Fangzhou Mu, Kuan Heng Lin +4 authors
FreeControl provides a training-free method to control text-to-image diffusion models with multiple conditions, architectures, and checkpoints, demonstrating competitive performance to training-based approaches.
Lihe Ding, Shaocong Dong, Zhanpeng Huang +5 authors
BiDiff combines 3D and 2D diffusion processes with bidirectional guidance to generate high-quality, diverse, and scalable 3D models efficiently.
Lior Yariv, Omri Puny, Natalia Neverova +2 authors
A new Mosaic-SDF representation for 3D shapes enables efficient, parameter-efficient, and parallelizable generation of 3D shapes using generative flow models.
Kaijie Zhu, Qinlin Zhao, Hao Chen +2 authors
PromptBench is a unified library designed to evaluate large language models with key components for prompt engineering, adversarial attacks, and flexible evaluation protocols.
Raviteja Anantha, Tharun Bethi, Danil Vodianik +1 authors
Context Tuning enhances Retrieval Augmented Generation by improving context retrieval, tool retrieval, and LLM-based planner accuracy using smart context signals and fusion techniques.
Jiarui Xu, Xingyi Zhou, Shen Yan +5 authors
The developed vision-language model can perform location-aware tasks by handling inputs and outputs as pixel coordinates or locations, achieving state-of-the-art performance on several benchmarks.
Roya Firoozi, Johnathan Tucker, Stephen Tian +12 authors
Pretrained foundation models, trained on internet-scale data, offer improved generalization and can enhance various aspects of robot autonomy, including perception, decision-making, and control, despite challenges in robotics-specific data, safety, and real-time execution.
Yue Yang, Fan-Yun Sun, Luca Weihs +11 authors
Holodeck automatically generates diverse and semantically detailed 3D environments using a large language model and 3D asset optimization, advancing Embodied AI training.
Arnav Chavan, Nahush Lele, Deepak Gupta
This paper presents a new compression technique for Large Language Models (LLMs) using reduced order modeling and low-rank decomposition that operates in a layer-wise manner without a GPU, achieving better results than structured pruning.
Yuqing Yang, Ethan Chern, Xipeng Qiu +2 authors
A paper proposes a framework and metrics to enhance the honesty of large language models by training them to accurately assess and reveal their knowledge limits.
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