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
John Thickstun, David Hall, Chris Donahue +1 authors
A method for anticipatory generation in music control tasks, using interleaved event and control sequences, matches the performance of autoregressive models and produces high-quality accompaniments.
Daniel McKee, Justin Salamon, Josef Sivic +1 authors
A text-synthesis approach using a large-scale language model and a new trimodal model enhances video-to-music recommendation with natural language inputs, outperforming prior methods in retrieval accuracy.
Sarah J. Zhang, Samuel Florin, Ariel N. Lee +12 authors
Large language models, including GPT-3.5 and GPT-4, can effectively solve a significant portion of MIT Mathematics and EECS curriculum questions, with GPT-4 achieving near-perfect performance through prompt engineering and fine-tuning.
Enric Boix-Adsera, Etai Littwin, Emmanuel Abbe +2 authors
The study examines incremental learning in transformers, proving that the difference in rank between trained and initial weights increases progressively.
Ian Huang, Vrishab Krishna, Omoruyi Atekha +1 authors
A system leverages foundation models to generate stylized 3D scene assets from descriptions, demonstrating higher fidelity than baselines through human evaluations.
Irene Solaiman, Zeerak Talat, William Agnew +15 authors
A framework is provided to evaluate the social impacts of generative AI systems across various modalities, both in terms of individual system and broader societal contexts.
Saminda Abeyruwan, Alex Bewley, Nicholas M. Boffi +9 authors
Two methodologies, Model Predictive Control with constrained trajectory optimization and Reinforcement Learning with zeroth-order optimization, are compared for high-speed object catching in robotics, highlighting performance metrics and potential integration strategies.
Ahmet Iscen, Mathilde Caron, Alireza Fathi +1 authors
RECO, a retrieval-enhanced technique using a lightweight fusion transformer, improves CLIP's performance on fine-grained tasks by refining embeddings with cross-modal information from external memory.
Orr Zohar, Shih-Cheng Huang, Kuan-Chieh Wang +1 authors
A new benchmark is introduced for efficient zero-shot evaluation of pre-trained vision-language models without requiring a downstream dataset, focusing on model selection based on text descriptions.
Yujin Tang, Wenhao Yu, Jie Tan +3 authors
A novel approach using-foot contact patterns interfaces natural language commands with LLMs to control quadrupedal robots, achieving high success rates and versatility in locomotion tasks.
Allan Jabri, Sjoerd van Steenkiste, Emiel Hoogeboom +2 authors
A video diffusion architecture adapted for 3D scene generation improves rendering quality and scalability while maintaining object-level editing capabilities.
Paul Couairon, Clément Rambour, Jean-Emmanuel Haugeard +1 authors
VidEdit is a zero-shot text-based video editing method combining atlas-based and pre-trained diffusion models with panoptic segmenters and edge detectors to achieve strong temporal and spatial consistency.
Haotong Lin, Qianqian Wang, Ruojin Cai +4 authors
A novel scene representation with temporal step function encoding enables reconstruction of time-varying 3D models from internet photos, achieving state-of-the-art view synthesis with independent control over viewpoint, time, and illumination.
Mingqiu Wang, Izhak Shafran, Hagen Soltau +4 authors
A joint speech and language model using a Speech2Text adapter and CTC-based blank-filtering improves dialog state tracking and ASR performance in the speech domain.
Gyungin Shin, Weidi Xie, Samuel Albanie
The paper introduces a task called automatic table verification (AutoTV) to ensure the accuracy of numerical data in tables through source cross-referencing, proposing a benchmark called arXiVeri and evaluating table and cell matching using LLMs.
Prannay Kaul, Weidi Xie, Andrew Zisserman
A novel open-vocabulary object detection method uses language descriptions and image exemplars to detect objects unseen during training, achieving superior performance in multi-modal classification.
Zhijing Jin, Jiarui Liu, Zhiheng Lyu +5 authors
A benchmark dataset Corr2Cause tests the causal inference skills of large language models, revealing their limitations and highlighting the need for improved generalizability.
Zhi-Hao Lin, Bohan Liu, Yi-Ting Chen +4 authors
UrbanIR computes an inverse graphics representation from video for realistic relighting and object insertion in urban scenes with novel lighting conditions.
Ziniu Hu, Ahmet Iscen, Chen Sun +5 authors
AUTONOMOUS INFORMATION SEEKING VISUAL QUESTION ANSWERING FRAMEWORK (AVIS) leverages LLMs to strategize and utilize external tools, achieving state-of-the-art results in knowledge-intensive VQA.
Sagar Vaze, Nicolas Carion, Ishan Misra
GeneCIS benchmark evaluates models' adaptability to various similarity conditions, revealing that existing methods struggle and proposing an automatic mining solution to enhance zero-shot performance.
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