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
Yongchao Zhou, Uri Alon, Xinyun Chen +3 authors
Transformers can extrapolate to longer test sequences with the right data format and position encodings, but length generalization is sensitive to initialization and data order.
Lichang Chen, Chen Zhu, Davit Soselia +6 authors
An evaluation protocol and reward model improvement address length bias in Reinforcement Learning from Human Feedback for LLMs.
Siddharth Karamcheti, Suraj Nair, Ashwin Balakrishna +3 authors
The paper provides standardized evaluations and investigations into visually-conditioned language models (VLMs), exploring factors affecting their performance and offering resources like a unified evaluation framework and optimized training code.
David Ruhe, Jonathan Heek, Tim Salimans +1 authors
Rolling Diffusion enhances temporal data generation by progressively corrupting frames with increasing noise, outperforming standard diffusion in complex tasks like video prediction and fluid dynamics forecasting.
Girish Sastry, Lennart Heim, Haydn Belfield +16 authors
Computing power, or "compute," is crucial for the development and deployment of artificial intelligence (AI) capabilities. As a result, governments and companies have started to leverage compute as a means to govern AI. For example, governments are investing in domestic compute capacity, controlling the flow of compute to competing countries, and subsidizing compute access to certain sectors. However, these efforts only scratch the surface of how compute can be used to govern AI development and deployment. Relative to other key inputs to AI (data and algorithms), AI-relevant compute is a particularly effective point of intervention: it is detectable, excludable, and quantifiable, and is produced via an extremely concentrated supply chain. These characteristics, alongside the singular importance of compute for cutting-edge AI models, suggest that governing compute can contribute to achieving common policy objectives, such as ensuring the safety and beneficial use of AI. More precisely, policymakers could use compute to facilitate regulatory visibility of AI, allocate resources to promote beneficial outcomes, and enforce restrictions against irresponsible or malicious AI development and usage. However, while compute-based policies and technologies have the potential to assist in these areas, there is significant variation in their readiness for implementation. Some ideas are currently being piloted, while others are hindered by the need for fundamental research. Furthermore, naive or poorly scoped approaches to compute governance carry significant risks in areas like privacy, economic impacts, and centralization of power. We end by suggesting guardrails to minimize these risks from compute governance.
Fei Deng, Qifei Wang, Wei Wei +2 authors
Proximal Reward Difference Prediction enables stable large-scale reward finetuning of diffusion models in vision, outperforming RL-based methods on complex prompts.
Nadav Schneider, Niranjan Hasabnis, Vy A. Vo +10 authors
Domain-specific fine-tuning of MonoCoder on MPI-related code significantly improves the generation of MPI-based parallel programs compared to general-purpose models.
Catherine Yeh, Gonzalo Ramos, Rachel Ng +2 authors
GhostWriter, an AI-enhanced writing system, learns users' writing style and provides personalization and control through both implicit learning and explicit annotations.
Yin Fang, Kangwei Liu, Ningyu Zhang +7 authors
ChatCell leverages natural language and sequence generation to enhance single-cell analysis, improving accessibility and performance in this field.
Luke Melas-Kyriazi, Iro Laina, Christian Rupprecht +4 authors
IM-3D improves text-to-3D generation by using video generators and Gaussian splatting to produce high-quality 3D outputs efficiently with minimized geometric inconsistencies.
Jakub Krajewski, Jan Ludziejewski, Kamil Adamczewski +9 authors
MoE models reduce computational cost through fine-grained scaling controlled by a hyperparameter called granularity, outperforming dense Transformers across varying scales and budgets.
Raunaq Bhirangi, Chenyu Wang, Venkatesh Pattabiraman +4 authors
Hierarchical State-Space Models (HiSS) stack structured state-space models to create a temporal hierarchy, outperforming sequence models like causal Transformers, LSTMs, S4, and Mamba in continuous sequential prediction across real-world sensor datasets.
Tiffany Tseng, Ruijia Cheng, Jeffrey Nichols
Keyframer, a natural language-driven animation design tool, allows users to create and refine SVG animations through prompting and direct editing, informed by professional insights and user studies.
Ta-Ying Cheng, Matheus Gadelha, Thibault Groueix +4 authors
Continuous 3D Words enable fine-grained control of 3D attributes in text-to-image models, allowing users to adjust image generation through continuous sliders alongside text prompts.
Fedor Borisyuk, Mingzhou Zhou, Qingquan Song +31 authors
LiRank, a large-scale ranking framework at LinkedIn, integrates advanced modeling techniques and optimization methods to enhance performance in feed ranking, job recommendations, and ads CTR prediction.
John Hewitt, Sarah Chen, Lanruo Lora Xie +3 authors
Model editing using canonical examples improves performance with LoRA and sense finetuning on Pythia and Backpack language models.
Zhenglin Zhou, Fan Ma, Hehe Fan +1 authors
HeadStudio uses 3D Gaussian splatting and FLAME representation to generate high-quality, animated digital avatars from text prompts with real-time rendering capabilities.
Xiaoyu Zhou, Xingjian Ran, Yajiao Xiong +5 authors
GALA3D is a generative framework for 3D content that uses large language models and conditioned diffusion for layout-guided, compositional scene generation, ensuring high fidelity and realistic object interactions.
Louis Castricato, Nathan Lile, Suraj Anand +3 authors
A novel method called Direct Principle Feedback fine-tunes language models to become controllable at inference time, demonstrated on the Pink Elephant Problem.
Amir Zandieh, Insu Han, Vahab Mirrokni +1 authors
A novel KV cache compression technique, SubGen, improves memory efficiency and performance in long-context token generation for large language models by using_online clustering and sampling.
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