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。
49 篇论文 · 按点赞排序
Adyasha Maharana, Dong-Ho Lee, Sergey Tulyakov +3 authors
A pipeline combining LLMs and human annotation generates long-term dialogues to evaluate model performance in understanding lengthy conversations and temporal dynamics.
Lu Ye, Ze Tao, Yong Huang +1 authors
ChunkAttention optimizes self-attention by sharing key/value tensors for identical system prompts across multiple requests, enhancing memory usage and reducing latency.
Simran Arora, Sabri Eyuboglu, Michael Zhang +6 authors
BASED, a hybrid architecture combining linear and sliding window attention, improves language model efficiency without sacrificing recall quality.
Dejan Grubisic, Chris Cummins, Volker Seeker +1 authors
Priority Sampling enhances code generation by producing unique, high-confidence samples and improving performance over Nucleus Sampling.
Yuxian Gu, Li Dong, Yaru Hao +3 authors
A theory for optimal language model learning is proposed, revealing properties of the learning process through a data compression objective and improving coefficients in the scaling law, with experimental validation.
Mosh Levy, Alon Jacoby, Yoav Goldberg
Studies the impact of varying input lengths on Large Language Models, revealing performance degradation at shorter lengths and introducing a QA reasoning framework to assess these effects.
Haoning Wu, Hanwei Zhu, Zicheng Zhang +11 authors
Co-Instruct, an open-source open-ended visual quality comparer trained on Co-Instruct-562K dataset, outperforms state-of-the-art models and its teacher GPT-4V on image quality assessment tasks.
Jianguo Zhang, Tian Lan, Rithesh Murthy +13 authors
AgentOhana standardizes and unifies agent trajectories for consistent training, enhancing performance with the xLAM-v0.1 large action model across diverse data sources.
Jianbin Zheng, Minghui Hu, Zhongyi Fan +4 authors
Trajectory Consistency Distillation (TCD) improves text-to-image synthesis by addressing errors in consistency models, leading to higher image quality and detail at low numerical flow evaluations.
Xuanyi Li, Daquan Zhou, Chenxu Zhang +3 authors
A new benchmark evaluates video generation fidelity by assessing the geometric consistency of reconstructed 3D models.
Mikayel Samvelyan, Sharath Chandra Raparthy, Andrei Lupu +9 authors
Rainbow Teaming, an approach to generating diverse adversarial prompts, enhances the robustness of large language models across multiple domains by improving their safety without compromising general performance.
Yazhou Xing, Yingqing He, Zeyue Tian +2 authors
A framework that integrates pre-trained video and audio generation models via a shared latent space to achieve joint and cross-modal content creation outperforms separate approaches.
Xianghui Yang, Yan Zuo, Sameera Ramasinghe +3 authors
ViewFusion, an auto-regressive, training-free algorithm, integrates with pre-trained diffusion models to synthesize consistent novel views by leveraging interpolated denoising from known views.
German Barquero, Sergio Escalera, Cristina Palmero
FlowMDM is a diffusion-based model for generating long, continuous human motions guided by changing textual descriptions, achieving high accuracy and smoothness with novel positional encodings and attention mechanisms.
Kinjal Basu, Ibrahim Abdelaziz, Subhajit Chaudhury +7 authors
API-BLEND is introduced as a large dataset for training and benchmarking tool-augmented Large Language Models by simulating real-world API tasks.
Santiago Castro, Amir Ziai, Avneesh Saluja +2 authors
A new framework enhances compositional language encoding in Vision-Language Models while maintaining or improving object recognition, achieving significant gains on benchmark tests.
Dave Epstein, Ben Poole, Ben Mildenhall +2 authors
Unsupervised disentanglement of 3D scenes into component objects using pretrained text-to-image models and joint optimization of NeRFs and layouts.
Jiaqi Lin, Zhihao Li, Xiao Tang +8 authors
VastGaussian addresses limitations in NeRF-based methods for large scene reconstruction by partitioning scenes and decoupling appearance modeling, achieving high-quality and real-time rendering.
Daiqing Li, Aleks Kamko, Ehsan Akhgari +3 authors
Enhancements to noise schedules, aspect ratio accommodation, and alignment with human perceptual preferences improve the aesthetic quality of text-to-image generative models, leading to state-of-the-art 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