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。
Zhihong Shao, Peiyi Wang, Qihao Zhu +6 authors
DeepSeekMath 7B improves mathematical reasoning through enhanced data pre-training and Group Relative Policy Optimization, achieving high scores on MATH benchmark without external tools.
50 篇论文 · 按点赞排序
Pei Zhou, Jay Pujara, Xiang Ren +7 authors
SELF-DISCOVER is a framework enabling LLMs to generate task-specific reasoning structures, enhancing performance on complex reasoning tasks with reduced computational cost.
Anian Ruoss, Grégoire Delétang, Sourabh Medapati +5 authors
A large-scale transformer model trained on a vast dataset of chess games outperforms traditional chess engines and other state-of-the-art models without domain-specific tweaks.
Yoad Tewel, Omri Kaduri, Rinon Gal +4 authors
ConsiStory achieves state-of-the-art text-to-image subject consistency without fine-tuning by using shared internal activations, attention blocks, and feature injection.
Junyou Li, Qin Zhang, Yangbin Yu +2 authors
A sampling-and-voting method enhances large language models' performance by increasing the number of agents, with effectiveness tied to task difficulty.
Wei Huang, Yangdong Liu, Haotong Qin +5 authors
BiLLM, a 1-bit post-training quantization scheme for pretrained LLMs, achieves high-accuracy inference with reduced computational and memory requirements.
David Grangier, Angelos Katharopoulos, Pierre Ablin +1 authors
The study evaluates various machine learning approaches under constrained inference and training budgets, demonstrating that hyper-networks and mixture of experts perform well with large pretraining budgets, while small models trained on importance sampled datasets are better with large specialization budgets.
Gilles Baechler, Srinivas Sunkara, Maria Wang +7 authors
ScreenAI, a vision-language model optimized with a flexible patching strategy, achieves state-of-the-art performance on UI and infographic tasks with 5B parameters.
Shihan Dou, Yan Liu, Haoxiang Jia +13 authors
StepCoder, a reinforcement learning framework, enhances code generation by breaking tasks into subtasks and optimizing only executed code segments, using the APPS+ dataset for training.
Xing Han Lù, Zdeněk Kasner, Siva Reddy
A benchmark and retrieval-inspired model for efficient conversational web navigation by digital agents across diverse real-world websites is introduced, highlighting challenges in generalization to unseen sites.
Jian Xie, Kai Zhang, Jiangjie Chen +5 authors
TravelPlanner is a new benchmark for evaluating the planning capabilities of language agents in complex, real-world travel planning scenarios, demonstrating that current models still struggle with information gathering and constraint management.
Shangmin Guo, Biao Zhang, Tianlin Liu +9 authors
Online AI feedback improves direct alignment from preferences methods by providing real-time feedback from a model, outperforming both offline DAP and RLHF methods.
Jongho Park, Jaeseung Park, Zheyang Xiong +5 authors
SSMs, including Mamba, offer competitive ICL capabilities in standard regression tasks but lag in non-standard retrieval; a hybrid model combining Mamba with attention blocks enhances ICL performance.
Sihao Hu, Tiansheng Huang, Ling Liu
Pok\'eLLMon, an LLM-embodied agent, achieves human-parity in tactical battle games through in-context reinforcement learning, knowledge-augmented generation, and consistent action generation, demonstrating human-like strategies and winning rates in various competitions.
Quan Sun, Jinsheng Wang, Qiying Yu +4 authors
EVA-CLIP-18B, a 18-billion parameter open-source CLIP model, achieves exceptional zero-shot performance across multiple image classification benchmarks with a relatively small dataset.
Zane Durante, Bidipta Sarkar, Ran Gong +17 authors
The proposed Interactive Agent Foundation Model uses a multi-task agent training paradigm to develop versatile AI agents capable of performing well across diverse domains and tasks.
Jiaxiang Tang, Zhaoxi Chen, Xiaokang Chen +3 authors
A novel Large Multi-View Gaussian Model (LGM) generates high-resolution 3D models from text or a single image using multi-view Gaussian features and an asymmetric U-Net backbone, achieved with a multi-view diffusion model.
Fuzhao Xue, Zian Zheng, Yao Fu +4 authors
OpenMoE, a series of open-sourced MoE-based decoder-only LLMs, demonstrates improved cost-effectiveness over dense LLMs and reveals insights into MoE routing mechanisms, including context-independent specialization, early routing learning, and drop-towards-the-end issues.
Weiming Ren, Harry Yang, Ge Zhang +4 authors
ConsistI2V, a diffusion-based method, enhances visual consistency in image-to-video generation by using spatiotemporal attention and noise initialization from the initial frame's low-frequency band.
Jiawei Wang, Yuchen Zhang, Jiaxin Zou +4 authors
Boximator enhances video synthesis by adding fine-grained motion control through box constraints, achieving superior video quality and alignment compared to base models.
Quentin Anthony, Yury Tokpanov, Paolo Glorioso +1 authors
BlackMamba, a novel architecture combining Mamba SSM and MoE models, achieves competitive performance in language modeling and improved inference and training efficiency.
Samy Jelassi, David Brandfonbrener, Sham M. Kakade +1 authors
Transformers outperform generalized state space models (GSSMs) in copying and retrieving information from context due to their ability to handle variable-length sequences without a fixed-size latent state.
Junhong Shen, Neil Tenenholtz, James Brian Hall +2 authors
A model-agnostic framework for learning custom input tags enables general LLMs to solve specialized domain tasks with zero-shot generalization and improved performance.
Zhuoyang Zhang, Han Cai, Song Han
EfficientViT-SAM combines SAM's prompt encoder and mask decoder with EfficientViT as the image encoder, achieving significant speedup without losing performance.
Liang Wang, Nan Yang, Xiaolong Huang +3 authors
Multilingual E5 text embedding models are released in different sizes and include an instruction-tuned variant, achieving performance comparable to state-of-the-art English-only models.
Chandan Singh, Jeevana Priya Inala, Michel Galley +2 authors
Large language models offer new opportunities in interpretable machine learning, enabling more complex explanations but also introducing challenges like hallucinations and high computational costs.
Jordan Juravsky, Bradley Brown, Ryan Ehrlich +3 authors
Hydragen improves LLM inference efficiency by decomposing attention operations into shared prefix and unique suffix computations, reducing memory reads and accelerating matrix multiplications, achieving significant throughput improvements.
Berivan Isik, Natalia Ponomareva, Hussein Hazimeh +3 authors
In transfer learning for machine translation, the size and alignment of pretraining data significantly impact downstream performance as measured by cross-entropy and BLEU score.
Tianqi Liu, Zhen Qin, Junru Wu +9 authors
The LiPO framework, which addresses LM alignment as a listwise ranking problem, outperforms DPO and SLiC by effectively learning from ranked lists of responses.
Shiyuan Yang, Liang Hou, Haibin Huang +5 authors
Direct-a-Video enables independent control of object motion and camera movement in text-to-video generation using spatial and temporal cross-attention modulation.
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