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
Jiaming Li, Longze Chen, Ze Gong +5 authors
PACS, a novel RLVR framework, reformulates RLVR as a supervised learning task, improving stability and efficiency in training large language models for reasoning tasks.
Shuzhou Yang, Xiaoyu Li, Xiaodong Cun +4 authors
A novel Diffusion Transformer pipeline automates video compositing by adaptively injecting identity and motion information, maintaining consistency and enabling user customization.
Tianjian Li, Yiming Zhang, Ping Yu +5 authors
DARLING, a diversity-aware reinforcement learning framework, enhances both the quality and diversity of large language model outputs across various tasks.
Andrei Semenov, Matteo Pagliardini, Martin Jaggi
A comprehensive evaluation of recent optimization techniques for Large Language Models provides guidance on selecting the best optimizer for different pretraining scenarios.
Daniela Gottesman, Alon Gilae-Dotan, Ido Cohen +4 authors
LMEnt is a suite for analyzing knowledge acquisition in language models during pretraining, providing annotated corpora, retrieval methods, and pretrained models to study knowledge representations and learning dynamics.
Delong Chen, Theo Moutakanni, Willy Chung +4 authors
The Vision Language World Model (VLWM) achieves state-of-the-art performance in visual planning by integrating language-based world modeling, action policy learning, and dynamics modeling with semantic and temporal abstraction.
Chenlu Ye, Zhou Yu, Ziji Zhang +5 authors
PROF, a data curation method, improves reinforcement learning for mathematical reasoning by harmonizing process and outcome rewards, enhancing both final accuracy and intermediate reasoning quality.
Monte Hoover, Vatsal Baherwani, Neel Jain +7 authors
Dynamic guardian models evaluate text based on user-defined policies, offering fast and accurate detection of both static harms and free-form policy violations.
Evan King, Adam Sabra, Manjunath Kudlur +2 authors
Monolingual ASR models trained on a balanced mix of high-quality, pseudo-labeled, and synthetic data outperform multilingual models for small model sizes, achieving superior error rates and enabling on-device ASR for underrepresented languages.
Jaewoo Ahn, Junseo Kim, Heeseung Yun +4 authors
FlashAdventure benchmark and COAST framework improve GUI agents' performance in completing full story arcs in Flash-based adventure games by addressing the observation-behavior gap.
Daria Kryvosheieva, Saba Sturua, Michael Günther +2 authors
Jina-code-embeddings uses an autoregressive backbone pre-trained on text and code to generate embeddings for code retrieval, question-answering, and similarity identification.
Orion Weller, Michael Boratko, Iftekhar Naim +1 authors
Vector embeddings face theoretical limitations in handling even simple queries, as demonstrated by a new dataset that shows state-of-the-art models fail due to the single vector paradigm.
Shunian Chen, Hejin Huang, Yexin Liu +10 authors
TalkVid, a large-scale, high-quality, and diverse dataset, improves audio-driven talking head synthesis by enhancing generalization across human diversity and revealing subgroup performance disparities.
Xuechao Zou, Shun Zhang, Xing Fu +6 authors
Face-MoGLE, a novel framework using Diffusion Transformers, achieves high-quality, controllable face generation through semantic-decoupled latent modeling, expert specialization, and dynamic gating.
Yuan Xie, Tianshui Chen, Zheng Ge +1 authors
Video-MTR, a reinforced multi-turn reasoning framework, improves long-form video understanding by iteratively selecting key segments and comprehending questions, outperforming existing methods in accuracy and efficiency.
Venkatesh Mishra, Amir Saeidi, Satyam Raj +5 authors
The IRMA framework improves the reliability and consistency of large language models in dynamic environments by reformulating user queries with domain rules and tool suggestions.
Guangzeng Han, Weisi Liu, Xiaolei Huang
Genetic Prompt enhances synthetic data quality and diversity in NLP by combining genetic algorithms with LLMs, improving downstream model performance.
Kaiyue Wen, David Hall, Tengyu Ma +1 authors
A systematic study reveals that fair comparisons of deep learning optimizers require rigorous hyperparameter tuning and evaluations across various model scales and data-to-model ratios, showing that matrix-based optimizers like Muon and Soap offer diminishing speedups as model size increases.
Dong She, Siming Fu, Mushui Liu +4 authors
MOSAIC framework enhances multi-subject image generation by ensuring precise semantic alignment and orthogonal feature disentanglement, achieving high fidelity even with multiple references.
Peter Belcak, Pavlo Molchanov
Universal Deep Research (UDR) is a flexible system that allows users to customize deep research strategies using any language model without additional training or fine-tuning.
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