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
Ziyin Zhang, Zihan Liao, Hang Yu +2 authors
F2LLM, a suite of large language models, achieves high embedding performance with efficient fine-tuning from foundation models using open-source datasets.
Yangzhou Liu, Yue Cao, Hao Li +13 authors
Sequential Diffusion Language Model (SDLM) enhances pre-trained autoregressive language models by adaptively determining generation length and maintaining KV-cache compatibility, achieving high efficiency and throughput.
Fanding Huang, Guanbo Huang, Xiao Fan +7 authors
Re-examining the exploration-exploitation trade-off in Reinforcement Learning for Verifiable Rewards through hidden-state analysis reveals opportunities for simultaneous enhancement using Effective Rank and its derivatives, leading to improved performance in diverse benchmarks.
Ali Hatamizadeh, Syeda Nahida Akter, Shrimai Prabhumoye +5 authors
RLP, an information-driven reinforcement pretraining objective, enhances reasoning models by integrating exploration into pretraining, leading to significant performance improvements across various benchmarks.
Yang Shi, Yuhao Dong, Yue Ding +23 authors
RealUnify evaluates the bidirectional synergy between understanding and generation in unified multimodal models, revealing that current models lack effective integration despite architectural unification.
Junlin Han, Shengbang Tong, David Fan +4 authors
LLMs develop visual priors during language pre-training, which can be leveraged for vision tasks with minimal additional data, and these priors are composed of separable perception and reasoning components.
Wentao Zhang, Yang Young Lu, Yuntian Deng
Interactive Training is a framework that allows real-time, feedback-driven intervention during neural network training, improving stability and adaptability.
Junyu Chen, Wenkun He, Yuchao Gu +12 authors
DC-VideoGen accelerates video generation by adapting pre-trained diffusion models to a deep compression latent space, reducing inference latency and enabling high-resolution video generation.
Shanghua Gao, Richard Zhu, Pengwei Sui +8 authors
ToolUniverse is an ecosystem that standardizes and integrates tools, models, and data for AI scientists, enabling automated refinement, creation, and composition of workflows.
Nicolas Boizard, Hippolyte Gisserot-Boukhlef, Kevin El-Haddad +2 authors
Reasoning models enhance performance across various tasks, surpassing instruction fine-tuned models in reasoning-intensive and open-ended tasks, despite higher computational costs.
Chetwin Low, Weimin Wang, Calder Katyal
Ovi is a unified audio-video generation model using twin-DiT modules with blockwise cross-modal fusion, enabling natural synchronization and high-quality multimodal outputs.
Penghao Wu, Yushan Zhang, Haiwen Diao +3 authors
Visual Jigsaw, a self-supervised reinforcement learning framework, enhances multimodal large language models' visual understanding through a permutation task without additional annotations or generative components.
Long Xing, Xiaoyi Dong, Yuhang Zang +6 authors
CapRL, a novel reinforcement learning framework, enhances image captioning by using a vision-free language model to evaluate caption quality through multiple-choice questions, leading to improved performance across benchmarks.
Minki Kang, Wei-Ning Chen, Dongge Han +5 authors
Agent Context Optimization (ACON) compresses context in large language models for efficient long-horizon tasks by analyzing failure cases and distilling the compressor into smaller models.
Yida Xue, Mingjun Mao, Xiangyuan Ru +9 authors
OceanGym is a benchmark for underwater embodied agents using Multi-modal Large Language Models to address challenges in perception, planning, and adaptability in harsh ocean environments.
Yanzhe Chen, Kevin Qinghong Lin, Mike Zheng Shou
Code2Video generates educational videos using a code-centric agent framework, improving coherence and interpretability compared to direct code generation.
Xin Xu, Cliveb AI, Kai Yang +4 authors
TFPI, a simple adaptation to RLVR, improves performance and reduces token usage by discarding thinking content during training, accelerating RL convergence and achieving higher accuracy with less computational cost.
Xin Luo, Jiahao Wang, Chenyuan Wu +6 authors
A specialized reward model, EditScore, enables effective reinforcement learning for instruction-guided image editing by providing a high-fidelity reward signal.
Xueliang Zhao, Wei Wu, Jian Guan +2 authors
PromptCoT 2.0 uses an EM loop to generate harder and more diverse synthetic prompts, improving reasoning capabilities in large language models through self-play and supervised fine-tuning.
Paul Teiletche, Quentin Macé, Max Conti +4 authors
ModernVBERT, a compact vision-language encoder, outperforms larger models in document retrieval by optimizing attention masking, image resolution, modality alignment, and contrastive objectives.
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