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
Penghui Qi, Zichen Liu, Tianyu Pang +3 authors
The AnytimeReasoner framework optimizes token efficiency and reasoning flexibility for large language models by introducing verifiable dense rewards and a decoupled policy optimization technique.
Zebin You, Shen Nie, Xiaolu Zhang +5 authors
A diffusion-based Multimodal Large Language Model (LLaDA-V) with integrated visual instruction tuning performs competitively on multimodal tasks and outperforms existing models in multimodal understanding.
Chenyu Zheng, Xinyu Zhang, Rongzhen Wang +5 authors
Maximal Update Parametrization (μP) is extended to diffusion Transformers, demonstrating efficient hyperparameter transferability and reduced tuning costs across various models and tasks.
Wei Liu, Ruochen Zhou, Yiyun Deng +5 authors
RL-based reward shaping methods, particularly LASER-D, enhance reasoning efficiency and performance in large reasoning models by dynamically adapting to difficulty and reducing redundancy.
Ziyu Liu, Yuhang Zang, Yushan Zou +6 authors
Visual Agentic Reinforcement Fine-Tuning enhances Large Vision-Language Models for flexible image manipulation and web-based reasoning, outperforming existing models on multi-modal benchmarks.
Yen-Chen Wu, Feng-Ting Liao, Meng-Hsi Chen +3 authors
The Latent Flow Transformer (LFT) compresses layers by replacing them with learned transport operators using flow matching and Flow Walking, showing improved performance over layer skipping and reducing the gap between autoregressive and flow-based models.
Elisei Rykov, Kseniia Petrushina, Kseniia Titova +3 authors
A new method, Through the Looking Glass (TLG), uses Large Vision-Language Models and Transformer-based encoders to improve image common sense consistency assessment on the WHOOPS! and WEIRD datasets.
Seungjun Oh, Younggeun Lee, Hyejin Jeon +1 authors
A new framework, 3D-4D Gaussian Splatting, improves computational efficiency in dynamic 3D scene reconstruction by adaptively representing static and dynamic regions, reducing parameter count and training time.
Chan-Jan Hsu, Davide Buffelli, Jamie McGowan +4 authors
A concurrent reasoning paradigm using a single LLM, called Group Think, enhances reasoning quality and reduces latency by allowing multiple reasoning threads to dynamically collaborate at the token level.
Chengqi Duan, Rongyao Fang, Yuqing Wang +5 authors
GoT-R1 enhances visual generation by using reinforcement learning to improve semantic-spatial reasoning, outperforming existing models on complex compositional tasks.
Dario Garcia-Gasulla, Jordi Bayarri-Planas, Ashwin Kumar Gururajan +10 authors
Aloe Beta models improve open-source medical Large Language Models through enhanced preprocessing, Direct Preference Optimization, Retrieval-Augmented Generation, and a rigorous evaluation methodology, achieving competitive performance and high safety standards.
Lingxiao Du, Fanqing Meng, Zongkai Liu +4 authors
MM-PRM, a process reward model with step-level annotations, enhances logical reasoning in multimodal language models by using automated supervision and achieves improved performance on various benchmarks.
Siqiao Huang, Jialong Wu, Qixing Zhou +2 authors
Vid2World repurposes pre-trained video diffusion models into interactive world models via causalization and action guidance, enhancing action controllability and scalability in complex environments.
Zhengqing Yuan, Weixiang Sun, Yixin Liu +13 authors
EfficientLLM evaluates efficiency techniques for LLMs across architecture pretraining, fine-tuning, and inference, demonstrating task-dependent optima and cross-modal generalization.
Igor Udovichenko, Olivier Croissant, Anita Toleutaeva +2 authors
Proposed Itakura-Saito divergence-based loss function enhances numerical stability in risk-averse reinforcement learning using exponential utility functions.
Yuechen Zhang, Jinbo Xing, Bin Xia +6 authors
Jenga, a novel inference pipeline for video Diffusion Transformer models, combines dynamic attention carving and progressive resolution generation to significantly speed up video generation while maintaining high quality.
Mohammad Reza Taesiri, Brandon Collins, Logan Bolton +4 authors
Analysis of 83k image editing requests reveals that AI editors, including GPT-4o, struggle with low-creativity tasks and precise editing, while performing better on open-ended tasks, and human and VLM judges differ in their preferences for AI versus human edits.
Kaizhi Zheng, Ruijian Zhang, Jing Gu +2 authors
A training-free framework named 3DTown generates realistic 3D scenes from a single top-down image using region-based generation and spatial-aware 3D inpainting techniques.
Xueguang Ma, Qian Liu, Dongfu Jiang +3 authors
General-Reasoner enhances large language model reasoning across diverse domains by using a large-scale dataset and generative model-based answer verification, outperforming existing methods.
Zongkai Liu, Fanqing Meng, Lingxiao Du +4 authors
A novel reinforcement learning algorithm, CPGD, stabilizes policy learning in language models by constraining policy drift and clipping updates, improving performance and stability.
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