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。
Chujie Zheng, Shixuan Liu, Mingze Li +9 authors
Group Sequence Policy Optimization (GSPO) is a reinforcement learning algorithm that improves training efficiency and performance of large language models by using sequence-level importance ratios and operations.
50 篇论文 · 按点赞排序
Xingxuan Li, Yao Xiao, Dianwen Ng +15 authors
The MiroMind-M1 series, built on Qwen-2.5, introduces fully open-source reasoning language models with state-of-the-art performance in mathematical reasoning through Context-Aware Multi-Stage Policy Optimization.
Fei Tang, Zhangxuan Gu, Zhengxi Lu +9 authors
A new reward framework, GUI-G$^2$, models GUI elements as continuous Gaussian distributions to improve autonomous interaction through dense gradient signals, outperforming existing methods in spatial reasoning tasks.
Hongyin Luo, Nathaniel Morgan, Tina Li +7 authors
The Thread Inference Model (TIM) and TIMRUN enable long-horizon reasoning in large language models by using reasoning trees and a rule-based subtask-pruning mechanism to manage working memory and GPU resources efficiently.
Dmitrii Mikhailov, Aleksey Letunovskiy, Maria Kovaleva +6 authors
NABLA, a Neighborhood Adaptive Block-Level Attention mechanism, enhances video diffusion transformers by reducing computational overhead without significantly impacting generative quality or visual fidelity.
Xiaofeng Mao, Shaoheng Lin, Zhen Li +7 authors
A framework for generating and exploring interactive, high-fidelity video worlds from images using a Masked Video Diffusion Transformer, advanced sampling techniques, and model acceleration.
Fang Wu, Weihao Xuan, Ximing Lu +2 authors
Reinforcement Learning with Verifiable Rewards (RLVR) enhances precision but may limit exploration and discovery of new solutions, suggesting potential limits to its effectiveness in expanding reasoning capabilities.
Boyong Wu, Chao Yan, Chen Hu +106 authors
Step-Audio~2, an end-to-end multi-modal large language model, integrates latent audio encoding and reinforcement learning to achieve state-of-the-art performance in ASR, audio understanding, and speech conversation, incorporating discrete audio token generation and retrieval-augmented generation.
Hongcheng Gao, Zihao Huang, Lin Xu +11 authors
The Turing Eye Test evaluates MLLMs' perceptual capabilities on synthetic images, revealing that vision tower generalization is a key gap compared to human perception.
Run-Ze Fan, Zengzhi Wang, Pengfei Liu
TextbookReasoning and MegaScience datasets, along with a comprehensive evaluation system, enhance scientific reasoning in AI by providing high-quality, verifiable data and outperforming existing models.
Zichen Wen, Jiashu Qu, Dongrui Liu +11 authors
DIJA is a framework that exploits safety weaknesses in diffusion-based large language models by constructing adversarial prompts, demonstrating significant vulnerabilities in their alignment mechanisms.
Maksim Kuprashevich, Grigorii Alekseenko, Irina Tolstykh +4 authors
An automated pipeline mines high-fidelity image editing triplets using generative models and a task-tuned validator, enabling large-scale training without human labeling.
Zhengwei Tao, Jialong Wu, Wenbiao Yin +10 authors
WebShaper, a formalization-driven framework, synthesizes information-seeking datasets using set theory and Knowledge Projections to enhance reasoning structure and achieve top performance in open-sourced benchmarks.
Kirill Borodin, Nikita Vasiliev, Vasiliy Kudryavtsev +4 authors
Balalaika, a new Russian speech dataset, improves speech synthesis and enhancement through comprehensive annotations and large-scale data.
Jooyeol Yun, Heng Wang, Yotaro Shimose +2 authors
DesignLab, an iterative design tool using fine-tuned language models, outperforms existing methods by refining design drafts through continuous error detection and correction.
Chilam Cheang, Sijin Chen, Zhongren Cui +18 authors
GR-3, a large-scale vision-language-action model, demonstrates exceptional generalization and fine-tuning capabilities, excelling in complex tasks and outperforming state-of-the-art methods.
Hang Yan, Fangzhi Xu, Rongman Xu +8 authors
Momentum Uncertainty-guided Reasoning (MUR) dynamically allocates computational resources to improve reasoning efficiency and accuracy in Large Language Models without additional training.
Chi-Pin Huang, Yueh-Hua Wu, Min-Hung Chen +2 authors
ThinkAct, a dual-system framework, uses reinforced visual latent planning to enable few-shot adaptation, long-horizon planning, and self-correction in embodied AI tasks by bridging high-level reasoning with low-level action execution.
Junfei Xiao, Ceyuan Yang, Lvmin Zhang +7 authors
Captain Cinema generates coherent short movies from textual descriptions using top-down keyframe planning and bottom-up video synthesis with Multimodal Diffusion Transformers.
Urchade Zaratiana, Gil Pasternak, Oliver Boyd +2 authors
GLiNER2 is a unified transformer-based framework that supports multiple NLP tasks with improved efficiency and accessibility compared to large language models.
Zhixiong Zhang, Shuangrui Ding, Xiaoyi Dong +7 authors
A concept-driven segmentation framework using Large Vision-Language Models improves video object segmentation by integrating high-level semantic reasoning and adapting to scene complexity.
Yu Li, Zhuoshi Pan, Honglin Lin +3 authors
A study investigates multi-domain reasoning in Reinforcement Learning with Verifiable Rewards (RLVR) using the GRPO algorithm and Qwen-2.5-7B model family, exploring in-domain improvements, cross-domain generalization, and the impact of curriculum learning and reward design.
Hayeon Kim, Ji Ha Jang, Se Young Chun
RoMaP, a novel framework, enhances precise local 3D Gaussian editing through robust 3D mask generation and regularized Score Distillation Sampling loss, achieving state-of-the-art results.
Yiwen Chen, Zhihao Li, Yikai Wang +4 authors
Ultra3D is an efficient 3D generation framework that uses VecSet for coarse layout and Part Attention for refined voxel features, achieving high-resolution generation with improved speed and quality.
Wongi Jeong, Kyungryeol Lee, Hoigi Seo +1 authors
Region-Adaptive Latent Upsampling (RALU) accelerates diffusion transformer inference by performing mixed-resolution sampling, reducing computation while maintaining image quality.
Ang Li, Charles Wang, Kaiyu Yue +11 authors
Zebra-CoT, a large-scale dataset with diverse visual and text reasoning tasks, improves multimodal model performance through fine-tuning, enhancing accuracy and visual chain of thought generation.
Shashanka Venkataramanan, Valentinos Pariza, Mohammadreza Salehi +5 authors
Franca, an open-source vision foundation model, achieves high performance using a transparent training pipeline and novel clustering and disentanglement techniques.
Yilong Xu, Xiang Long, Zhi Zheng +1 authors
RAVine is a new evaluation framework for agentic LLMs with search, focusing on realistic queries, accurate ground truth, and iterative process efficiency.
Cheng-Han Chiang, Xiaofei Wang, Linjie Li +7 authors
Stitch is a novel method that enables simultaneous thinking and talking in spoken language models by alternating between generating unspoken reasoning and spoken response chunks, achieving low latency and improved performance on math reasoning tasks.
Hao Luo, Yicheng Feng, Wanpeng Zhang +7 authors
Being-H0, a Vision-Language-Action model, uses human videos and physical instruction tuning to achieve high dexterity in manipulation tasks and scales well with data and model size.
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