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。
Ruhan Wang, Yucheng Shi, Zongxia Li +7 authors
The Harness Handbook and Behavior-Guided Progressive Disclosure automatically map agent behaviors to source code, improving localization and edit planning for evolving AI harnesses.
50 篇论文 · 按点赞排序
Changhai Zhou, Kieran Liu, Yuhua Zhou +17 authors
LongStraw is an execution framework that enables million-token reinforcement learning post-training with fixed GPU resources by reducing live training graphs through replay and selective state retention.
Xinhao Li, Yuhan Zhu, Xiangyu Zeng +24 authors
VideoChat3 is an open, efficient generalist video-language model that uses an inflated 3D vision transformer and adaptive frame resolution for efficient spatiotemporal modeling, alongside curated synthetic datasets, to achieve broad video understanding with only 4B parameters.
Shiyuan Feng, Huan-ang Gao, Haohan Chi +7 authors
Direct On-Policy Distillation transfers reinforcement learning improvements from smaller to larger models by using the policy shift induced by RL as an implicit reward signal, enabling efficient scaling of training without re-running expensive RL on the target model.
Guoxuan Chen, Chufeng Xiao, Haoran Yang +30 authors
Boogu-Image-0.1 is an open-source unified multimodal model family that achieves strong text-to-image generation, editing, and bilingual rendering with low training cost by improving multimodal understanding, data quality, and agentic inference scaling.
Yubo Wang, Jiarong Liang, Yuxuan Zhang +5 authors
Function-aware fill-in-the-middle mid-training on code improves coding-agent tool integration and mitigates capability erosion from agentic post-training.
Jinyang Wu, Shuo Yang, Zhengxi Lu +8 authors
SEED improves long-horizon agent training by extracting hindsight skills from completed trajectories and distilling them into dense token-level supervision alongside outcome-based reinforcement learning.
Ruiyan Gong, Yingnan Guo, Junjun Hu +43 authors
ABot-N1 improves visual language navigation by separating reasoning from control through a slow-fast architecture that uses explicit chain-of-thought reasoning and pixel goals to guide continuous waypoint generation, achieving strong results across diverse embodied tasks.
Xinyu Tang, Gangqiang Cao, Yurou Liu +13 authors
Scaling zero-reinforcement-learning to 1 trillion parameters improves reasoning quality and reveals emergent cognitive behaviors, supported by optimized training and structured evaluation.
Letian Wang, Chuhan Zhang, Rishabh Kabra +9 authors
GenCeption uses a video generative diffusion backbone to build a generalist vision model that achieves strong performance across diverse perception tasks with high data efficiency and emergent generalization.
Haozhe Wang, Weijia Feng, Jinpeng Yu +8 authors
Researchers introduce benchmarks and a co-training framework to enable search-augmented visual generation that addresses open-ended world-knowledge gaps.
Runhui Huang, Qihui Zhang, Zhe Liu +3 authors
SpectraReward leverages pretrained multimodal language models as training-free reward models for image-generation reinforcement learning by measuring prompt recovery likelihood, with a self-improving variant that uses the policy's own understanding branch to close the loop.
Jiayi Tian, Shiao Liu, Yuting Xu +30 authors
ABot-AgentOS is a general robotic agent operating system that provides deliberative planning, isolated skill execution, verification, and persistent multi-modal memory for long-horizon embodied tasks, evaluated on the EmbodiedWorldBench benchmark.
Yunxin Li, Jinchao Li, Shibo Su +7 authors
KnowAct-GUIClaw introduces a Know-Route-Act-Reflect framework that enhances cross-platform GUI automation and recursive self-improvement through experience-driven memory and skill libraries.
Yuliang Liu, Zhang Li, Ziyang Zhang +11 authors
MonkeyOCRv2 is a document-oriented visual-text pretrained encoder that improves performance across document analysis and multimodal document understanding tasks while using a much smaller vision backbone.
Zongxia Li, Zhongzhi Li, Yucheng Shi +10 authors
A new terminal benchmark evaluates long-horizon AI agents through graded subtasks that provide dense rewards and reveal substantial performance gaps.
Abhigya Verma, Khyati Mahajan, Amit Kumar Saha +4 authors
SynthDocBench is a synthetic long-context visual document benchmark that isolates failure modes in vision-language models, revealing sharp length degradation, positional sensitivity, and chart comprehension breakdowns.
Yuyao Zhang, Junjie Gao, Zhengxian Wu +11 authors
SearchOS is a multi-agent framework that externalizes search progress into shared structured states to prevent repetitive loops and improve information-seeking efficiency.
Yiming Zhang, Zhonghan Zhao, Wenwei Zhang +13 authors
Visual pretraining on document images outperforms text-only pretraining for scalable language intelligence without extracting text.
Shiyin Lu, Yinglun Li, Yu Xia +10 authors
OvisOCR2 is a compact end-to-end document parsing model that uses supervised fine-tuning, reinforcement learning, and distillation to generate structured Markdown from document images with state-of-the-art accuracy.
Minhyuk Hwang, Sangmin Kim, Seunguk Do +2 authors
StudioRecon reconstructs dynamic human scenes from sparse cameras by separating background and human reconstruction, using synthesized views and deformable Gaussian initialization with consistency refinement.
Qi Li, Xingyi Yang, Xinchao Wang
BadWAM reveals that world-action models are vulnerable to adversarial perturbations that desynchronize imagined futures from executed actions, enabling both overt and stealthy control failures.
Yijun Chen, Boyi Xiao, Yixian Zhao +10 authors
LightMem-Ego is a lightweight streaming multimodal memory system that organizes continuous egocentric visual and audio data into hierarchical memory levels to support real-time retrieval and personalized assistance on mobile and wearable devices.
Xinghang Li, Jun Guo, Qiwei Li +21 authors
Xiaomi-Robotics-U0 is a large multimodal autoregressive model that unifies embodied image and video generation tasks while preserving pre-trained visual knowledge and enabling multi-view consistency across robot embodiments.
Zichen Ding, Jiaye Ge, Shufan Jiang +20 authors
AgentCompass is an open-source framework that unifies evaluation of LLM-based agents through modular benchmarks, harnesses, and environments with fault-tolerant execution and trajectory analysis.
Yuqi Tang, Tengfei Liu, Yizheng Lai +18 authors
KeyFrame-Compass benchmarks keyframe-conditioned video generation, revealing trade-offs between faithful keyframe reproduction and video quality across diverse settings.
Yufei Cai, Xuesong Niu, Hao Lu +3 authors
MetaView is a diffusion-based framework for monocular novel view synthesis that integrates implicit geometry modeling with minimal explicit 3D cues to achieve consistent, controllable rendering under large viewpoint changes.
Matteo Santelmo, Xiuying Wei, Israa Fakih +5 authors
A new benchmark of simple human tasks reveals persistent weaknesses in modern AI systems and highlights performance gaps between open and closed models.
Zhen Li, Zian Meng, Shuwei Shi +4 authors
This survey examines interactive game world modeling through action-state-observation loops, analyzing control, dynamics, persistence, and real-time generation while introducing a large-scale gameplay dataset.
Zhengpeng Xie, Li Lyna Zhang, Zeke Xie +1 authors
Trust Region Policy Distillation stabilizes on-policy distillation via a proximal teacher, reducing gradient variance and improving convergence for mathematical reasoning tasks without extra computation.
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