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
Yufei Xu, Fanxu Meng, Fan Jiang +11 authors
HISA improves sparse attention efficiency by replacing the traditional indexer with a hierarchical approach that reduces computational complexity from O(L²) to sub-quadratic scaling while maintaining selection fidelity.
Haonan Han, Jiancheng Huang, Xiaopeng Sun +7 authors
ViGoR benchmark addresses limitations in current AIGC evaluation by introducing a comprehensive framework for assessing visual generative reasoning across multiple modalities and cognitive dimensions.
Jinkun Hao, Mingda Jia, Ruiyan Wang +5 authors
We introduce EgoSim, a closed-loop egocentric world simulator that generates spatially consistent interaction videos and persistently updates the underlying 3D scene state for continuous simulation. Existing egocentric simulators either lack explicit 3D grounding, causing structural drift under viewpoint changes, or treat the scene as static, failing to update world states across multi-stage interactions. EgoSim addresses both limitations by modeling 3D scenes as updatable world states. We generate embodiment interactions via a Geometry-action-aware Observation Simulation model, with spatial consistency from an Interaction-aware State Updating module. To overcome the critical data bottleneck posed by the difficulty in acquiring densely aligned scene-interaction training pairs, we design a scalable pipeline that extracts static point clouds, camera trajectories, and embodiment actions from in-the-wild large-scale monocular egocentric videos. We further introduce EgoCap, a capture system that enables low-cost real-world data collection with uncalibrated smartphones. Extensive experiments demonstrate that EgoSim significantly outperforms existing methods in terms of visual quality, spatial consistency, and generalization to complex scenes and in-the-wild dexterous interactions, while supporting cross-embodiment transfer to robotic manipulation. Codes and datasets will be open soon. The project page is at egosimulator.github.io.
Kyudan Jung, Jihwan Kim, Soyoon Kim +3 authors
Full-duplex speech language models require high-quality multi-speaker conversational data, which is scarce, necessitating a robust open-source data processing pipeline to address challenges in natural dialogue dynamics and system accuracy.
Chongyang Zhao, Mingsong Li, Haodong Lu +1 authors
LLaVA-DyMoE addresses routing-drift-induced forgetting in multimodal continual instruction tuning by dynamically expanding mixture of experts with token-level assignment guidance and routing score regularizations.
Esakkivel Esakkiraja, Sai Rajeswar, Denis Akhiyarov +1 authors
Reasoning models appear to encode action choices before beginning textual deliberation, as evidenced by early decision detection and activation steering effects.
Aleksandar Cvejic, Rameen Abdal, Abdelrahman Eldesokey +2 authors
Researchers developed a novel framework using Near-identity distractors to improve identity-focused vision tasks by creating a dataset and evaluation protocol that better isolates identity from background context, leading to more reliable representations and metrics.
Samin Mahdizadeh Sani, Max Ku, Nima Jamali +23 authors
ImagenWorld introduces a comprehensive benchmark for image generation and editing across multiple tasks and domains, featuring human annotations and explainable evaluation to assess model performance and failure modes.
Shihua Zhang, Qiuhong Shen, Shizun Wang +2 authors
GeoSR enhances vision-language models' spatial reasoning by strategically utilizing geometry tokens through masking and guided fusion mechanisms.
Jiachun Jin, Zetong Zhou, Xiao Yang +4 authors
LatentUM is a unified model that represents all modalities in a shared semantic latent space, enabling efficient cross-modal reasoning and generation without pixel-space mediation.
Gleb Rodionov
Reasoning behaviors in large language models compress under varied contextual conditions, potentially impacting performance on complex tasks despite maintaining accuracy on simpler ones.
Zhangqi Jiang, Zheng Sun, Xianfang Zeng +7 authors
A new benchmark and evaluation model for image editing are introduced to better assess visual consistency and human alignment in complex editing tasks.
Jiaqi Liu, Zipeng Ling, Shi Qiu +9 authors
An autonomous research pipeline discovers Omni-SimpleMem, a unified multimodal memory framework that significantly improves lifelong AI agent performance through automated architectural modifications, bug fixes, and prompt engineering rather than traditional hyperparameter tuning.
Muyang He, Hanzhong Guo, Junxiong Lin +1 authors
Video generation models capable of simulating complex physical dynamics and long-horizon causality require efficient frameworks to become practical world simulators for interactive applications.
Weixiang Shen, Yanzhu Hu, Che Liu +8 authors
MEDOPENCLAW and MEDFLOWBENCH enable evaluation of vision-language models in medical imaging by allowing dynamic interaction with 3D medical volumes through standard viewers, revealing limitations in spatial reasoning when accessing professional tools.
Zhe Yang, Shulin Tian, Kairui Hu +9 authors
HippoCamp is a multimodal file management benchmark that evaluates agents' capabilities in user-centric environments, revealing significant performance gaps in long-horizon retrieval and cross-modal reasoning within dense personal file systems.
Jovana Kondic, Pengyuan Li, Dhiraj Joshi +24 authors
ChartNet is a large-scale multimodal dataset featuring 1.5 million chart samples with aligned visual, textual, and numerical components, designed to enhance chart interpretation and reasoning capabilities in multimodal models.
Jiajun Zhang, Yuying Li, Zhixun Li +13 authors
Vision-Language Models face significant challenges in generating complex multi-panel charts from real-world data, as demonstrated by a new large-scale benchmark that reveals performance gaps between proprietary and open-weight models.
Yongkang Li, Lijun Zhou, Sixu Yan +11 authors
UniDriveVLA is a unified vision-language-action model for autonomous driving that decouples spatial perception and semantic reasoning through a mixture-of-transformers architecture with expert coordination and progressive training.
Weixian Xu, Tiantian Mi, Yixiu Liu +6 authors
ASI-Evolve is an agentic framework that demonstrates AI-driven discovery across key AI development components, achieving superior performance in neural architecture design, data curation, and reinforcement learning algorithm design.
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