今日从 arXiv 订阅中筛选 10 篇论文。

Arxiv ID2605.15725幻觉翻译2605.15725

⚡ DiLA: Disentangled Latent Action World Models

内容-结构解耦潜在动作世界模型，解决动作抽象与生成保真度矛盾。无标签自监督，支持跨体现动作迁移。

DiLA: Disentangled Latent Action World Models
Arxiv ID2605.15618幻觉翻译2605.15618

⚡ Latent Video Prediction Learns Better World Models

首次系统评估四种视频基础模型五维度鲁棒性。潜在预测模型V-JEPA在所有维度一致优于像素重建/对比方法。

Latent Video Prediction Learns Better World Models
Arxiv ID2605.15477幻觉翻译2605.15477

⚡ EgoExo-WM: Unlocking Exo Video for Ego World Models

利用外部视角视频增强自我中心世界模型，解决ego数据稀缺与域间差距。

EgoExo-WM: Unlocking Exo Video for Ego World Models
Arxiv ID2605.15466幻觉翻译2605.15466

⚡ Entity-Centric World Models: Interaction-Aware Masking for Causal Video Prediction

交互感知掩码因果视频预测世界模型，以实体为中心建模交互关系。

Arxiv ID2605.16079幻觉翻译2605.16079

⚡ VideoSeeker: Incentivizing Instance-level Video Understanding via Native Agentic Tool Invocation

原生agentic工具调用做实例级视频理解，RL激励工具使用提升细粒度感知。

VideoSeeker: Incentivizing Instance-level Video Understanding via Native Agentic Tool Invocation
Arxiv ID2605.15342幻觉翻译2605.15342

⚡ Minerva-Ego: Spatiotemporal Hints for Egocentric Video Understanding

自我中心视频推理基准+时空推理轨迹标注。定位提示显著提升模型表现。

Minerva-Ego: Spatiotemporal Hints for Egocentric Video Understanding
Arxiv ID2605.15735幻觉翻译2605.15735

⚡ UAM: A Dual-Stream Perspective on Forgetting in VLA Training

双流视角分析VLA训练灾难性遗忘，揭示视觉-语言-动作微调中遗忘模式。

Arxiv ID2605.15458幻觉翻译2605.15458

⚡ Video Models Can Reason with Verifiable Rewards

可验证奖励训练视频扩散模型推理，将推理信号注入视频生成过程。

Video Models Can Reason with Verifiable Rewards
Arxiv ID2605.15391幻觉翻译2605.15391

⚡ PanoWorld: Geometry-Consistent Panoramic Video World Modeling

从单图+描述生成几何一致的全景360°视频世界模型。

Arxiv ID2605.15424幻觉翻译2605.15424

⚡ Social-Mamba: Socially-Aware Trajectory Forecasting with State-Space Models

基于SSM的社交感知轨迹预测，平衡精度与效率。

Social-Mamba: Socially-Aware Trajectory Forecasting with State-Space Models

自动生成于 2026-05-18 · 基于 arXiv Daily Digest