今日从 arXiv 订阅中筛选 10 篇论文。

Arxiv ID2609.20709幻觉翻译2609.20709

⚡ MoWAM: Explicit Future Motion Prediction for Efficient World Action Models

WAM 的未来表示：不用生成视频，直接预测结构化机器人运动。

MoWAM: Explicit Future Motion Prediction for Efficient World Action Models
Arxiv ID2609.20377幻觉翻译2609.20377

⚡ MM-Future: Multi-Mode Joint World-Action Modeling for Autonomous Driving

自驾世界-动作模型：多条"场景-动作"假设成对演化，再按配对未来排序。

MM-Future: Multi-Mode Joint World-Action Modeling for Autonomous Driving
Arxiv ID2609.19934幻觉翻译2609.19934

⚡ Beyond Depth Truncation: Controlled Evaluation of Depth Utilization in Recursive Language Models

常用指标"截断深度看质量坡率"其实混了三个因素。

Arxiv ID2609.18176幻觉翻译2609.18176

⚡ MoRE: Mixture of Reused Experts

MoE 参数效率：相邻层组共享专家池，路由各管各的。

MoRE: Mixture of Reused Experts
Arxiv ID2609.20034幻觉翻译2609.20034

⚡ Astronex-World 1.0: Real-Time Interactive World Model Foundation

实时可交互视频世界模型基础：5 阶段全部在 2×L20 上完成。

Astronex-World 1.0: Real-Time Interactive World Model Foundation
Arxiv ID2609.18462幻觉翻译2609.18462

⚡ CSWAM: Better Causal Semantic Representations for Out-of-Distribution Generalization in World Action Models

WAM 的表示修正：重建导向表示换成因果语义表示，OOD 泛化翻数倍。

CSWAM: Better Causal Semantic Representations for Out-of-Distribution Generalization in World Action Models
Arxiv ID2609.19729幻觉翻译2609.19729

⚡ Recency Forcing: Bridging the Long-Horizon Gap in Autoregressive Video Generation

自回归视频生成的 train-inference 差异：KV 淘汰不匹配。

Arxiv ID2609.18259幻觉翻译2609.18259

⚡ M2Tok: Multi-head Multi-codebook Discrete Action Tokenization for Vision-Language-Action Models

离散动作 tokenizer：瓶颈在重建损失，修复在多头+多码本。

M2Tok: Multi-head Multi-codebook Discrete Action Tokenization for Vision-Language-Action Models
Arxiv ID2609.20776幻觉翻译2609.20776

⚡ GeoAAC: Geometry-Based Adaptive Action Chunking from Denoising Trajectories in VLA Policies

动作块开多长？用去噪轨迹的几何自己说。

GeoAAC: Geometry-Based Adaptive Action Chunking from Denoising Trajectories in VLA Policies
Arxiv ID2609.20480幻觉翻译2609.20480

⚡ Worst-Case Hidden-Vehicle Trajectory Search in Spatiotemporal Occlusion Regions

遮挡里藏着什么车？把"最坏合法轨迹"搜出来。

Worst-Case Hidden-Vehicle Trajectory Search in Spatiotemporal Occlusion Regions

自动生成于 2026-09-19 · 基于 arXiv Daily Digest