今日从 arXiv 订阅中筛选 10 篇论文。

Arxiv ID2609.18916幻觉翻译2609.18916

⚡ Higher-order pruning of experts in mixture-of-experts language models

专家剪枝要二阶：REAP 忽略了专家之间的交互项。

Higher-order pruning of experts in mixture-of-experts language models
Arxiv ID2609.18016幻觉翻译2609.18016

⚡ Causal-History Test-Time Scaling for Failure Recovery in Autoregressive World-Action Models

自回归 WAM 出错后：把"历史改哪里"也当作搜索问题。

Causal-History Test-Time Scaling for Failure Recovery in Autoregressive World-Action Models
Arxiv ID2609.18487幻觉翻译2609.18487

⚡ ActionPiece: Rethinking Action Tokenization for Autoregressive Vision-Language-Action Models

动作 tokenizer：重建 MSE 会骗人，物理关系会。

ActionPiece: Rethinking Action Tokenization for Autoregressive Vision-Language-Action Models
Arxiv ID2609.17909幻觉翻译2609.17909

⚡ Zing-0.5: Toward Playable Worlds with Real-Time Joint Action and Text Control

可玩世界模型：键盘+文本联合实时控制，成本按分钟计价。

Zing-0.5: Toward Playable Worlds with Real-Time Joint Action and Text Control
Arxiv ID2609.18430幻觉翻译2609.18430

⚡ StrucPhysVideo: Learning Physical Dynamics from Structured Captions and Robot Actions

视频世界模型的物理课：结构化标注 + 稀疏 MoE 课程。

StrucPhysVideo: Learning Physical Dynamics from Structured Captions and Robot Actions
Arxiv ID2609.18442幻觉翻译2609.18442

⚡ Risk-Aware World Modeling with Flow-Guided Occupancy Evolution for Selective Trajectory Planning in Automated Driving

什么时候该改道：风险增量触发，而不是每步重规划。

Risk-Aware World Modeling with Flow-Guided Occupancy Evolution for Selective Trajectory Planning in Automated Driving
Arxiv ID2609.17728幻觉翻译2609.17728

⚡ RAF-VLA: Representation Alignment with the Future for End-to-End Autonomous Driving

未来不必生成：对齐未来帧的表示就够了。

RAF-VLA: Representation Alignment with the Future for End-to-End Autonomous Driving
Arxiv ID2609.18197幻觉翻译2609.18197

⚡ WholeBodyWAM: Learning Whole-Body World Action Models with Scalable Motion Priors

人形全身 WAM：先把人类/异构机体的运动学成先验。

WholeBodyWAM: Learning Whole-Body World Action Models with Scalable Motion Priors
Arxiv ID2609.18080幻觉翻译2609.18080

⚡ Decodability is Not Causality: Dissociating Probe Readouts from Behavioral Drivers via SAE Decomposition

探针能读到，不等于模型在用它。

Decodability is Not Causality: Dissociating Probe Readouts from Behavioral Drivers via SAE Decomposition
Arxiv ID2609.19142幻觉翻译2609.19142

⚡ PointZero: 3D Point Track Completion for Learning Transferable 3D Dynamics

点轨迹补全：不用机器人数据学会 3D 动力学。

PointZero: 3D Point Track Completion for Learning Transferable 3D Dynamics

自动生成于 2026-09-18 · 基于 arXiv Daily Digest