今日从 arXiv 订阅中筛选 10 篇论文。

Arxiv ID2609.31394幻觉翻译2609.31394

⚡ InternW0-Δ: A World Action Model Bridging Predictive Dynamics and Actions with 20K+ Hours of Open Data

世界模型×动作大系统：MoT 双专家 + 冻结 VLM 语义 + 训练期蒸馏，23,072h 全栈开源。

InternW0-Δ: A World Action Model Bridging Predictive Dynamics and Actions with 20K+ Hours of Open Data
Arxiv ID2609.30436幻觉翻译2609.30436

⚡ WALT: Learning World-Model-Aligned Latent Trajectories for Autonomous Driving

轨迹潜空间对齐冻结世界模型：NAVSIM 涨分的同时规划 FLOPs −30.5%。

WALT: Learning World-Model-Aligned Latent Trajectories for Autonomous Driving
Arxiv ID2609.30595幻觉翻译2609.30595

⚡ Action Forcing: Training World Models on Unsupervised Video by Recovering Underlying Egomotion Bases

无标注视频训可控世界模型：PCA egomotion 基当动作监督 + 四维免参考评测。

Action Forcing: Training World Models on Unsupervised Video by Recovering Underlying Egomotion Bases
Arxiv ID2609.31161幻觉翻译2609.31161

⚡ I Act Therefore I Am: When Is JEPA’s Action-Conditioning Enough to Learn Causal Mechanisms?

动作条件 JEPA 的可辨识性：何时能恢复因果状态（充分动作变异）。

I Act Therefore I Am: When Is JEPA's Action-Conditioning Enough to Learn Causal Mechanisms?
Arxiv ID2609.30946幻觉翻译2609.30946

⚡ OneWorld: Learning Consistent Physics Across Actions in World Models

跨动作物理一致性：共享潜机制 + 多干预评估协议。

OneWorld: Learning Consistent Physics Across Actions in World Models
Arxiv ID2609.30996幻觉翻译2609.30996

⚡ The Linear Representation Hypothesis for Vision-Language-Action Models

VLA 的线性表示假设：表示-策略统一的 signature 框架 + 线性操控。

The Linear Representation Hypothesis for Vision-Language-Action Models
Arxiv ID2609.31383幻觉翻译2609.31383

⚡ Guiding End-to-End Driving Models with Endpoint-Constrained Trajectory Optimization

驾驶闭环差距修复：端点保持、重塑中间 waypoint，即插即用。

Guiding End-to-End Driving Models with Endpoint-Constrained Trajectory Optimization
Arxiv ID2609.30465幻觉翻译2609.30465

⚡ RAZOR: Pruning Replaceable Experts in LLMs

MoE 专家剪枝的「功能可替代性」口径：8 组设置全胜 REAP。

RAZOR: Pruning Replaceable Experts in LLMs
Arxiv ID2609.30783幻觉翻译2609.30783

⚡ Skip the Talk, Re-Focus on Vision: Latent Reasoning for Reasoning Segmentation in Multimodal Large Language Models

推理分割的潜空间推理：潜 token 替代显式 CoT，token ↓16×、gIoU ↑。

Arxiv ID2609.30557幻觉翻译2609.30557

⚡ Auditing Latent-Space Monitors for Autonomous Driving

潜空间监控器审计：能预测≠有增量，输出-only 同样强。

自动生成于 2026-09-29 · 基于 arXiv Daily Digest