今日从 arXiv 订阅中筛选 10 篇论文。

Arxiv ID2609.24626幻觉翻译2609.24626

⚡ Relationally Grounded Latent World Models for Autonomous Driving

给潜世界模型喂"关系结构"特权监督：只训练时用场景图，推理时全删。

Relationally Grounded Latent World Models for Autonomous Driving
Arxiv ID2609.23478幻觉翻译2609.23478

⚡ Algebraic Consistency Alone Does Not Certify Temporal Structure in Latent Action Models

拆掉潜动作模型的"免标签证书"：代数恒等式对任意配对都成立，证明不了时序结构。

Algebraic Consistency Alone Does Not Certify Temporal Structure in Latent Action Models
Arxiv ID2609.24749幻觉翻译2609.24749

⚡ D-JEPA: A Decision-Aligned Latent World Model

预测更准不等于选得对：把"决策相关序关系"从执行结果里学出来。

D-JEPA: A Decision-Aligned Latent World Model
Arxiv ID2609.22317幻觉翻译2609.22317

⚡ EditWM: Event-Decomposed World Modeling with Incremental Correction for End-to-End Autonomous Driving

把未来拆成"常规演化＋事件驱动增量修正"，只在对的地方花算力。

EditWM: Event-Decomposed World Modeling with Incremental Correction for End-to-End Autonomous Driving
Arxiv ID2609.24745幻觉翻译2609.24745

⚡ Beyond Visual Quality: A Study of Test-Time Planning with World Action Models

WAM 的想象未来能用于规划吗？oracle 上界给了，选择器回收不了。

Beyond Visual Quality: A Study of Test-Time Planning with World Action Models
Arxiv ID2609.23252幻觉翻译2609.23252

⚡ Robot World Models Are Not Invariant to How the Actions Are Written

动作写法（绝对关节 vs 增量）是静默继承的灾难：世界模型换编码就答错问题。

Robot World Models Are Not Invariant to How the Actions Are Written
Arxiv ID2609.22582幻觉翻译2609.22582

⚡ Beyond the Leaderboard: Counterfactual Diagnosis of End-to-End and VLA Driving Policies Under Domain Shift

排行榜名次预测不了新地点行为：用反事实编辑把驾驶策略读成"剖面"。

Beyond the Leaderboard: Counterfactual Diagnosis of End-to-End and VLA Driving Policies Under Domain Shift
Arxiv ID2609.23601幻觉翻译2609.23601

⚡ PREM: Prefix-Steered Recurrent Memory for Long-Video Understanding

长视频记忆从 token 挪进 KV 调制：冻结 VLM，write-once、query-many。

PREM: Prefix-Steered Recurrent Memory for Long-Video Understanding
Arxiv ID2609.23048幻觉翻译2609.23048

⚡ Anatomy of a Closed-Loop Collapse: A Causal Case Study of a Compressed VLA Policy

离线 0.996 的蒸馏策略闭环 0/72：一例崩溃的因果解剖。

Arxiv ID2609.22199幻觉翻译2609.22199

⚡ Improving Parameter Utilization by Sharing Neural Experts Across Layers in Transformers

层间参数冗余的另一种吃法：全局共享专家池。

Improving Parameter Utilization by Sharing Neural Experts Across Layers in Transformers

自动生成于 2026-09-23 · 基于 arXiv Daily Digest