今日从 arXiv 订阅中筛选 10 篇论文。

Arxiv ID2609.38140幻觉翻译2609.38140

⚡ Breaking the Uniformity Trap: Scaling Video Diffusion Model via SplitMoE

视频扩散 MoE 的 uniformity trap：把专家池拆成两种角色。

Breaking the Uniformity Trap: Scaling Video Diffusion Model via SplitMoE
Arxiv ID2609.37250幻觉翻译2609.37250

⚡ V-JEPA Policy: Building Effective World-Action Models on Predictive Visual Latents

WAM 底座之战：不继承生成器，预测式潜空间够不够？

V-JEPA Policy: Building Effective World-Action Models on Predictive Visual Latents
Arxiv ID2609.36851幻觉翻译2609.36851

⚡ RoXDrive: Closed-Loop Reinforcement Learning for End-to-End Autonomous Driving via Action-Faithful Rollouts

世界模型当 RL 环境之前，先过「动作忠实性」滤网。

RoXDrive: Closed-Loop Reinforcement Learning for End-to-End Autonomous Driving via Action-Faithful Rollouts
Arxiv ID2609.37970幻觉翻译2609.37970

⚡ PhysWAM: Physically Consistent World Action Model for Autonomous Driving

用「深度×运动」的几何关系把世界与动作焊在一起。

PhysWAM: Physically Consistent World Action Model for Autonomous Driving
Arxiv ID2609.37871幻觉翻译2609.37871

⚡ ExceptionDrive: A Planning-Oriented Counterfactual Corner-Case Benchmark for Autonomous Driving

平均分不测危险：把 hazard 注入真实场景再考规划器。

ExceptionDrive: A Planning-Oriented Counterfactual Corner-Case Benchmark for Autonomous Driving
Arxiv ID2609.37378幻觉翻译2609.37378

⚡ Do-JEPA: From Masking to Intervention in Latent World Models

干预物理世界，而不是模型能看见什么。

Do-JEPA: From Masking to Intervention in Latent World Models
Arxiv ID2609.38111幻觉翻译2609.38111

⚡ From Routing Signals to Selective Review: Visual regrounding in MoE VLMs

路由概率不只是负载表：它是随身的感知证据。

From Routing Signals to Selective Review: Visual regrounding in MoE VLMs
Arxiv ID2609.37046幻觉翻译2609.37046

⚡ Speed in the Blind Spot: An Interpretability Analysis of Dynamic Perception in VLMs for Autonomous Driving

动态感知审计：能读出来 ≠ 说得出来 ≠ 时间接地。

Speed in the Blind Spot: An Interpretability Analysis of Dynamic Perception in VLMs for Autonomous Driving
Arxiv ID2609.37349幻觉翻译2609.37349

⚡ TAEC: Trajectory-Aware Evidence Coordination for Multi-Step Visual RAG

检索到证据 ≠ 用得上：多步视觉 RAG 的证据生命周期管理。

TAEC: Trajectory-Aware Evidence Coordination for Multi-Step Visual RAG
Arxiv ID2609.37035幻觉翻译2609.37035

⚡ Watch-Think-Interact: Bootstrapping Long-Horizon Multi-Turn Streaming Video Reasoning with Reinforcement Learning

流式视频多轮推理：看、想、按需回找。

Watch-Think-Interact: Bootstrapping Long-Horizon Multi-Turn Streaming Video Reasoning with Reinforcement Learning

自动生成于 2026-10-01 · 基于 arXiv Daily Digest