今日从 arXiv 订阅中筛选 10 篇论文。

Arxiv ID2609.39245幻觉翻译2609.39245

⚡ ReWAM: Reciprocal World Action Models for Interactive Autonomous Driving

WAM 的交互问：把「他车」从布景升格为决策者。

ReWAM: Reciprocal World Action Models for Interactive Autonomous Driving
Arxiv ID2609.39973幻觉翻译2609.39973

⚡ EWAM: Emergent Depth-Wise Specialization in a Unified Embodied Model

没人监督分工，分工自己出现：动作 token 的深度角色分化。

EWAM: Emergent Depth-Wise Specialization in a Unified Embodied Model
Arxiv ID2609.39137幻觉翻译2609.39137

⚡ ID Balancing: Stable Training of Extremely Sparse MoE via PID-Based Load Control

负载均衡不是玄学，是控制问题。

Arxiv ID2609.39235幻觉翻译2609.39235

⚡ The Planning Limits of Latent World Models

「预测式潜空间够不够」：今天给出边界本身。

The Planning Limits of Latent World Models
Arxiv ID2609.38463幻觉翻译2609.38463

⚡ TrafficSignBench: Rule-Centric Closed-Loop Evaluation of Traffic-Sign Compliance in Autonomous Driving

高分≠合法：以交通标志为中心的合规性闭环评估。

TrafficSignBench: Rule-Centric Closed-Loop Evaluation of Traffic-Sign Compliance in Autonomous Driving
Arxiv ID2609.38984幻觉翻译2609.38984

⚡ Sparse-WAM: Accelerating World Action Models via Action-Guided Sparse Imagination

WAM 加速：剪「未来帧 token」，且按动作相关性剪。

Sparse-WAM: Accelerating World Action Models via Action-Guided Sparse Imagination
Arxiv ID2609.39026幻觉翻译2609.39026

⚡ Search Shapes Conclusions: Auditing Evidence Selection Bias in Deep Research Agents

引用都对，结论仍可能歪：审计「读到的东西有没有代表性」。

Search Shapes Conclusions: Auditing Evidence Selection Bias in Deep Research Agents
Arxiv ID2609.38900幻觉翻译2609.38900

⚡ MEMO: Multi-Level Entity-Aware Memory for Streaming Video Understanding

流式视频记忆：不是更大，是更有结构。

MEMO: Multi-Level Entity-Aware Memory for Streaming Video Understanding
Arxiv ID2609.39665幻觉翻译2609.39665

⚡ ChronoGraph: Functional 4D Scene Graphs with Vision-Language Models for Interaction Understanding and Grounded Planning

理解与规划共用一张图：功能性 4D 场景图。

ChronoGraph: Functional 4D Scene Graphs with Vision-Language Models for Interaction Understanding and Grounded Planning
Arxiv ID2609.38823幻觉翻译2609.38823

⚡ DecoMoE: Decoupling Visual Propagation and Expert Computation for Efficient Multimodal MoE Inference

多模态 MoE 推理：视觉传播与专家计算各自压缩。

DecoMoE: Decoupling Visual Propagation and Expert Computation for Efficient Multimodal MoE Inference

自动生成于 2026-10-02 · 基于 arXiv Daily Digest