今日从 arXiv 订阅中筛选 10 篇论文。

Arxiv ID2610.00558幻觉翻译2610.00558

⚡ Redundancy Meets Synergy: Dependency-aware Expert Selection for MoE via Submodular Optimization

专家选择的正式化：从「单专家打分」到「组合结构」。

Redundancy Meets Synergy: Dependency-aware Expert Selection for MoE via Submodular Optimization
Arxiv ID2610.01698幻觉翻译2610.01698

⚡ ActiveWAM: Evidence-Aware Active Vision for World-Action Models

WAM 的「看哪里、留什么」：把相机升格为决策对象。

ActiveWAM: Evidence-Aware Active Vision for World-Action Models
Arxiv ID2610.00757幻觉翻译2610.00757

⚡ Video Evidence Indexing: Learning Where to Look from Video Previews for Token-Budgeted Long-Video Question Answering

token 预算下的「看哪几帧」：把证据选择当策略学。

Video Evidence Indexing: Learning Where to Look from Video Previews for Token-Budgeted Long-Video Question Answering
Arxiv ID2610.00367幻觉翻译2610.00367

⚡ MoRA: MoE Pruning via Router Bias Learning and Expert Approximation

MoE 剪枝：别只看单个专家，看它们的路由行为。

MoRA: MoE Pruning via Router Bias Learning and Expert Approximation
Arxiv ID2610.02120幻觉翻译2610.02120

⚡ SkeleWAM: Skeleton World-Action Modeling for Efficient Robotic Manipulation

WAM 的状态换成「稀疏 3D 骨架」。

SkeleWAM: Skeleton World-Action Modeling for Efficient Robotic Manipulation
Arxiv ID2610.00992幻觉翻译2610.00992

⚡ Closed-Loop Refinement and Execution for Learned Driving Planners

闭环里「轨迹好看」和「车能开」是两回事。

Closed-Loop Refinement and Execution for Learned Driving Planners
Arxiv ID2610.01762幻觉翻译2610.01762

⚡ OneStreamer: Unifying Perception, Memory, and Proactive Response in Streaming Video Interaction

流式视频：记录与响应，共享一个生成过程。

OneStreamer: Unifying Perception, Memory, and Proactive Response in Streaming Video Interaction
Arxiv ID2610.00327幻觉翻译2610.00327

⚡ Actions with Receipts: Jointly Binding Claims, Evidence, and Execution for Replayable Tool-Agent Auditing

审计的最后一公里：claim、证据、执行必须绑成一条链。

Actions with Receipts: Jointly Binding Claims, Evidence, and Execution for Replayable Tool-Agent Auditing
Arxiv ID2610.01083幻觉翻译2610.01083

⚡ WBAG: A Whole-Body and Attached-Geometry Safety Framework for Vision-Language-Action Manipulation

VLA 安全：不建全身体+抓持几何的检查都是漏的。

WBAG: A Whole-Body and Attached-Geometry Safety Framework for Vision-Language-Action Manipulation
Arxiv ID2610.00317幻觉翻译2610.00317

⚡ DriftOPD: Sequence-Level Reverse-KL Distillation for One-Step VLA Policies

序列级蒸馏：teacher 和 rollout 都可以不要。

DriftOPD: Sequence-Level Reverse-KL Distillation for One-Step VLA Policies

自动生成于 2026-10-03 · 基于 arXiv Daily Digest