今日从 arXiv 订阅中筛选 10 篇论文。

Arxiv ID2609.25809幻觉翻译2609.25809

⚡ You Only Need 2/3 of the Chosen Experts: An Empirical Study of Dynamic Expert Pruning in Fine-Grained MoE LLMs

细粒度 MoE 动态专家剪枝的系统实证：保留 2/3 选中专家 = 98.8% 性能 + 1.2-1.7× 加速。

Arxiv ID2609.25820幻觉翻译2609.25820

⚡ Beyond Reconstruction Error: Analytical and Data-Driven Action Tokenization for Autoregressive Vision-Language-Action Models

动作 tokenization 的重建误差 ≠ 闭环控制效用。

Arxiv ID2609.26299幻觉翻译2609.26299

⚡ ForeDrive: Foresight-Guided End-to-End Autonomous Driving with a Planning-Relevant Latent World Model

规划相关的 latent world model：把预测表示直接做成规划条件信号。

ForeDrive: Foresight-Guided End-to-End Autonomous Driving with a Planning-Relevant Latent World Model
Arxiv ID2609.26292幻觉翻译2609.26292

⚡ RoboTwin-Phys: Do WAMs and VLAs Understand the Physical World?

把物理条件多样性做成显式评测维度：13 个物理属性连续变化。

RoboTwin-Phys: Do WAMs and VLAs Understand the Physical World?
Arxiv ID2609.25366幻觉翻译2609.25366

⚡ From Decorative to Load-Bearing: Task Difficulty Shapes the Causal Role of Chain-of-Thought

CoT 的承重因果测试：简单题上模型绕过自己的推理，难题上才跟着污染步骤走。

From Decorative to Load-Bearing: Task Difficulty Shapes the Causal Role of Chain-of-Thought
Arxiv ID2609.26314幻觉翻译2609.26314

⚡ TriWorldBench: A Tri-View Consistency Perspective on Embodied World Models

具身世界模型三视角一致性基准：单看合理≠同一动作的一致预测。

TriWorldBench: A Tri-View Consistency Perspective on Embodied World Models
Arxiv ID2609.25969幻觉翻译2609.25969

⚡ Predict Before You Step: Auditable Occupancy Forecasting for Dynamic Obstacle Avoidance under Sparse Guidance

可审计占用预测做动态避障：预测接口显式可检查、可替换。

Predict Before You Step: Auditable Occupancy Forecasting for Dynamic Obstacle Avoidance under Sparse Guidance
Arxiv ID2609.25827幻觉翻译2609.25827

⚡ Protocol before progress: leakage-aware evaluation of AIS trajectory prediction

AIS 轨迹预测：进展该归给协议还是架构？先测协议。

Protocol before progress: leakage-aware evaluation of AIS trajectory prediction
Arxiv ID2609.26007幻觉翻译2609.26007

⚡ Skytopia: Monocular Drone Navigation with Action-Conditioned Latent World Models

单目无人机导航："要的不是预测，而是产生预测的表示"。

Skytopia: Monocular Drone Navigation with Action-Conditioned Latent World Models
Arxiv ID2609.25773幻觉翻译2609.25773

⚡ Video-HopChain: Multi-Hop Questions and Confidence-Gated Exploration for Video Reasoning Models

多跳视频问答：把视觉证据链做成可验证 RLVR 信号。

Video-HopChain: Multi-Hop Questions and Confidence-Gated Exploration for Video Reasoning Models

自动生成于 2026-09-24 · 基于 arXiv Daily Digest