今日从 arXiv 订阅中筛选 10 篇论文。

Arxiv ID2609.12441幻觉翻译2609.12441

⚡ IMPLY: Physically Anchored Consistency for World-Model Rollouts

世界模型 rollout 的物理锚定一致性：不只看"自洽"，要看反演出的物理量是否被证据锚定。

Arxiv ID2609.01343幻觉翻译2609.01343

⚡ SMELT: Scaling Laws for Compute-Matched MoE Looped Transformers

looped MoE 的 compute-matched 缩放律：中间一半层循环两次，三预算对齐后仍省 6.8–18% FLOPs。

SMELT: Scaling Laws for Compute-Matched MoE Looped Transformers
Arxiv ID2608.27406幻觉翻译2608.27406

⚡ CLAP: Cross-Embodiment Video World Models are Zero-Shot Physical Simulators

跨具身视频世界模型＝零样本物理模拟器：统一动作接口 + 课程式跨具身学习。

CLAP: Cross-Embodiment Video World Models are Zero-Shot Physical Simulators
Arxiv ID2609.12317幻觉翻译2609.12317

⚡ Sampling via Decision-Flow: Training-Free Extraction of Improved Latent Reasoning Paths in Large Language Models

训练-free 提取 latent 推理路径：全局轨迹评估 + 效用回传，GPQA 45.6 反超 RL 训练。

Sampling via Decision-Flow: Training-Free Extraction of Improved Latent Reasoning Paths in Large Language Models
Arxiv ID2609.12090幻觉翻译2609.12090

⚡ Does Video Memory Use What It Retrieves? A Causal Audit of Memory Specificity

视频记忆的因果审计：读时替换分离"记忆增益"与"记忆特异性"。

Does Video Memory Use What It Retrieves? A Causal Audit of Memory Specificity
Arxiv ID2609.12371幻觉翻译2609.12371

⚡ READ: Learning Risk-Informed Fields for End-to-End Autonomous Driving

E2E 自动驾驶的学习式风险场：显式、可微、planning-aligned 的时空风险表示。

READ: Learning Risk-Informed Fields for End-to-End Autonomous Driving
Arxiv ID2609.11947幻觉翻译2609.11947

⚡ Scenario-Independent Criticality Assessment and Prediction for Vulnerable Road Users in Autonomous Driving

VRU 专用 criticality 度量 + 场景无关预测框架。

Scenario-Independent Criticality Assessment and Prediction for Vulnerable Road Users in Autonomous Driving
Arxiv ID2609.06746幻觉翻译2609.06746

⚡ Reason Through the Latent! Making Latent Visual Reasoning Necessary

让 latent 视觉推理"因果必要"：解码前删掉视觉 KV，只让递归状态携带图像信息。

Reason Through the Latent! Making Latent Visual Reasoning Necessary
Arxiv ID2609.12641幻觉翻译2609.12641

⚡ Breaking the Vision-Action Shortcut: Latent Interface Training for Generalizable Robotics Foundation Models

拆"视觉-动作捷径"：两阶段 latent interface 训练。

Breaking the Vision-Action Shortcut: Latent Interface Training for Generalizable Robotics Foundation Models
Arxiv ID2609.12036幻觉翻译2609.12036

⚡ Pelican-Sim 1.0: A General World Model Simulator for Embodied Intelligence

通用世界模型模拟器技术报告：统一动作空间 + 动作-视觉注入 + 稀疏 MoE。

Pelican-Sim 1.0: A General World Model Simulator for Embodied Intelligence

自动生成于 2026-09-15 · 基于 arXiv Daily Digest