今日从 arXiv 订阅中筛选 10 篇论文。

Arxiv ID2609.28991幻觉翻译2609.28991

⚡ Beneath the Scores: Rethinking Hallucination Evaluation for Video Understanding Models

视频理解 agent 幻觉评估的因果重建：grounding 是下游错误主源（≈破坏视觉观测的 4 倍），错误证据比缺失证据更害。

Beneath the Scores: Rethinking Hallucination Evaluation for Video Understanding Models
Arxiv ID2609.28931幻觉翻译2609.28931

⚡ HelloWorld: Towards Practical Applications of Generative Driving World Models

驾驶世界模型的规模化系统：2B 模型、7 摄像头同步生成 + 条件 LiDAR、少步蒸馏，全维度评估。

HelloWorld: Towards Practical Applications of Generative Driving World Models
Arxiv ID2609.30264幻觉翻译2609.30264

⚡ AD-WM: Action-Discriminative World Models for Counterfactual Model Predictive Control

反事实 MPC 的动作判别世界模型：OGBench-Cube hard-start 3.7%→52.0%，并给出"预测误差不跟随闭环成功"的诊断。

AD-WM: Action-Discriminative World Models for Counterfactual Model Predictive Control
Arxiv ID2609.29555幻觉翻译2609.29555

⚡ Visual Representation and History Modeling for Navigation World Models

导航世界模型的表示之问：5 种冻结表示的优劣随目标而变，历史计算可复用。

Visual Representation and History Modeling for Navigation World Models
Arxiv ID2609.30036幻觉翻译2609.30036

⚡ Aim Short to Reach Far: Your Frozen World Model Can Plan Better Than You Think

冻结世界模型的目标位置诊断：Anchored Planning 零训练全面超原版 LeWM planner。

Arxiv ID2609.29171幻觉翻译2609.29171

⚡ Representation World Model: Learning States, Transition and Executable Plans in Representation

表示世界模型：规划直接在表示几何里发生，无递归 rollout、无动作空间搜索。

Arxiv ID2609.29706幻觉翻译2609.29706

⚡ Safety-oriented pedestrian trajectory prediction at urban intersections using time-to-collision and crossing-zone context

安全导向行人轨迹预测：TTC+过街上下文，"超差计数"指标压大误差。

Arxiv ID2609.29738幻觉翻译2609.29738

⚡ Combining Evasive and Braking Reactions for Safety Reference Models in Automated Vehicles

ADS 场景测试的"谨慎人类驾驶员"基准：制动+避让混合状态机，可避免碰撞包络扩大。

Combining Evasive and Braking Reactions for Safety Reference Models in Automated Vehicles
Arxiv ID2609.28984幻觉翻译2609.28984

⚡ CrossSafe: Towards Cross-Embodiment Latent Safety Filters

跨形态潜空间安全滤波：安全推理共享、安全实现随身体，5 形态零样本泛化。

CrossSafe: Towards Cross-Embodiment Latent Safety Filters
Arxiv ID2609.29964幻觉翻译2609.29964

⚡ World Action Agent: Harnessing VLMs for Robot Manipulation via World Action Rehearsal

VLM 开机器人：视觉动作工作台（接触视图/动作彩排/视野内纠偏）+ 技能演化，LIBERO-Pro 75.6%。

World Action Agent: Harnessing VLMs for Robot Manipulation via World Action Rehearsal

自动生成于 2026-09-26 · 基于 arXiv Daily Digest