今日从 arXiv 订阅中筛选 10 篇论文。

Arxiv ID2609.09925幻觉翻译2609.09925

⚡ Time-Frequency Geometric Cross-Attention for Chunked Vision-Language-Action Models

动作 chunk 的表征补丁：时频分解+楔积注意力，对近正交相位敏感，零初始化 drop-in，真机 OOD 增益大于 ID。

Time-Frequency Geometric Cross-Attention for Chunked Vision-Language-Action Models
Arxiv ID2609.10540幻觉翻译2609.10540

⚡ Programmable World Model

世界状态做成"可编程程序"显式维护：状态演化与渲染解耦，可暴力改离屏实体，CombatStateBench 94/98。

Programmable World Model
Arxiv ID2609.09985幻觉翻译2609.09985

⚡ VLX-VR: An Agentic-Aware Video Reasoning Model

视频推理的 agentic 证据闭环：Think–Memory–Observation 循环 + RL，96.2% 推理轨迹与参考证据一致。

VLX-VR: An Agentic-Aware Video Reasoning Model
Arxiv ID2609.09528幻觉翻译2609.09528

⚡ MotionBlind: Probing the Illusion of Motion Understanding in Video-LLMs

Video-LLM 运动理解幻觉的硬证据：开源全在 6.25% 随机线，删视频=0、打乱帧=随机，只有 Gemini3.1 Pro 过线且速度挂。

MotionBlind: Probing the Illusion of Motion Understanding in Video-LLMs
Arxiv ID2609.09928幻觉翻译2609.09928

⚡ Structural Process Supervision for Latent Chain-of-Thought Reasoning

隐 CoT 的过程监督：原型锚做多对多软对齐防坍塌，token<50% 且比 SIM-CoT 平均 +2.08%。

Structural Process Supervision for Latent Chain-of-Thought Reasoning
Arxiv ID2609.10464幻觉翻译2609.10464

⚡ Semigroup-JEPA: Latent Dynamics Consistency for Zero-Shot Physics Generalization

物理参数进 JEPA：重力场 conditioning+latent rollout 联合训练；增益来自 encoder 学到"可携带"特征。

Semigroup-JEPA: Latent Dynamics Consistency for Zero-Shot Physics Generalization
Arxiv ID2609.09163幻觉翻译2609.09163

⚡ World-Time Compute with Verified Code World Models

可验证代码世界当数据制造机：0.5B +29 点、跨世界 +34 点，增益来自"标签精确性"而非多样性。

World-Time Compute with Verified Code World Models
Arxiv ID2609.09881幻觉翻译2609.09881

⚡ CLFTv2: Efficient Camera-LiDAR Fusion for Semantic Segmentation via Hierarchical Feature Pyramids

相机-LiDAR 融合效率向增量：Swin 多尺度+FPN 残差解码，ZOD 行人 IoU +9.4、比 Mask2Former 省 1.4× GFLOPs。

CLFTv2: Efficient Camera-LiDAR Fusion for Semantic Segmentation via Hierarchical Feature Pyramids
Arxiv ID2609.10297幻觉翻译2609.10297

⚡ TRACE: Trajectory-robust Admission with Evidence Ordering for Efficient GUI Agents

剪枝=不可逆准入决策：布局先验+相关性+新颖度排序，预留覆盖修复，嵌套收缩全程可复用。

TRACE: Trajectory-robust Admission with Evidence Ordering for Efficient GUI Agents
Arxiv ID2609.09790幻觉翻译2609.09790

⚡ LogiScope-VQA: Benchmarking Vision-Language Models for Logistics Hazard Identification in Industrial Scenarios

工业仓库危险识别基准：10274 VQA、39 子任务；GPT-5.5/Gemini-3.1-Pro/Claude-Opus-4.7 全低于人类。

LogiScope-VQA: Benchmarking Vision-Language Models for Logistics Hazard Identification in Industrial Scenarios

自动生成于 2026-09-11 · 基于 arXiv Daily Digest