今日从 arXiv 订阅中筛选 10 篇论文。

Arxiv ID2609.27296幻觉翻译2609.27296

⚡ Teach-to-Crash: A Closed-Loop Student-Teacher LLM Framework for Collision-Inducing Test Scenario Generation

碰撞场景生成闭环：双 LLM 教师-学生架构在搜索停滞时介入，CARLA 上 90.79% 碰撞命中率与最高多样性。

Teach-to-Crash: A Closed-Loop Student-Teacher LLM Framework for Collision-Inducing Test Scenario Generation
Arxiv ID2609.27070幻觉翻译2609.27070

⚡ The Gaussian Is Enough: Flow-Matching Priors Do Not Help When Fine-Tuning Large Behavior Models

负面结果：微调 LBM 时更"接近目标"的非高斯先验并不比标准高斯好，编码器训练才是主导。

The Gaussian Is Enough: Flow-Matching Priors Do Not Help When Fine-Tuning Large Behavior Models
Arxiv ID2609.27247幻觉翻译2609.27247

⚡ Memory That Changes Action Is Not Memory That Guides It: Counterfactual Auditing of History-Conditioned Robot Policies

机器人策略记忆的决定层审计：记忆改变动作 ≠ 记忆引导决策，Mem-0 上仅 20/64 对完全可靠。

Memory That Changes Action Is Not Memory That Guides It: Counterfactual Auditing of History-Conditioned Robot Policies
Arxiv ID2609.28414幻觉翻译2609.28414

⚡ Frozen Flows Forget: Diagnosing and Restoring Lost Motion in a Latent-flow World Model

冻结流世界模型的静默丢运动：失效在训练信号而非表示，DART 只重训流即可修复。

Frozen Flows Forget: Diagnosing and Restoring Lost Motion in a Latent-flow World Model
Arxiv ID2609.27455幻觉翻译2609.27455

⚡ Latent evolving World Action Model

WAM 表示之问：JEPA 预测嵌入比 VAE 压缩潜更能支撑动作生成，LeWAM 0.4B 参数达 92.28% RoboTwin 2.0。

Latent evolving World Action Model
Arxiv ID2609.27038幻觉翻译2609.27038

⚡ Are Stated Reasoning Steps Causally Load-Bearing?

CoT 承重的激活级因果测试：Qwen3-4B 76.9% 陈述步为因果承重，行为测试高估 11.4pp。

Arxiv ID2609.27747幻觉翻译2609.27747

⚡ Less Language, More Latents: Annotation-Efficient VLAs for Driving

驾驶 VLA 的潜动作路线：<5% 语言标注 + VQ 码本，Bench2Drive 闭环 DS 87.98。

Less Language, More Latents: Annotation-Efficient VLAs for Driving
Arxiv ID2609.28366幻觉翻译2609.28366

⚡ AnchorReasoning: A Visual Grounding and Causal Reasoning Dataset in Long-Tail Autonomous Driving Scenarios

长尾 AD 的视觉落地推理数据集：416K 帧 VG-CoT，8 backbone 上 ADE/FDE 显著下降且推理更省。

AnchorReasoning: A Visual Grounding and Causal Reasoning Dataset in Long-Tail Autonomous Driving Scenarios
Arxiv ID2609.28161幻觉翻译2609.28161

⚡ Dissecting Advantage-Guided Post-Training for Vision-Language-Action Policies

VLA 后训练的受控拆解：分离耦合设计选择，模组配方在 4 个真机双臂任务超 SFT 初值。

Dissecting Advantage-Guided Post-Training for Vision-Language-Action Policies
Arxiv ID2609.28086幻觉翻译2609.28086

⚡ LAYERSCOPE: A Layerwise Characterization of Video and Multimodal Learned Representations

视频/多模态表示的分层刻画：无标签几何指标跨层扫描，中层可超末层、无单一指标稳定预测下游。

LAYERSCOPE: A Layerwise Characterization of Video and Multimodal Learned Representations

自动生成于 2026-09-25 · 基于 arXiv Daily Digest