今日从 arXiv 订阅中筛选 10 篇论文。

Arxiv ID2609.21474幻觉翻译2609.21474

⚡ MT-WAM: Reorienting the One-Pass Predictive Representation Toward Action Generation

单次前向的预测表示，按"动作需要什么"重新定向。

MT-WAM: Reorienting the One-Pass Predictive Representation Toward Action Generation
Arxiv ID2609.21662幻觉翻译2609.21662

⚡ When Steering Fails in Latent Reasoning: A Latent-to-Language Transition Gap

latent CoT 的干预为什么打不动输出？作者把它做成一个可检验的"转移间隙"假设。

When Steering Fails in Latent Reasoning: A Latent-to-Language Transition Gap
Arxiv ID2609.21211幻觉翻译2609.21211

⚡ Stochastic Neural Signed Swept Volume for Real-time Chance-Constrained Trajectory Optimization

把连续碰撞检查的概率性缺口补上：扫掠体积＝概率符号距离场。

Stochastic Neural Signed Swept Volume for Real-time Chance-Constrained Trajectory Optimization
Arxiv ID2609.21605幻觉翻译2609.21605

⚡ Trading Depth for Time in Recurrent Transformers

受控对照：多算几步 vs 多堆几层，参数效率差一半。

Trading Depth for Time in Recurrent Transformers
Arxiv ID2609.21748幻觉翻译2609.21748

⚡ World Modeling in Transformers

行为失败判不了世界模型：TaxiGPT 的机理翻案。

World Modeling in Transformers
Arxiv ID2609.21155幻觉翻译2609.21155

⚡ Same World, Different Knowledge: When Isolated Audits Misjudge World-Model Repairs

孤立审计会选错修复：世界模型的"信息接口审计"。

Same World, Different Knowledge: When Isolated Audits Misjudge World-Model Repairs
Arxiv ID2609.21486幻觉翻译2609.21486

⚡ Driving on Registers, Reasoning on Risk: Risk-Aware Occupancy for Register-Based End-to-End Autonomous Driving

把"候选差、排序乱"量化，再用风险占位同时修两者。

Driving on Registers, Reasoning on Risk: Risk-Aware Occupancy for Register-Based End-to-End Autonomous Driving
Arxiv ID2609.21514幻觉翻译2609.21514

⚡ Skel-WAM: A Hand-Skeleton-Conditioned World Action Model for Human-to-Robot Manipulation Transfer

human video 变机器人数据：用骨架统一运动接口。

Skel-WAM: A Hand-Skeleton-Conditioned World Action Model for Human-to-Robot Manipulation Transfer
Arxiv ID2609.21346幻觉翻译2609.21346

⚡ IntBMoE: Integrating Block-Level Conditioning into Expert Composition for Full-Participation Mixture-of-Experts

MoE 设计空间三量：参与、执行、物化。

Arxiv ID2609.21521幻觉翻译2609.21521

⚡ VidOmni-Bench: A Benchmark for Fine-Grained Video Understanding via Spatio-Temporal Event Verification across Complexity and Duration

视频理解评测：从"答对问题"换成"验证事件"。

VidOmni-Bench: A Benchmark for Fine-Grained Video Understanding via Spatio-Temporal Event Verification across Complexity and Duration

自动生成于 2026-09-22 · 基于 arXiv Daily Digest