视频选题榜 2026-08-21

• https://arxiv.org/abs/2608.18746 · Decision-Metric Alignment in Latent World Models: Diagnostics and Action-Conditioned Objectives for MPC Planning · JEPA-MPC 潜码欧氏距离 ≠ 真实奖励排序 · Plan-Real Spearman + CEM-stage Spearman 双相诊断 + DA-LeWM(inverse-dynamics + demo-conditioned goal-action)双头训练 + 决策度量对齐 decision-metric alignment · round=R1
• https://arxiv.org/abs/2608.18613 · CTIFoundry: An Agent-Native Corpus Scaffold for Cyber Threat Intelligence · Agent 原生语料架构 · CTI 语料从 chunk 集合升级为可遍历本体图 + 7 typed tool + 3 procedural skill · 仅替换 action surface F1 +0.19~+0.28 · 一半 tool call 拿更高准确率 · corpus 是 agent 第二杠杆 · round=R2
• https://arxiv.org/abs/2608.18852 · SkillGate: Training In-Policy Skill Selection in Long-Horizon Agents · 长视野 Agent 技能选择的结构性 RL 缺陷(selector credit starvation)+ dual-channel credit 修正 · 9B policy trial success 40.8%→53.2% · round=R3