2026.06.18 | 多模态大模型记忆成短板；语言指令驱动3D轨迹预测

【赞助商】OpenClaw快报每天五分钟，听听 OpenClaw 快报，带你了解最新动态和业内讨论传送门 https://www.xiaoyuzhoufm.com/podcast/6a1732a2dffa135d0ab5ef43【目录】本期的 15 篇论文如下：[00:32] 🧠 Beyond the Current Observation: Evaluating Multimodal Large Language Models in Controllable Non-Markov Games（超越当前观测：评估多模态大语言模型在可控非马尔可夫博弈中的表现）[01:29] 🎯 MolmoMotion: Forecasting Point Trajectories in 3D with Language Instruction（MolmoMotion：基于语言指令的3D点轨迹预测）[02:15] 🌍 Kairos: A Native World Model Stack for Physical AI（Kairos：面向物理智能的原生世界模型栈）[03:05] 🛠 Guava: An Effective and Universal Harness for Embodied Manipulation（番石榴：一种有效且通用的具身操作框架）[03:58] ⚡ EfficientRollout: System-Aware Self-Speculative Decoding for RL Rollouts（高效展开：面向强化学习展开的系统感知自推测解码）[04:45] 🎯 The Reward Was in Your Data All Along: Correcting Flow Matching with Discriminator-Guided RL（奖励一直就在你的数据中：用判别器引导的强化学习纠正流匹配）[05:51] 🔍 SAE Interventions are Unreliable: Post-Intervention Recovery of Suppressed Behavior（SAE干预不可靠：抑制行为在干预后的恢复）[06:36] 🤖 From Trainee to Trainer: LLM-Designed Training Environment for RL with Multi-Agent Reasoning（从受训者到训练者：面向多智能体推理的LLM设计训练环境强化学习框架）[07:27] 🧠 Reinforcing Dual-Path Reasoning in Spatial Vision Language Models（增强空间视觉语言模型中的双路径推理）[08:25] 🎯 Trust the Right Teacher: Quality-Aware Self-Distillation for GUI Grounding（信任正确的教师：面向GUI定位的质量感知自蒸馏方法）[09:25] 👁 Native Active Perception as Reasoning for Omni-Modal Understanding（原生主动感知作为全模态理解的推理）[10:15] 🐱 MaineCoon: Pursuing A Real-Time Audio-Visual Social World Model（缅因猫：追求实时的音视频社交世界模型）[11:08] 🖌 Sumi: Open Uniform Diffusion Language Model from Scratch（Sumi：从头构建的开放均匀扩散语言模型）[11:51] 🎯 STARE: Surprisal-Guided Token-Level Advantage Reweighting for Policy Entropy Stability（STARE：基于惊异度的令牌级优势重加权以实现策略熵稳定性）[12:46] 🌍 Beyond Alignment: Value Diversity as a Collective Property in Multicultural Agent Systems（超越对齐：价值多样性作为多元文化智能体系统中的集体属性）【关注我们】您还可以在以下平台找到我们，获得播客内容以外更多信息小红书: AI速递在小宇宙查看该单集文稿