2026.05.22 | 大模型内化地理空间；判别性令牌优化推理

【目录】本期的 15 篇论文如下：[00:25] 🚌 TransitLM: A Large-Scale Dataset and Benchmark for Map-Free Transit Route Generation（TransitLM: 面向无地图公交路线生成的大规模数据集与基准）[01:21] 🎯 DelTA: Discriminative Token Credit Assignment for Reinforcement Learning from Verifiable Rewards（DelTA：面向可验证奖励强化学习的判别性令牌信用分配）[02:15] 🤖 $π$-Bench: Evaluating Proactive Personal Assistant Agents in Long-Horizon Workflows（π-Bench：评估长时工作流中的主动式个人助理代理）[03:04] 🤔 Perception or Prejudice: Can MLLMs Go Beyond First Impressions of Personality?（感知还是偏见：多模态大语言模型能否超越对个性的第一印象？）[03:51] 🔥 Full Attention Strikes Back: Transferring Full Attention into Sparse within Hundred Training Steps（全注意力回归：在百步训练内将全注意力迁移至稀疏注意力）[04:35] 🤖 ACC: Compiling Agent Trajectories for Long-Context Training（ACC：为长上下文训练编译智能体轨迹）[05:35] 🧊 PhysX-Omni: Unified Simulation-Ready Physical 3D Generation for Rigid, Deformable, and Articulated Objects（PhysX-Omni：面向刚体、可变形体和关节物体的统一仿真就绪物理3D生成）[06:37] 🧠 LatentOmni: Rethinking Omni-Modal Understanding via Unified Audio-Visual Latent Reasoning（LatentOmni：通过统一音频-视觉潜在推理重新思考全模态理解）[07:37] 🌍 WorldKV: Efficient World Memory with World Retrieval and Compression（世界KV：结合世界检索与压缩的高效世界记忆）[08:22] 📊 Spreadsheet-RL: Advancing Large Language Model Agents on Realistic Spreadsheet Tasks via Reinforcement Learning（Spreadsheet-RL：通过强化学习推进大型语言模型智能体在真实电子表格任务中的应用）[09:27] 🎥 FlowLong: Inference-time Long Video Generation via Manifold-constrained Tweedie Matching（FlowLong：基于流形约束Tweedie匹配的推理时长视频生成）[10:35] 🧠 SpaceDG: Benchmarking Spatial Intelligence under Visual Degradation（SpaceDG：视觉退化下的空间智能基准测试）[11:35] 🎯 Maestro: Reinforcement Learning to Orchestrate Hierarchical Model-Skill Ensembles（Maestro：强化学习编排层次化模型-技能集成）[12:29] 🚗 Sensor2Sensor: Cross-Embodiment Sensor Conversion for Autonomous Driving（传感器到传感器：面向自动驾驶的跨本体传感器转换）[13:21] 🎥 Q-ARVD: Quantizing Autoregressive Video Diffusion Models（Q-ARVD：量化自回归视频扩散模型）【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2026.05.22 | 大模型内化地理空间；判别性令牌优化推理在小宇宙查看该单集文稿