本期的 13 篇论文如下：

[00:25] 🧩 IterResearch: Rethinking Long-Horizon Agents via Markovian State Reconstruction（IterResearch：基于马尔可夫状态重构的长程智能体再思考）

[01:16] 🏆 DRIVE: Data Curation Best Practices for Reinforcement Learning with Verifiable Reward in Competitive Code Generation（DRIVE：面向可验证奖励强化学习的竞赛级代码生成数据精选最佳实践）

[02:03] 🔬 The Station: An Open-World Environment for AI-Driven Discovery（“站”：面向AI驱动科学发现的开放世界环境）

[02:43] 🚀 RedOne 2.0: Rethinking Domain-specific LLM Post-Training in Social Networking Services（RedOne 2.0：社交网络场景下领域大模型后训练新范式）

[03:15] 🧠 SofT-GRPO: Surpassing Discrete-Token LLM Reinforcement Learning via Gumbel-Reparameterized Soft-Thinking Policy Optimization（SofT-GRPO：用Gumbel重参数化软思考策略优化让离散Token强化学习望尘莫及）

[03:53] 🧭 Routing Manifold Alignment Improves Generalization of Mixture-of-Experts LLMs（路由流形对齐提升混合专家大语言模型的泛化能力）

[04:30] 🔍 Reasoning with Confidence: Efficient Verification of LLM Reasoning Steps via Uncertainty Heads（以置信度推理：通过不确定性头高效验证大模型推理步骤）

[05:10] 🎬 MVU-Eval: Towards Multi-Video Understanding Evaluation for Multimodal LLMs（MVU-Eval：面向多模态大模型的多视频理解评测基准）

[05:50] 🎨 MPJudge: Towards Perceptual Assessment of Music-Induced Paintings（MPJudge：面向音乐诱发绘画的感知评估）

[06:57] 🔄 RLoop: An Self-Improving Framework for Reinforcement Learning with Iterative Policy Initialization（RLoop：一种通过迭代策略初始化自我提升的强化学习框架）

[07:36] 🤖 Robot Learning from a Physical World Model（基于物理世界模型的机器人学习）

[08:21] 🛠 NURBGen: High-Fidelity Text-to-CAD Generation through LLM-Driven NURBS Modeling（NURBGen：基于大模型驱动NURBS建模的高保真文本转CAD生成）

[08:52] 🚀 SWE-fficiency: Can Language Models Optimize Real-World Repositories on Real Workloads?（SWE-fficiency：语言模型能否在真实工作负载下优化真实仓库性能？）

【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