2026.08.24 | 大模型超参迁移降本；图工程引领系统协作

【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] 🚀 Let's Scale Step by Step: Compute-Efficient Hyperparameter Transfer for Large-Scale Mixture-of-Experts（让我们一步一步扩展规模：面向大规模混合专家模型的高效超参数迁移）[01:24] 🕸 Graph Engineering in the Era of LLM Agents: From Individual Intelligence to System Intelligence（大语言模型智能体时代的图工程：从个体智能到系统智能）[02:21] 🤖 OmniAssistBench: Assistant-style Interaction Benchmark for Omni-LLMs（OmniAssistBench：全模态大语言模型的助手式交互基准）[03:17] ⚡ ParaTempo: Efficient Parallel Reasoning via Temporal Confidence（ParaTempo：基于时间置信度的高效并行推理）[04:00] ♾ InfinityEdit: Infinite Video Editing with a Lightweight Edit-Ignition Adapter（InfinityEdit：基于轻量级编辑触发适配器的无限视频编辑）[04:53] 🪙 Every Coin Has Two Sides: On the Dual Nature of Generalization in On-Policy Distillation of Large Language Models（每一枚硬币都有两面：论大型语言模型同策略蒸馏中泛化的双重性）[05:37] 🧩 EviRank: Structured Relevance Evidence for Multimodal Image Re-ranking（EviRank：面向多模态图像重排序的结构化相关性证据）[06:46] 🧠 Beyond Correctness: Benchmarking and Aligning Response Behaviors in Hybrid-Thinking MLLMs（超越正确性：混合思考多模态大语言模型的响应行为基准测试与对齐）[07:45] 🎨 UniSpace: Unified Visual Representation and Scalable Multimodal Modeling（UniSpace：统一视觉表示与可扩展多模态建模）[08:39] 🛒 Towards Faithful Simulation of Human Shopping Behavior（面向人类购物行为的忠实模拟）[09:35] ⚙ AgentMercury: Your Agent Can Synthesize Verifiable Environments for Business Scenarios at scale（AgentMercury：你的智能体能够大规模合成可验证的商业场景环境）[10:41] ⚡ Daedalus-150M: A Convolution-Attention Hybrid Designed for CPU Inference（Daedalus-150M：为CPU推理设计的卷积-注意力混合模型）[11:35] 🛡 CLEAR: Continuous Latent Adapter Routing for Utility-Preserving LLM Safety Alignment（CLEAR：面向保持实用性的大语言模型安全对齐的连续潜在适配器路由）[12:42] 📱 Llama-Mobile: Efficient 2.7-Bit Quantization of VLMs（Llama-Mobile：高效2.7比特视觉语言模型量化）[13:32] 🍳 FlavourBench: Ranking Frontier Language Models with Executable Culinary Ground Truth（FlavourBench：用可执行的烹饪真值对前沿语言模型进行排名）【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