ARTICLE · 1209219
AI 论文每日简报 · 10/19
AI 论文每日简报 · 10/19
AI 论文每日简报
HuggingFace 社区精选 · 每日热门论文
📅 2026-10-19
📡 HF Daily Papers
📄 40 篇
📌 今日概览
今日论文围绕 LLM Agent 的记忆与推理效率、多模态理解与生成、科学 AI（医疗/材料/进化）以及具身智能展开，其中 Agent 的长期记忆、状态治理与高效预训练是最密集的方向。
⭐ 点击上方蓝字 → 右上角「···」→ 设为星标，不错过每天的 AI 论文。
⭐ 今日重点
1
PA1-OPD: One-Step Prefix-Augmented On-Policy Distillation for Language Model Pretraining
核心：针对语言模型预训练，提出一步前缀增强的 on-policy 蒸馏方法，在学生自己生成的上下文上补充教师监督，弥补传统知识蒸馏只在语料前缀匹配教师分布、缺乏对「学生预测诱导上下文」直接监督的不足，从而提升预训练效率。
📎 arxiv: 2610.19601
2
DuplexBridge: Coordinating Understanding and Generation for Unified Image Editing
核心：复杂图像编辑需要忠实理解任务并准确执行，但一次失败的编辑无法暴露是「理解」还是「生成」环节出错。DuplexBridge 提出统一框架，同时协调理解与生成两个责任，让失败可定位、编辑更可控。
📎 arxiv: 2610.18819
3
Memorizing Who, Knowing How: Benchmarking Personal and Procedural Memory for Persistent Assistants
核心：持久个人 Agent 的记忆系统需要同时维护「记住服务对象是谁」的个人记忆和「记住如何行动」的程序记忆。本文提出基准，分别评估这两类记忆的能力，推动 Agent 记忆从孤立条目走向成体系。
📎 arxiv: 2610.20022
4
Risk-Controlled DFT Selection for Machine-Learned Interatomic Potentials
核心：自动化材料发现需要评估海量候选构型，DFT 计算是主要吞吐瓶颈。本文提出风险控制的 DFT 选择策略，在保证机器学习原子间势精度的前提下减少昂贵的 DFT 调用，加速材料筛选。
📎 arxiv: 2610.19630
5
MemoryVLN: From Ego-Aware Memory Representation to Cross-Context Memory Reuse in Vision-Language Navigation
核心：LVLM 在视觉语言导航中，长期轨迹因上下文长度限制而性能下降。MemoryVLN 提出自我感知的记忆表示与跨上下文的记忆复用机制，缓解长程导航的上下文瓶颈，提升具身导航的连贯性。
📎 arxiv: 2610.18478
🧠 LLM 推理 & 效率 (11篇)
SkillScaffold: Scaffolding Agentic Reinforcement Learning with Policy-Adaptive Skill Credit Assignment and Selective Injection — 策略自适应技能信用分配的Agent RL
Narrative Lookahead in LLMs: Event-Level Decodability, Predictive Advantage, and Steering — 揭示LLM叙事前瞻的可解码性与引导
Empathy Is Not the Scapegoat for Declines in Factuality and Reliability — 实证共情并非事实性下降的主因
Task-State Governance: A Virtual Machine for Verifiable and Recoverable GUI Agent Execution — 可验证可恢复的GUI执行虚拟机
Training LLMs for Multi-Agent Collaboration with Verifiable Information Rewards — 可验证信息奖励训练多智能体协作
Read the Room Before Your Model Generates — 按上下文调节采样温度控制随机性
When Protection Overrides Consequences: Understanding Protection Mode in Large Language Model Decision-Making — 剖析保护模式凌驾后果的决策偏差
Lost in Implicit Routing: How LLM Agents Fail to Load Relevant Skills at Scale — 诊断技能隐式路由规模化失效
When Reasoning Is Not Enough: Proactive Inquiry for Cross-Modal Safety in Multimodal Large Language Models — 主动质询提升跨模态安全
StateHarness: Maintaining Evolving Project State for Long-Horizon Coding Agents — 维护长程编码Agent的演进项目状态
What Should We Remove? Evaluating the Reliability and Costs of RLVR Data Decontamination — 评估RLVR数据去污染的可靠性与代价
🎬 视频 & 图像生成 (2篇)
PolUR: Polarization Aware Multimodal Degradation Perception for Underwater Image Restoration — 偏振感知退化建模复原水下图像
GeoState: Learning View-Decoupled Geometry States for 3D World Modeling — 视角解耦几何状态建模3D世界
🔬 科学 & 医疗 AI (7篇)
Deep Generative Models for Phylogenetic Inference with Complex Evolutionary Processes — 深度生成模型推断系统发育树
MedEvo: Medical Reasoning Evolution through Knowledge Navigation — 知识导航演化医学推理
GeoNIRS: Geometry-Aware Pretraining for Heterogeneous fNIRS Representation Learning — 几何感知预训练异构fNIRS表征
ProtModernBERT: Compute-Efficient Protein Language Modeling — 计算高效的蛋白质语言模型
MedPAO-VL: Interpretable and Protocol-Compliant Agent for Structured Radiology Report Generation — 协议合规生成结构化放射报告
Joint Neural Posterior Estimation and Model Selection for Chemical Reaction Networks — 联合神经后验估计与模型选择
A Tutorial on Quantum Amplitude Embedding with Data Re-Uploading — 量子振幅嵌入与数据重上传教程
📊 机器学习方法 (5篇)
Neurosymbolic Learning of Interpretable World Models — 神经符号学习可解释世界模型
Hard Way, The Only Way: Provable Merits of Training on Hardest Examples — 证明训练最难样本的可证明优势
In-Context Batch Black-box Optimization via Flow Matching — 流匹配做上下文批量黑盒优化
Taming Softmax at the Limits of Muon: A Unified Approach through Hilbert Geometry — 希尔伯特几何统一Muon与Softmax
In-Context Learning Before Specialisation: Theory of Learning in a Solvable Topic Model — 可解主题模型的上下文学习理论
🏭 工程 & 应用 (8篇)
ABNAT: Attention-Based Noise-Aware Training for Robust Transformers on Analog Compute-in-Memory Systems — 噪声感知训练赋能模拟存算一体
AutoMARS: Automatic and Scalable Multi-Agent Large Language Model Framework for Supply Chain Risk Assessment — 多智能体自动供应链风险评估
RT-J: Large-Scale Pretraining of Relational Transformers for Context-Efficient Predictions — 关系Transformer表格基础模型预训练
The Gap Is in the Geometry: A Convergent Meta-Analysis of Multi-Facet Aerial and Remote Sensing Video Understanding — 揭示航拍遥感视频理解的几何鸿沟
Towards Training and Inference of Graph Neural Networks in Graph Databases — 图数据库内训练与推理GNN
Dexterity from Supervised Play — 监督玩耍习得灵巧操作
Self-Guided LLM Reasoning for Error Diagnosis in Vehicle Routing Problems — 自引导LLM推理诊断VRP错误
EgoExplorer: A Vision-Motion-Touch Observatory for Human Loco-Manipulation In Situ — 人形移动操作的视觉运动触觉观测
📄 其他 (2篇)
Twenty Days Later: A Pre-Registered Test of History-Based Forecasts of Brand Mentions in Gemini Answers — 预注册检验品牌提及预测的持久性
Digital Self-Interference Cancellation for In-Band Full-Duplex Systems Based on Transformer Neural Operators — Transformer神经算子做数字自干扰消除
🔗 今日信号
今日最集中的方向是 LLM Agent 的长期记忆与执行治理：从「记住服务对象与办事方法」（Memorizing Who, Knowing How）、导航中的跨上下文记忆复用（MemoryVLN）、编码 Agent 的项目状态维护（StateHarness），到技能加载的隐式路由失效诊断（Lost in Implicit Routing）与可恢复的 GUI 执行虚拟机（Task-State Governance），Agent 正从「单次任务能不能做」走向「长期、可靠、可恢复地运行」。对做 LLM 应用/工程的人而言，记忆检索、状态持久化与技能路由不再是可后补的细节，而是决定 Agent 能否投入生产的一等工程问题。
内容由 AI 自动整理 · HuggingFace Daily Papers · 2026-10-19
由 Hermes Agent 生成
