How Compliant is Sepsis Treatment? An Expert-Guided Neuro-symbolic Pipeline for Generating Clinical Compliance Insights

arXiv:2608.…

• 13617v1 Announce Type: new Abstract: Verifying whether clinical care f…
• We present an expert-guided pipeline that constrains a large language …
• The fuzzy layer encodes eight Surviving Sepsis Campaign bundle rules a…

RSS 官方收录 · 可信分层展示

SemPlan: Benchmarking Structured Semantic Planning for LLM-Based Queries over Enterprise Data

arXiv:2608.…

• 13612v1 Announce Type: new Abstract: Natural-language interfaces to en…
• SemPlan Benchmark evaluates this architectural design space with a det…
• Four architectures are compared under the same model configuration: di…

RSS 官方收录 · 可信分层展示

Evaluating Agentic Learning Harness Capabilities Without Labels via the Scaling Hypothesis

arXiv:2608.…

• 13608v1 Announce Type: new Abstract: Agentic "Continual Learning Harne…
• But their value is conventionally measured by gains against labeled be…
• Benchmark labels are scarce, stale, and unrepresentative, so a practit…

RSS 官方收录 · 可信分层展示

No Universal Signal Predicts Sample-Level LLM Regression under Version Updates

arXiv:2608.13607v1 Announce Type: new Abstract: Frontier LLMs are updated frequently and typically outperform their predecessors in aggregate.…

• But aggregate gains say little about individual samples: an update can…
• This paper studies how to predict such regressions from signals availa…
• We compare single-model signals (confidence, logit margin, attention e…

RSS 官方收录 · 可信分层展示

MobileMem: Learning from a Year of Mobile Experiences

arXiv:2608.…

• 13606v1 Announce Type: new Abstract: The next generation of AI agents …
• Such assistants require long-term memory to accumulate and leverage us…
• We introduce MobileMem, a benchmark and framework for studying on-devi…

RSS 官方收录 · 可信分层展示

调用成功不等于判断正确：KDC 的行动治理主张

点击查看原文>

RSS 官方收录 · 可信分层展示

Vibe check：你的AI产品真的能落地吗

点击查看原文>

RSS 官方收录 · 可信分层展示

拒绝「差不多」，电商模型测评迎来「最严厉的父亲」

测评 AI 这件事，如今越来越难做了。几年前，AI 还是聊天机器人时，测评主要就是做题：写作的、数学的、编程的……最后形成一套跑分，用分数代表模型能力。但当 AI 从聊天框里走出来，变成能够调用工具、操作网页、处理文件、跨系统执行任务的 Agent，情况就变得复杂起来。…

• 平时我们看到的，是 AI 出现在手机上，能够刷淘宝、加购物车，最后下单。
• 实际上在商家的那一端，也有 AI 的参与，承担着比选购和下单更复杂的工作。
• 在专门考察电商环境中模型表现的 RealReplicaBench 测试中，经过 107 个真实商业任务的考核，所有参评模型都没有达到 60 …

RSS 官方收录 · 可信分层展示

Cross-Disciplinary Taxonomy and Modeling of Misunderstanding Generation, Amplification, and Detection, from Pragmatics to AI Agents

arXiv:2608.…

• 13604v1 Announce Type: new Abstract: Detection of misunderstanding is …
• This shift cuts communicators off from the resources repair depends on…
• In this paper we analyse misunderstanding as a layered process in whic…

RSS 官方收录 · 可信分层展示

Active Perception for Embodied Disambiguation

arXiv:2608.…

• 13605v1 Announce Type: new Abstract: Natural language provides robots …
• Existing interactive disambiguation methods primarily obtain additiona…
• We propose an active-perception framework for embodied target disambig…

RSS 官方收录 · 可信分层展示

Measuring Cross-Task Behavioral Consistency in Language Model Agents

arXiv:2608.…

• 13598v1 Announce Type: new Abstract: Agent evaluation relies almost en…
• We argue that behavioral consistency across tasks is a distinct and me…
• BCM trains a model to predict task success from behavioral features of…

RSS 官方收录 · 可信分层展示

Stable Miscalibration in Large Language Models: A Practical View of High-Confidence Errors

arXiv:2608.13591v1 Announce Type: new Abstract: High-confidence errors in large language models are often treated as evidence of fragile internal inference.…

• We study a different possibility: stable miscalibration, where a confi…
• We combine two diagnostics: a label-aware output-level audit score tha…
• On a multi-domain binary factual audit set, this audit score tracks wh…

RSS 官方收录 · 可信分层展示

AI Evaluation Should Work With Humans

arXiv:2608.…

• 13577v1 Announce Type: new Abstract: This position paper argues that t…
• Instead, the AI community should pivot to evaluating the performance o…
• We argue that this collaborative shift will foster AI systems that act…

RSS 官方收录 · 可信分层展示

Agentao: A Governed Local-First Runtime for Tool-Using LLM Agents

arXiv:2608.…

• 13574v1 Announce Type: new Abstract: LLM agents increasingly operate a…
• These capabilities make agents useful, but they also introduce risks r…
• This paper presents Agentao, a governed local-first runtime for tool-u…

RSS 官方收录 · 可信分层展示

A Year in LLM Serving: Workload Evolution, Caching and Load-Balancing

arXiv:2608.…

• 13573v1 Announce Type: new Abstract: Large Language Model (LLM) servin…
• However, existing LLM serving workload studies remain limited in scale…
• They often observe short time periods and provide limited visibility i…

RSS 官方收录 · 可信分层展示

Modular Cognitive Architecture Emerges in Large Language Models

arXiv:2608.…

• 13567v1 Announce Type: new Abstract: The human brain exhibits a striki…
• Is this modular organization a fundamental principle of how intelligen…
• Here, we test whether a similar organization emerges in Large Language…

RSS 官方收录 · 可信分层展示

Depth-Aware Sensitivity Analysis of Mixture-of-Experts Models via Magnitude-Based Expert Masking

arXiv:2608.…

• 13565v1 Announce Type: new Abstract: Mixture-of-Experts (MoE) architec…
• Despite their widespread adoption, the relative importance of individu…
• This paper presents a systematic layer-wise sensitivity analysis of th…

RSS 官方收录 · 可信分层展示

Inducing Reward-Free Judging Rubrics that Reduce Over-Crediting in Agent Evaluation

arXiv:2608.…

• 13564v1 Announce Type: new Abstract: Evaluating language-model agents …
• Such a judge is a reward-free proxy whose value depends on whether it …
• We instead induce the text of an agent-judging rubric from a small set…

RSS 官方收录 · 可信分层展示

闭门、举牌、真话，这场路演有点不一样

作者｜Li Yuan 编辑｜郑玄 会算命的非遗挂件？专挑膝盖下手的外骨骼？会跳的机器人？8 月 7 日，地瓜机器人办的这场闭门路演「地心引力 DemoDay 2026」，十个项目轮番上台，画风一个比一个野。…

• 有做具身智能的，也有做智能硬件的，大部分都还初具雏形：一些项目，半年前才刚正式启动。
• 地瓜选择本场 Demoday 选手的调性，与其他不太一样。
• CEO 王丛开场就说：他不太在乎地瓜自己出货量多大、市场占有率多高，他更想看到的画面是，被他们孵化出来的那些五花八门的东西，能被随便拿一个回…

RSS 官方收录 · 可信分层展示

世界机器人大会今年不一定「Wow！」，但有五个问题很值得关注

作者｜Li Yuan 编辑｜张鹏 8 月 10 日，宇树科技完成科创板申购，发行价 150.80 元，发行市值约 610 亿元，发行市盈率 219 倍，是通用设备行业平均值的近六倍。…

• 多家机构给出的挂牌后估值直接标到千亿以上，敲钟的日子很可能就落在这个月下旬，几乎与世界机器人大会撞在一起。
• 对标电动车，意味着一个万亿级的制造业；对标 AI，则意味着根本没有天花板——毕竟具身智能被寄予的，是让 AI 真正落进物理世界。
• 热的是估值，冷的是订单。

RSS 官方收录 · 可信分层展示

Dropbox 集成 MCP 与 Dash，将安全设计与代码审查连接起来

点击查看原文>

RSS 官方收录 · 可信分层展示

编程能力提高50%！GLM-5.3 满分通过了GPT-5.6给的Coding 测试

点击查看原文>

RSS 官方收录 · 可信分层展示

中行回应“Token贷”：已投放3户800万元；头部AI大厂员工：90小时工作制成常态；宇树科技中签者不敢发朋友圈：怕被嫉妒｜AI 周报

点击查看原文>

RSS 官方收录 · 可信分层展示

问界「童车」上市，华为联合设计；DeepSeek 涨价策略今日实行；大学生用 AI 人脸视频盗刷 5 万元被判刑

IPO 前 Anthropic CEO 达里奥 · 阿莫迪罕见发长文回应质疑，预告未来 5-10 年 AI 将治愈多数疾病 Anthropic CEO 达里奥 · 阿莫迪（Dario Amodei）几乎从不碰社交媒体。…

• 但就在 IPO 估值被推到 2 万亿美元的节骨眼上，他罕见地发了一篇长文，正面回应 Anthropic 如何看待监管。
• 长文中，达里奥 · 阿莫迪回「我们设计的每一条政策，都让前沿公司承担更多成本，同时有利于规模更小的竞争者。
• 」他同样支持 Demis Hassabis 提出的、建立一个类似 FINRA（美国金融业监管局）的 AI 监管机构的想法。

RSS 官方收录 · 可信分层展示