llm-echo 0.5 ↗
Release: llm-echo 0.5 I'm working towards a new LLM release, and testing PRs caused me to revisit the llm-echo debug plugin. This plugin provides a "model" which doesn't involve…
AI 雷达 / 周报
2026-W42
按原始发布周整理来源摘录与阅读线索，不自动生成个人观点或趋势结论。
本周概览
- 信号
- 49
- 主题
- 6
- 来源
- 7
重点信号
先看这些信号
本周动态 / 来源信号
模型
Anthropic 如何在 Claude 写下 Claude Code 约 90% 代码后重新设计代码评审 ↗
作者解读 Anthropic 关于 Claude Code 代码评审重构的博客：Claude 写了约 90% 的代码，2026 年 6 月到 9 月合并 PR 翻倍以上而 revert 率下降超 40%。
Anthropic 披露 Claude 编写了约 90% 的 Claude Code 代码 ↗
Anthropic 的 ClaudeDevs 账号称 Claude 编写了约 90% 的 Claude Code 代码。6 月至 9 月间合并 PR 翻了一倍以上，回退比例下降超过 40%，团队发文介绍了为达到这一效果如何改造代码评审及采用方法，链接见 claude.dev/blog/solving-code-review-at-anthropic/。
Chronoberg: A 250-Year Diachronic Corpus with Temporal Affective Lexicons for Language Model Analysis ↗
Large language models (LLMs) excel at operating at scale by leveraging social media and data crawled from the web. Whereas existing corpora are diverse, their frequent lack of l…
Diffeomorphic Hidden-State Steering: Improving Imitation Learning Policy Robustness through Expert Hidden-State Geometry ↗
Small observation or action errors can compound during closed-loop execution and disrupt a trained imitation learning policy. We study whether the original expert demonstrations…
LIN-SLAM: Lightweight LiDAR--Inertial Neural SLAM for Real-Time Onboard Navigation ↗
Geometric light detection and ranging (LiDAR)--inertial simultaneous localization and mapping (SLAM) provides efficient state estimation but lacks the continuous map representat…
Claude Opus 5.5 fast mode 结束研究预览并移除等待名单 ↗
Anthropic 宣布 Claude 平台上的 Opus 5.5 fast mode 不再处于研究预览阶段，并已移除等待名单。该模式速度最高提升 2.5 倍，定价为 $8 / $40 per million tokens；现已登陆 Microsoft Foundry（Azure 托管），Google Cloud 即将支持。Claude Code 中的…
Chirp Pong ↗
Tool: Chirp Pong Another game experiment - I wanted to see if it was feasible to have a two player game on two separate phones running in static HTML where the networking code u…
Microsoft AI 公布 MAI-Base-2 预训练细节，2.8T 参数 MoE 模型 FLOPs 效率达 DeepSeek-V4-Pro-Base 的 4.2 至 12.3 倍 ↗
Microsoft AI 宣布完成 MAI-Base-2 预训练，这是一个 800 亿激活、2.8T 总参数的 MoE 基础模型，在 held-out 评测中 BPB 低于所测全部开源权重模型，预训练 FLOPs 效率比 DeepSeek-V4-Pro-Base 高 4.2 至 12.3 倍。
阶跃星辰开放 Step 5 Preview 权重 ↗
阶跃星辰（StepFun）宣布 Step 5 Preview 开放权重，BF16 和 FP8 权重已上线 Hugging Face 和 ModelScope，并附技术博客。
LintEval: Benchmarking LLMs on Linter-Guided File Refactoring ↗
Linters are widely used in code maintenance and real-time IDE editing to surface lint alerts related to performance, style, error-prone, and other patterns in source files. Howe…
PersEvalBench: Can LLM Evaluators Judge Contextually Appropriate Use of Persona Information? ↗
Large language model (LLM) evaluators are increasingly used to assess personalized assistants, yet whether they can reliably judge personalization has rarely been tested. We int…
Claude Code 空闲压缩从摘要继续，可用 idleCompaction 关闭并自定义阈值 ↗
Claude Code 的空闲压缩会让 Claude 从摘要而非完整上下文继续。作者说明可在 settings. 中设置 "idleCompaction": false 关闭，并通过 CLAUDE_CODE_IDLE_COMPACT_MIN_TOKENS 自定义阈值，默认为 200k。
Tree Appreciator ↗
Tool: Tree Appreciator Another game experiment with Claude. This one used the default Claude iPhone app (chat mode, not Claude Code mode), where I prompted Claude Opus 5.5 Max w…
Databricks 讲解如何在 AI Runtime 上将 Qwen 3.8 27B 微调时间缩短 6 倍 ↗
Databricks 演示通过系统性性能优化，将 Qwen 3.8 27B 在 AI Runtime 上的微调时间从 7 小时 44 分钟降至 1 小时 17 分钟，提速 6 倍。
PatTransBench: an ipc-stratified, multi-dimensional benchmark for patent machine translation evaluation ↗
We present PatTransBench, a 12,000-item Chinese-English/Japanese/Korean patent machine translation (PatMT) benchmark with a four-dimensional evaluation protocol. The benchmark c…
Xbow 用 GPT-5.6 挖出 Chrome 三个漏洞，获 Google 25 万美元全链奖励 ↗
Xbow 工程师 Brendan Dolan-Gavitt 用 OpenAI 的 gpt-5.6-sol 分析 Chrome 源码，历时六天、花费约 4 万美元算力，找到三个独立漏洞组成漏洞链，使标签页内普通 JavaScript 代码穿越两个防护边界、逃出所有沙箱。
Google DeepMind and Singapore Government collaborate to build fine-tuned AI model for scam detection ↗
Learn how Google DeepMind and GovTech Singapore uses lightweight on-device Gemma models to intercept conversational scams in private chats with zero cloud data egress.
llm-gemini 0.35 ↗
Release: llm-gemini 0.35 I added support to llm-gemini for the new Gemini 4 Argon model, made available today via the Gemini API and for customers of their $100/month subscripti…
Hello Python 3.15, goodbye 3.10, and now I can use 3.11 ↗
Python 3.15 is here ! Here's the essential What's new in Python 3.15 guide, this year curated by 3.14 and 3.15 release manager Hugo van Kemenade. As someone who builds open sour…
What Should SFT Learn? Conditional Information Gain for Generalizable Post-Training ↗
Supervised fine-tuning (SFT) remains a widely used approach for post-training large language models, because fixed demonstrations can directly introduce desired reasoning strate…
VICAR: Video Imitation with Contact-Aware Retargeting for Humanoid Interactive Skills ↗
Third-person human videos offer a scalable alternative to task-specific reward engineering, teleoperation, and manually scripted demonstrations for humanoid loco-manipulation. R…
本周动态 / 来源信号
AI 编程
WorkBuddy Intl 更新支持 DeepSeek-V4.1-Flash-1017，10 月 31 日前免费 ↗
WorkBuddy Intl 已更新至最新的 DeepSeek-V4.1-Flash-1017 模型，10 月 31 日前免费开放。引用的 DeepSeek 官方内容称该模型相比上一版性能更强、任务完成更快、输出更短，已上线网站、App 和 API，并针对 DeepSeek Harness、Tencent CodeBuddy、WorkBuddy 等做了优化。
HyCo: Hybrid Drafting for Cost-Efficient Mobile-Cloud Speculative Decoding ↗
Device-cloud speculative decoding reduces server computation, but growing a deep draft tree with a small language model (SLM) requires repeated passes on a mobile neural process…
Benefit Street Partners answers questions faster with Codex ↗
Anant Kumar builds tools to compare loan valuations and analyze credit agreements, using Codex for code review and faster follow-ups.
Gemini 4 Argon 在 ARC-AGI (Verified) 评测中公布：ARC-AGI-3 得分 97.5% ↗
ARC Prize 公布 Gemini 4 Argon 在 ARC-AGI (Verified) 的评测成绩。ARC-AGI-3 上，provider adapter harness 得分 97.5%、成本 $4.8K，标准 harness 得分 49.8%、成本 $5.8K；ARC-AGI-2 得分 91.3%、$0.32/task；ARC-AGI-…
ANCHOR: Assessing Naturalistic Constraint Hierarchies and Ordering in Repository-Grounded Agentic Coding ↗
For repository-grounded coding agents, a functionally correct patch is not necessarily an acceptable one: developers also constrain how the task must be carried out-modification…
Visual Studio Copilot 与 SSMS 远程代码执行漏洞（CVE-2026-103572）技术分析 ↗
Johann Rehberger 披露 Visual Studio Copilot 与 SQL Server Management Studio 中的远程代码执行漏洞 CVE-2026-103572，根因是利用文件名末尾多加一个点绕过路径校验，写入 .mcp. 或 settings.VisualStudio. 等受保护配置文件，实现任意命令执行。
用 Claude Opus 5.5、Blender、Unreal Engine 和 Meshy 构建游戏的七步教程 ↗
一篇教程介绍用 Claude Opus 5.5 配合 Claude Code、Blender、Unreal Engine 5 和 Meshy 构建游戏的方法。核心思路是给每个工具分配单一职责：通过 MCP 连接 Blender 和 Meshy，由 Meshy 生成道具和角色、Blender 清理和放置、Unreal 负责光照物理和游戏本身、Opus 在…
本周动态 / 来源信号
智能体
Code, Test, and Deploy on One Machine with NVIDIA RTX Spark ↗
AI agents are changing how software gets built. Developers have gone from writing every line of code by hand to directing agents that write most of the code. As... AI agents are…
Auditing Escalation Reliability in Healthcare LLM Agents: A Perturbation-Controlled Benchmark of Workflow-State Failures ↗
Trustworthy biomedical AI requires reliable action selection when evidence is missing, stale, conflicting, or late, not only accurate answers on static inputs. We instantiate an…
Meta 开放权重模型 Muse Spark Open 上架 OpenRouter ↗
Meta 发布开放权重模型 Muse Spark Open，现已可在 OpenRouter 使用，定位为面向编码与智能体工作的美国最强开放权重模型。模型总参数 816B、激活 42B，支持 1M token 上下文窗口，原生多模态，OpenRouter 定价为输入 $1.25/M、输出 $4.25/M，正文提到支持图像、视频、音频和 PDF 输入。
Step 5 Preview 上线 FeatherlessAI，官方转发开放权重模型部署信息 ↗
阶跃星辰转发 FeatherlessAI 的消息，Step 5 Preview 已上线该平台，以 FP8 精度提供服务。该开放权重模型为 600B 稀疏 MoE 架构、每 token 激活 27B，面向长程智能体工作，在软件工程和专业知识工作（尤其金融）上有前沿级表现，按 Artificial Analysis 数据其每任务成本约为 Claude Op…
Optimizing Agent Reliability and Tool Calling Efficiency for Chip Design Using NVIDIA NeMo Gym ↗
AI agents offer remarkable productivity gains in complex chip design workflows, but their tendency to generate plausible yet incorrect artifacts makes... AI agents offer remarka…
Amp puts AI agents to work beyond code with OpenAI models ↗
See how Amp pushes the frontier of what agents can do with GPT-6 Astra, Ultrafast, GPT-Live-1, and Sign in with ChatGPT.
How to Rewrite Apache Spark UDFs for GPU Acceleration with AI Agents ↗
Enterprise applications have become increasingly demanding. Agentic AI has compounded the need for vast amounts of enterprise data with frequent, complex... Enterprise applicati…
LangChain 发布 langchain-scaledown 集成，用 ScaleDown 约束解码模型管理长时智能体上下文 ↗
LangChain 发布 langchain-scaledown 集成，并在 LangSmith LLM Gateway 中支持 ScaleDown 的约束解码模型（CDM），用于降低长时智能体上下文管理的成本和延迟。
Liquid AI 发布 LFM2.5-24B-A2B 端侧推理模型 ↗
Liquid AI 发布开源权重模型 LFM2.5-24B-A2B，总参数 24B、每 token 激活 2B，定位端侧快速智能体。
Let Large Language Models Judge Each Other: Multi-Agent Peer-Reviewed Reasoning for Medical Question Answering ↗
Objective: To enhance the accuracy, interpretability, and robustness of large language models (LLMs) in medical question answering (MedQA). Method: We designed a multi-agent pee…
本周动态 / 来源信号
具身智能
De-Risking Robotic Welding for In-Space Metal Joining Through Physics-Grounded Digital Twins ↗
Growing interest in in-space servicing, manufacturing, and assembly has renewed attention on reliable metal joining for modular construction and on-orbit repair. Robotic welding…
Safe Reinforcement Learning with Receding-Horizon Safety Critic and Differentiable Action Shields ↗
Despite the Markov assumption, safety failures in robot reinforcement learning rarely occur instantaneously; they typically emerge over several steps as the policy enters unfami…
TrAct: Bridging Robot Control and Visual Prediction with Visual Tracks ↗
Robot actions are inherently embodiment-specific and only weakly aligned with image-space visual changes, limiting their effectiveness as conditioning signals for robot world mo…
Trust-Calibrated VLA Policy Post-Training with Evolving Imagination ↗
Vision-Language-Action (VLA) models are often pretrained via imitation learning, making reinforcement learning (RL) post-training important when deployment moves beyond demonstr…
DexStitch: Learning and Composing Dexterous Skills from Human Videos via Contact-Aware Robotic Inpainting ↗
A long-horizon manipulation task consists of primitive skills and the transitions connecting them. Learning skills that can be executed in sequence requires supervision for both…
SAP-CoPE: Social-Aware Planning using Cooperative Pose Estimation with Infrastructure Sensor Nodes ↗
Autonomous mobile robots must operate smoothly in human-populated indoor environments, where onboard-only sensing suffers from limited perception range and occlusions, and socia…
本周动态 / 来源信号
基础设施
vLLM 解析 Kimi K3 在 AMD Instinct MI355X 上的推理优化，单 GPU 吞吐超 NVIDIA HGX B300 24.4% ↗
AMD 与 Embedded LLM 在 vLLM 官方博客详解 Kimi K3 在八卡 AMD Instinct MI355X 上的推理优化。
vLLM 实现 Reflection Beam 501B MoE 模型 Day 0 支持 ↗
vLLM 宣布从 Day 0 支持 Reflection 的 Beam 模型，这是一个 501B 参数、每 token 激活 23B 的 MoE 模型，在 NVIDIA Blackwell 上完成优化。
微软发布自研芯片 Maia 300 并将大规模扩展 Maia 推理部署 ↗
微软宣布明年起将 Maia 芯片家族扩展到多吉瓦级部署，作为自家产品和服务的首选推理系统，并发布 Maia 300，每瓦 token 产出比主流 GPU 系统高 40%。Maia 200 已在美国、欧洲和澳大利亚部署，在 Teams 和 Copilot 场景中每瓦 token 产出提升最高 1.7x。
本周动态 / 来源信号
研究
AWS 提出 priming 方法，可将预训练 Transformer 改造成混合状态空间架构 ↗
AWS 团队发布论文《Priming: Hybrid state space models from pre-trained Transformers》，提出 priming 方法，复用预训练 Transformer 权重构建混合架构，将预训练成本降低逾 99%，全流程消耗不足源模型 0.5% 的 token 预算。
值得跟进
继续跟踪
- llm-echo 0.5 · 待补独立核验依据
- WorkBuddy Intl 更新支持 DeepSeek-V4.1-Flash-1017，10 月 31 日前免费 · 待补独立核验依据
- Anthropic 如何在 Claude 写下 Claude Code 约 90% 代码后重新设计代码评审 · 待补独立核验依据
- Anthropic 披露 Claude 编写了约 90% 的 Claude Code 代码 · 待补独立核验依据
- Code, Test, and Deploy on One Machine with NVIDIA RTX Spark · 待补独立核验依据
- Chronoberg: A 250-Year Diachronic Corpus with Temporal Affective Lexicons for Language Model Analysis · 待补独立核验依据
- Auditing Escalation Reliability in Healthcare LLM Agents: A Perturbation-Controlled Benchmark of Workflow-State Failures · 待补独立核验依据
- De-Risking Robotic Welding for In-Space Metal Joining Through Physics-Grounded Digital Twins · 待补独立核验依据
- HyCo: Hybrid Drafting for Cost-Efficient Mobile-Cloud Speculative Decoding · 待补独立核验依据
- Diffeomorphic Hidden-State Steering: Improving Imitation Learning Policy Robustness through Expert Hidden-State Geometry · 待补独立核验依据
- LIN-SLAM: Lightweight LiDAR--Inertial Neural SLAM for Real-Time Onboard Navigation · 待补独立核验依据
- vLLM 解析 Kimi K3 在 AMD Instinct MI355X 上的推理优化，单 GPU 吞吐超 NVIDIA HGX B300 24.4% · 待补独立核验依据
- vLLM 实现 Reflection Beam 501B MoE 模型 Day 0 支持 · 待补独立核验依据
- Benefit Street Partners answers questions faster with Codex · 待补独立核验依据
- Claude Opus 5.5 fast mode 结束研究预览并移除等待名单 · 待补独立核验依据
- Chirp Pong · 待补独立核验依据
- Meta 开放权重模型 Muse Spark Open 上架 OpenRouter · 待补独立核验依据
- Gemini 4 Argon 在 ARC-AGI (Verified) 评测中公布：ARC-AGI-3 得分 97.5% · 待补独立核验依据
- Microsoft AI 公布 MAI-Base-2 预训练细节，2.8T 参数 MoE 模型 FLOPs 效率达 DeepSeek-V4-Pro-Base 的 4.2 至 12.3 倍 · 待补独立核验依据
- Step 5 Preview 上线 FeatherlessAI，官方转发开放权重模型部署信息 · 待补独立核验依据
- 阶跃星辰开放 Step 5 Preview 权重 · 待补独立核验依据
- Safe Reinforcement Learning with Receding-Horizon Safety Critic and Differentiable Action Shields · 待补独立核验依据
- TrAct: Bridging Robot Control and Visual Prediction with Visual Tracks · 待补独立核验依据
- LintEval: Benchmarking LLMs on Linter-Guided File Refactoring · 待补独立核验依据
- PersEvalBench: Can LLM Evaluators Judge Contextually Appropriate Use of Persona Information? · 待补独立核验依据
- Claude Code 空闲压缩从摘要继续，可用 idleCompaction 关闭并自定义阈值 · 待补独立核验依据
- Tree Appreciator · 待补独立核验依据
- Optimizing Agent Reliability and Tool Calling Efficiency for Chip Design Using NVIDIA NeMo Gym · 待补独立核验依据
- Amp puts AI agents to work beyond code with OpenAI models · 待补独立核验依据
- Databricks 讲解如何在 AI Runtime 上将 Qwen 3.8 27B 微调时间缩短 6 倍 · 待补独立核验依据
- PatTransBench: an ipc-stratified, multi-dimensional benchmark for patent machine translation evaluation · 待补独立核验依据
- Trust-Calibrated VLA Policy Post-Training with Evolving Imagination · 待补独立核验依据
- DexStitch: Learning and Composing Dexterous Skills from Human Videos via Contact-Aware Robotic Inpainting · 待补独立核验依据
- ANCHOR: Assessing Naturalistic Constraint Hierarchies and Ordering in Repository-Grounded Agentic Coding · 待补独立核验依据
- Xbow 用 GPT-5.6 挖出 Chrome 三个漏洞，获 Google 25 万美元全链奖励 · 待补独立核验依据
- Google DeepMind and Singapore Government collaborate to build fine-tuned AI model for scam detection · 待补独立核验依据
- llm-gemini 0.35 · 待补独立核验依据
- Visual Studio Copilot 与 SSMS 远程代码执行漏洞（CVE-2026-103572）技术分析 · 待补独立核验依据
- How to Rewrite Apache Spark UDFs for GPU Acceleration with AI Agents · 待补独立核验依据
- LangChain 发布 langchain-scaledown 集成，用 ScaleDown 约束解码模型管理长时智能体上下文 · 待补独立核验依据
- Liquid AI 发布 LFM2.5-24B-A2B 端侧推理模型 · 待补独立核验依据
- Hello Python 3.15, goodbye 3.10, and now I can use 3.11 · 待补独立核验依据
- AWS 提出 priming 方法，可将预训练 Transformer 改造成混合状态空间架构 · 待补独立核验依据
- 微软发布自研芯片 Maia 300 并将大规模扩展 Maia 推理部署 · 待补独立核验依据
- 用 Claude Opus 5.5、Blender、Unreal Engine 和 Meshy 构建游戏的七步教程 · 待补独立核验依据
- What Should SFT Learn? Conditional Information Gain for Generalizable Post-Training · 待补独立核验依据
- SAP-CoPE: Social-Aware Planning using Cooperative Pose Estimation with Infrastructure Sensor Nodes · 待补独立核验依据
- VICAR: Video Imitation with Contact-Aware Retargeting for Humanoid Interactive Skills · 待补独立核验依据
- Let Large Language Models Judge Each Other: Multi-Agent Peer-Reviewed Reasoning for Medical Question Answering · 待补独立核验依据