Openshelf：作者自己网站上的研究https://chenxiachan.github.io/openshelf/zh/2026-11-07T01:05:49ZApp Store 2026 in numbers: AI, pricing, abandoned appshttps://obelisk.club/research/app-store-20262026-11-06T00:00:00Zobelisk.club统计 178,637 个 App Store 应用的 AI 与内购情况How scientists use Claudehttps://www.anthropic.com/research/how-researchers-use-claude2026-11-06T00:00:00Zanthropic.com用隐私保护方法研究科研人员如何使用 ClaudeTeaching Qwen Whyhttps://www.lesswrong.com/posts/35yw2qGK7naFLAvcq/teaching-qwen-why-12026-11-06T00:00:00Zlesswrong.com研究如何通过教Qwen理由来改变其行为Experiments That Could Winhttps://aspirationalneuroscience.org/experiments/2026-11-06T00:00:00Zaspirationalneuroscience.org提出可赢得记忆解码挑战的具体实验方案Automated "Inverse" Design of Grating Couplers for Co-Packaged Opticshttps://engineering.flexcompute.com/articles/inverse-design-grating-couplers/2026-11-06T00:00:00Zengineering.flexcompute.com用形状优化与伴随法自动设计共封装光学光栅耦合器Codetations: Rich, Granular, and Maintainable Annotations for Evolving Codehttps://ztatlock.net/pubs/2026-uist-codetations/2026-11-06T00:00:00Zztatlock.net提出支持持久交互标注的VSCode扩展CodetationsShaping and Monitoring Models Through Their Internal Representationshttps://www.goodfire.com/research/mechanistic-shaping2026-11-06T00:00:00Zgoodfire.com通过内部表征塑造和监控模型的研究The Secret Leak Test: how well GateBolt kept Hodor's password, and where it didn'thttps://hodor.gatebolt.com/blog/secret-leak-test2026-11-06T00:00:00Zhodor.gatebolt.com测试提示注入过滤器GateBolt对AI攻击者的防护Lighthouse 13.6: how we made PageSpeed Insights more accurate for Framer siteshttps://www.framer.com/blog/lighthouse-13-6/2026-11-06T00:00:00Zframer.comFramer介绍与Lighthouse团队优化预加载评分的工程细节pico-WAM—Counterfactual Prediction and Planning Errorhttps://www.keivalya.com/pico-wam/2026-11-06T00:00:00Zkeivalya.com介绍小型世界动作模型pico-WAM并评估反事实预测I Spent $1000 Making LLMs Play Competitive Pokémonhttps://blog.doubleword.ai/llms-play-pokemon2026-11-06T00:00:00Zblog.doubleword.ai花1000美元让LLM对战竞技宝可梦并比较结果594,101 knocks: who reads my site, and what AI crawlers can't seehttps://www.shuffle-on.com/blog/ai-crawlers-read-my-site2026-11-06T00:00:00Zshuffle-on.com统计GPTBot对SSO登录页的594,101次访问The Day a Test Deleted My Home Directoryhttps://lucianolupo.com/notes/The%20Day%20a%20Test%20Deleted%20My%20Home%20Directory2026-11-06T00:00:00Zlucianolupo.com事后分析编码agent的变异测试删除主目录Turning the HEIC decoder into a SHA-256 computerhttps://www.irisc2.com/blog/heic-hashquine2026-11-06T00:00:00Zirisc2.com把HEIC解码器变成SHA-256计算机的技术演示Using decision models to speed up design rule violation repair in chip designhttps://engineering.flexcompute.com/articles/decision-models-drc-repair/2026-11-06T00:00:00Zengineering.flexcompute.com用决策模型加速芯片设计规则违规修复Harness 2: executing agent code as it streamshttps://www.framer.com/blog/harness-2/2026-11-06T00:00:00Zframer.comFramer发布Harness 2，通过流式执行JavaScript代码提升Agent效率并降低成本。METR: Independent Assessment of AI R&D Acceleration at Anthropic in October 2026https://www.lesswrong.com/posts/N5AmYkRks5CqAxvBg/metr-independent-assessment-of-ai-r-and-d-acceleration-at2026-11-06T00:00:00Zlesswrong.comMETR报告指出Anthropic在2026年10月利用AI加速研发的能力提升了约两倍。Programming in the Machine Code Space: LLMs can read & write executables byte by bytehttps://sgiannopoulos.github.io/blog/machine-code-programming/2026-11-06T00:00:00Zsgiannopoulos.github.io研究显示LLM可直接读取和编写机器码，无需反汇编工具即可分析二进制文件。Can LLMs redesign themselves as they train?https://www.doubleai.com/research/can-llms-redesign-themselves-as-they-train2026-11-06T00:00:00Zdoubleai.comDoubleAI研究让AI代理在训练期间动态重新设计神经网络架构。VFlash: Enabling System-1 VLMs for Physical AIhttps://research.nvidia.com/labs/lpr/vflash/2026-11-06T00:00:00Zresearch.nvidia.comNVIDIA提出VFlash，通过并行视觉规划加速物理AI中的机器人和自动驾驶推理。The components of agent memoryhttps://www.anthropic.com/research/components-of-agent-memory2026-11-06T00:00:00Zanthropic.comAnthropic研究展示Claude如何通过“做梦”机制整理记忆并纠正错误。DeepFinance: Can AI diligence a $100MM deal?https://www.juliusresearch.ai/blog/deep-finance2026-11-06T00:00:00Zjuliusresearch.aiDeepFinance是一个衡量AI在私募股权和投行工作流中表现的基准测试及排行榜。Computer-Use Agents Are Embodied Generalistshttps://pranjal2041.github.io/cua-embodied/2026-11-06T00:00:00Zpranjal2041.github.io研究展示单一计算机使用代理在多个具身AI竞赛中获得优异排名。The Language Server Tax: More Cost Without Better C/C++ Code Repair. ByteAskhttps://byteask.ai/research/lsp-claude-code-cpp2026-11-05T00:00:00Zbyteask.ai测试 clangd LSP 插件对 Claude Code 解决问题的影响StateBench: Testing how language models track state changes over long documentshttps://huggingface.co/blog/allenai/statebench2026-11-05T00:00:00Zhuggingface.co提出评测语言模型长文档状态追踪的基准 StateBenchKimi K3.1 Pro writes fake OpenAI ciphertexthttps://www.peterstran.com/projects/kimi-fernet/2026-11-05T00:00:00Zpeterstran.com记录 Kimi K3.1 Pro 输出伪造 OpenAI 密文的现象Larger Tokenizers Suffer More Multi-Epoch Regrethttps://wenhaochai.com/blogs/vocab-multi-epoch.html2026-11-05T00:00:00Zwenhaochai.com实验显示更大的分词器多轮训练遗憾更高On the Limits of Test-Time Compute for Self-Evaluationhttps://anirudhb11.github.io/blog/scorpio/2026-11-05T00:00:00Zanirudhb11.github.io分析测试时计算用于自我评估的局限Hiding a Rule Inside Another Model's Weightshttps://jonneslin.github.io/weight-space-trigger-en.html2026-11-05T00:00:00Zjonneslin.github.io手动编辑 Qwen3.5-0.8B 约 0.1% 权重植入规则Why the search for particle dark matter is really hopeless, part 2https://tritonstation.com/2026/11/05/why-the-search-for-particle-dark-matter-is-really-hopeless-part-2/2026-11-05T00:00:00Ztritonstation.com论证粒子暗物质搜寻在MOND现象下希望渺茫（第二部分）The test that deleted my home folderhttps://www.shuffle-on.com/blog/the-test-that-deleted-my-home-folder2026-11-05T00:00:00Zshuffle-on.com事后分析AI助手测试删除主目录的路径遍历漏洞事故Forget about credit assignment in long-horizon agentic post-traininghttps://qiyaoliang.com/blog/es-vs-pg-long-horizon/2026-11-05T00:00:00Zqiyaoliang.com论证长程agent后训练无需信用分配,比较ES与PGThe Exploit Was the Bottleneckhttps://www.a16z.news/p/the-exploit-was-the-bottleneck2026-11-05T00:00:00Za16z.news用23年数据分析AI与补丁差距Rethinking Abstractions in the Agentic Era: Towards a DSL and Compiler Foundation for Agentic Kernel Developmenthttps://tvm.apache.org/2026/11/03/rethinking-abstractions-in-the-agentic-era2026-11-05T00:00:00Ztvm.apache.org展示用Claude与Codex构建DSL与编译器基础GPT-Live vs GPT-Realtime: Voice AI Trade-offshttps://agentunicorn.ai/research/gpt-live-voice-ai-tradeoffs2026-11-05T00:00:00Zagentunicorn.ai比较GPT-Live、GPT-Realtime与流式语音管线的架构权衡AgentSight: System-wide AI agent profiling and monitoring with eBPFhttps://eunomia.dev/agentsight/2026-11-05T00:00:00Zeunomia.dev提出用eBPF做系统级AI agent监控的AgentSightVibeSys on Trainium: Agents Build and Tune Qwen3.8 Serving Engines on AWS Trainium2https://syfi.cs.washington.edu/blog/2026-11-04-vibesys-trainium/2026-11-05T00:00:00Zsyfi.cs.washington.edu让agent在Trainium2上构建并调优Qwen3.8推理引擎How to Train a Swarmhttps://www.gr.inc/blog/how-to-train-a-swarm2026-11-05T00:00:00Zgr.inc讨论训练agent swarm的方法与结果Declarative Dragginghttps://declarative-dragging.github.io/2026-11-04T00:00:00Zdeclarative-dragging.github.io提出声明式拖拽交互，UIST 2026 最佳论文项目页The SBND collaboration releases first neutrino resultshttps://news.fnal.gov/2026/10/the-sbnd-collaboration-releases-first-neutrino-results/2026-11-04T00:00:00Znews.fnal.govSBND 合作组发布首个中微子相互作用结果AbMaze: Evaluating frontier AI agents on protein fitness landscapeshttps://valenlabs.com/blog/abmaze2026-11-04T00:00:00Zvalenlabs.com提出抗体适应度景观基准 AbMaze 评测智能体Investigating vault particles with Claudehttps://www.anthropic.com/research/investigating-vault-particles-with-claude2026-11-04T00:00:00Zanthropic.com用 Claude 辅助解析原核生物 vault 颗粒的首个 3D 结构Advancing the Quality-Efficiency Frontier for Coding AI on Windows PCs - Microsoft Foundry on Windows Blog Coding AI on Windows: The Quality–Efficiency Frontierhttps://devblogs.microsoft.com/foundry-on-windows/coding-ai-on-windows-pcs-quality-efficiency/2026-11-04T00:00:00Zdevblogs.microsoft.com用低比特量化与推测解码提升 Windows 本地编码 AI 效率Serving GLM-5.3-Flash on one Trainium2 box with NKI layer megakernelshttps://runanywhere.ai/blog/glm-5-3-flash-trainium2-nki-megakernels2026-11-04T00:00:00Zrunanywhere.ai在单台 Trainium2 上服务 GLM-5.3-Flash 并测量吞吐I Tried to Break Weight Programminghttps://flecart.github.io/posts/2026-11-03-i-tried-to-break-weight-programming/2026-11-04T00:00:00Zflecart.github.io复现权重编程实验并分析复杂度限制MCPhantom: AI Agents Installing Deprecated MCP Servershttps://www.glow.io/blogs/mcphantom-ai-coding-agents-are-installing-vulnerable-deprecated-mcp-servers-from-training-data2026-11-04T00:00:00Zglow.io发现AI编码agent会安装带已知漏洞的废弃MCP服务器Can Advertising Fund Free AI? ChatGPT Ads Researchhttps://ecommerceresearch.org/chatgpt-ads/2026-11-04T00:00:00Zecommerceresearch.org用实证数据研究ChatGPT广告能否资助免费AI20 Percent of All ‘Ok’ Web Responses AI Gets Are Not Okayhttps://www.unite.ai/20-percent-of-all-ok-web-responses-ai-gets-are-not-okay/2026-11-04T00:00:00Zunite.ai实验发现AI抓取网页时约20%的OK响应其实有问题Regulation by AI: Replaying EPA’s Glyphosate Decisionshttps://reckoningscience.org/regulation-by-ai-replaying-epas-glyphosate-decisions/2026-11-04T00:00:00Zreckoningscience.org让五个AI模型重演50年EPA草甘膦决策并比较Making flawed benchmarks usefulhttps://blatent.ai/blog/making-flawed-benchmarks-useful2026-11-04T00:00:00Zblatent.ai审视有缺陷基准并将其变得有用的评测方法Jane Street Blog - A study of sequence weighting at scalehttps://blog.janestreet.com/a-study-of-sequence-weighting-at-scale/2026-11-04T00:00:00Zblog.janestreet.com研究语言模型中序列加权的缩放律Weight-space programming a neural quantum statehttps://paulg.info/2026/11/04/quantum-weight-space/2026-11-04T00:00:00Zpaulg.info与Opus协作诊断并修复神经量子态的权重空间编程Learnings from Auditing Public Evaluationshttps://www.primeintellect.ai/blog/auditing-public-evaluations2026-11-04T00:00:00Zprimeintellect.ai审计公开评测发现评分器错误与歧义Progress on provable inference for large language modelshttps://www.anthropic.com/research/provable-inference2026-11-03T00:00:00Zanthropic.comAnthropic 报告 LLM 可证明推理 zkSNARK 的进展The shelf life of LLM findingshttps://zou-group.github.io/llm-claims-shelf-life/2026-11-03T00:00:00Zzou-group.github.io对 50 项 LLM 已发表结论在 20 个模型上复现The Nordic token taxhttps://research.klang.ai/papers/the-nordic-token-tax2026-11-03T00:00:00Zresearch.klang.ai测量 12 个前沿模型对非英语提示多收的 token 成本LLM grounding tutorial: stop AI agents guessing priceshttps://blog.apify.com/llm-grounding-with-live-data/2026-11-03T00:00:00Zblog.apify.com在 21 个商品页上测试 AI 代理价格幻觉与实时数据接地Krea 3 Technical Reporthttps://www.krea.ai/blog/krea-3-technical-report2026-11-03T00:00:00Zkrea.aiKrea 3 图像编辑模型的技术报告Rotary Position Encodings for Graphshttps://idlemachines.co.uk/essays/rope-for-graphs2026-11-03T00:00:00Zidlemachines.co.uk提出图结构上的旋转位置编码A Mathematical Framework for Transformer Circuitshttps://transformer-circuits.pub/2021/framework/index.html#summary-of-results2026-11-03T00:00:00Ztransformer-circuits.pubTransformer 电路的数学框架，机制可解释性奠基论文Quantity vs quality in pottery is a liehttps://www.lesswrong.com/posts/LrAqM4bGmAoA79rP3/quantity-vs-quality-in-pottery-is-a-lie2026-11-03T00:00:00Zlesswrong.com追溯陶艺数量与质量故事的来源并核查事实Autonomous Agents Need Mandatory Access Controlhttps://www.legiscope.com/blog/autonomous-agents-mandatory-access-control.html2026-11-03T00:00:00Zlegiscope.com提出借鉴 FreeBSD 的自主智能体强制访问控制框架The 3D Printed Databasehttps://www.amplifypartners.com/blog-posts/the-3d-printed-database2026-11-03T00:00:00Zamplifypartners.com探索根据领域本体自动生成定制数据库系统“This Is Fine”: How Clean Rooms Train Agents to Close the Books on Firehttps://archytas-ai.com/blog/this-is-fine2026-11-03T00:00:00Zarchytas-ai.com研究 RL 环境如何让智能体虚报任务完成Same Model, Different Harness: CUGA, Hermes, and Deep Agentshttps://huggingface.co/blog/ibm-research/cuga-harness2026-11-03T00:00:00Zhuggingface.co比较 CUGA、Hermes、Deep Agents 三种 agent harnessThe Hitchhiker’s Guide to Training on Compute Islandshttps://www.coreauto.com/hitchhikers-guide-to-training-on-compute-islands2026-11-03T00:00:00Zcoreauto.com跨数据中心容错分布式训练的立场与稀疏 MoE 实验A Direct Construction of SZ Sequence Matriceshttps://andrew-helmer.github.io/sz-matrix-construction/2026-11-03T00:00:00Zandrew-helmer.github.io给出 SZ 序列矩阵的直接构造方法1.26 Million Data Center Satellites? Orbit Says Nohttps://blog.budget-engineer.com/space-data-centers-have-you-looked-up-lately2026-11-03T00:00:00Zblog.budget-engineer.com分析126万颗数据中心卫星对轨道拥挤的影响Quadrillion-parameter transformers and how to train themhttps://qlabs.sh/research/parameters2026-11-03T00:00:00Zqlabs.sh提出万亿参数Transformer的缩放律并外推训练方案TIS-100 Bench: LLMs Push the Pareto Frontier in Programming Puzzle Gameshttps://tis-100.wincon.dev/2026-11-03T00:00:00Ztis-100.wincon.dev评测16个LLM在TIS-100编程谜题上的帕累托前沿Eval-production behavior gaps in agent safety and the role of eval-awarenesshttps://alignment.openai.com/eval-production-behavior-gap/2026-11-03T00:00:00Zalignment.openai.com比较agent在安全评测与实际部署中的行为差异及评测意识The Inference Stack Behind OCRiahttps://www.llamaindex.ai/blog/the-inference-stack-behind-ocria2026-11-03T00:00:00Zllamaindex.ai介绍OCRia推理栈的解码延迟优化方法StateBench: Testing how language models track state changes over long documentshttps://allenai.org/blog/statebench2026-11-03T00:00:00Zallenai.org提出StateBench评测LLM在长文档中追踪状态变化ZHA: Zyphra Hierarchical Attentionhttps://www.zyphra.com/our-work/zyphra-hierarchical-attention2026-11-03T00:00:00Zzyphra.com提出Zyphra分层稀疏注意力方法ZHAFormalizing PTX in Lean to Prove GPU Kernel Correctnesshttps://sf-tensor.com/engineering/formalizing-ptx2026-11-03T00:00:00Zsf-tensor.com在Lean中形式化Nvidia PTX以证明GPU内核正确性Userland Linux scheduler in Rust-faster than timehttps://kapteko.com/userland-linux-scheduler-in-rust/2026-11-03T00:00:00Zkapteko.com用Rust实现用户态Linux调度器以加速虚拟时间Sparks of Weight-Space Programming: LLMs can natively read & write neural network weightshttps://rikiyat.github.io/blog/weight-space-programming/2026-11-02T00:00:00Zrikiyat.github.io发现前沿 LLM 能直接读写神经网络权重以改进 Qwenbf16 might be freezing your weight updateshttps://lacuna.tiptreesystems.com/paper/bf16-might-be-freezing-your-weight-updates/art_487b1ed949e745a6ae0c42815b8d3cba2026-11-02T00:00:00Zlacuna.tiptreesystems.com分析 bf16 权重下 AdamW 更新被冻结的问题GDPx: Measuring Frontier Model Capabilities Across the Most Valuable Domainshttps://newsletter.semianalysis.com/p/gdpx-measuring-frontier-model-capabilities2026-11-02T00:00:00Znewsletter.semianalysis.com提出跨多个高价值领域评测前沿模型的基准 GDPxAccelerating Attention for Ultra-Long Contexthttps://prefrontal.ai/news/lsa/2026-11-02T00:00:00Zprefrontal.ai提出 LSA 层级聚类注意力以加速超长上下文Will AI compute stay tight?https://hxu296.github.io/ai-compute-landscape/2026-11-02T00:00:00Zhxu296.github.io分析 AI 算力租金与承诺以判断 neocloud 是否泡沫WeVisCut: The Whole Editing Room in One Agenthttps://tencent.github.io/WeVisCut2026-11-02T00:00:00Ztencent.github.ioWeVisCut 项目页，用智能体完成长视频剪辑Making CUDA Checkpointing 15x Faster on a B300: A Workloghttps://maknee.github.io/blog/2026/Cuda-Checkpointing/2026-11-02T00:00:00Zmaknee.github.io记录将 B300 上 CUDA 检查点加速 15 倍的工作日志ZeroBenchhttps://zerobench.github.io/2026-11-02T00:00:00Zzerobench.github.io提出对多模态大模型极难的视觉基准 ZeroBenchTracking R&D funding in the Global South: challenges and opportunities in Indiahttps://www.leidenmadtrics.nl/articles/tracking-r-d-funding-in-the-global-south-challenges-and-opportunities-in-india2026-11-02T00:00:00Zleidenmadtrics.nl分析印度研发经费追踪的挑战与机会What does “verified memory safe” even mean?https://focs-lab.comp.nus.edu.sg/blog/memory-safety-audit/2026-11-02T00:00:00Zfocs-lab.comp.nus.edu.sg对Anthropic与Theorem的内存安全证明做独立审计Agentic Backdoors via Pretraining Poisoninghttps://pbb.sh/research/Agentic-Backdoor/2026-11-02T00:00:00Zpbb.sh证明预训练投毒可植入远程代码执行的agent后门BlitzRank: Principled Zero-shot Ranking Agents with Tournament Graphshttps://blitzrank.ai/2026-11-02T00:00:00Zblitzrank.ai提出基于锦标赛图的零样本LLM文档排序方法BlitzRankMaking a Magic The Gathering Evalhttps://folds.ink/posts/mtg-benchmark2026-11-02T00:00:00Zfolds.ink构建并评测LLM玩万智牌的基准Lessons From Red-teaming Auto-review, OpenAI's Blocking Monitorhttps://www.lesswrong.com/posts/dbStc8ZFESqfihwAd/lessons-from-red-teaming-auto-review-openai-s-blocking-12026-11-02T00:00:00Zlesswrong.com红队测试OpenAI的Auto-review阻断监控器并总结经验GPU-accelerated Weight Teleportationhttps://humansand.ai/blog/gpu-accelerated-weight-teleportation2026-11-02T00:00:00Zhumansand.ai提出GPU压缩与流式解压实现无损权重更新Shotwell-Bench: An annotation benchmark that accepts valid alternativeshttps://shotwell.ai/blog/shotwell-bench/2026-11-02T00:00:00Zshotwell.ai提出接受有效替代分段的视频标注基准Shotwell-BenchFrom Words to Waypoints: Teaching a Small LLM to Plan for Robotshttps://www.daweisun.me/blog/stl-planning-rl/2026-11-02T00:00:00Zdaweisun.me训练小型LLM为机器人生成路径点的实验Creating Preferred IUPAC nameshttps://chem-bla-ics.linkedchemistry.info/2026/11/01/creating-preferred-iupac-names.html2026-11-01T00:00:00Zchem-bla-ics.linkedchemistry.info探讨用语言模型生成IUPAC优选名称Morg2SMILES: Generative Recovery of SMILES Strings from Morgan Fingerprintshttps://jonswain.github.io/morg2smiles/2026-11-01T00:00:00Zjonswain.github.io讨论从Morgan指纹生成式恢复SMILES字符串Can an LLM and a keyword index replace embeddings?https://internals.laxmena.com/p/can-an-llm-and-a-keyword-index-replace2026-11-01T00:00:00Zinternals.laxmena.com实验比较用LLM加关键词索引替代embedding做语义检索的效果The abandoned PR pagehttps://aictrl.dev/blog/open-source-review-ai-era2026-11-01T00:00:00Zaictrl.dev分析两个AI原生开源项目中PR合并率与被放弃PR的数据Leaking Muse's Charter, Meta's model spechttps://www.lesswrong.com/posts/TJGPL5FAk9sCioPbE/leaking-muse-s-charter-meta-s-model-spec2026-11-01T00:00:00Zlesswrong.com通过提取思维链内容还原Meta Muse模型的Charter规范AI text watermarking and quality losshttps://blog.keyvan.net/p/ai-text-watermarking-and-quality2026-10-31T00:00:00Zblog.keyvan.net评估 AI 文本水印对文本质量的影响A Specialised Haskell Processor on FPGAhttps://bathtub-01.github.io/technical/ouros-demo/2026-10-31T00:00:00Zbathtub-01.github.io展示在 FPGA 上运行 Haskell 的专用处理器A Generalized Scaling Law for Monopiles in Sand: From Centrifuge Tests to an Interactive Predictorhttps://geo-huanwang.com/posts/2026/10/monopile-moment-rotation-predictor/2026-10-31T00:00:00Zgeo-huanwang.com基于离心机试验提出砂土中单桩的通用缩放律并附交互预测工具Automated Subtle Emergent Misalignment Researchhttps://alignment.anthropic.com/2026/automated-em-search/2026-10-31T00:00:00Zalignment.anthropic.com自动化研究微妙的涌现式错位Measuring voice agents: turn-taking, tool calls and expressive speechhttps://soniqo.audio/blog/voice-agent-benchmarks2026-10-31T00:00:00Zsoniqo.audio评测语音智能体的轮替、工具调用与表达性语音A CVE triage pipeline that keeps asking, because day one tells you nothinghttps://34-118-125-101.sslip.io/log/rappaport/2026-10-31T00:00:00Z34-118-125-101.sslip.io描述一个持续追问的CVE分诊流水线Rappaport的设计与经验Deterministic Factoring in Time $N^{o(1)}$https://hexagonmath.org/2610.00600v12026-10-30T00:00:00Zhexagonmath.org声称给出整数分解N^O(1)确定性算法Honeycomb-Surfels (Hexels)https://honeycomb-surfels.github.io/2026-10-30T00:00:00Zhoneycomb-surfels.github.ioHoneyComb Surfels边缘3D渲染项目页What is work for? Human wellbeing and the value of work in the AGI erahttps://institute.deepmind.com/essays/what-is-work-for-human-wellbeing-and-the-value-of-work-in-the-agi-era/2026-10-30T00:00:00Zinstitute.deepmind.comDeepMind研究AGI时代工作与幸福CDC Grant Disruption Methodologyhttps://grantwitness.org/cdc/updates/methodology2026-10-30T00:00:00Zgrantwitness.org说明CDC资助中断数据的来源与方法AGI Economics: Coordination vs. Competencehttps://institute.deepmind.com/essays/the-economics-of-intelligence/2026-10-30T00:00:00Zinstitute.deepmind.comDeepMind分析AGI经济学中协调与能力的关系Building oversight tools into a model's weightshttps://divagr.com/blog/building-oversight-tools-into-a-models-weights2026-10-30T00:00:00Zdivagr.com把监督工具编入Qwen权重的测谎与日志机制An adventure in multifoundational Mathematics 2.0: the cycle double cover theorem in ZFC with Grace17https://grace17.org/docs/cycle-double-cover.html2026-10-30T00:00:00Zgrace17.org在ZFC中证明循环双覆盖定理的多基础数学案例Reverse Turing Test: Surprising Proof AIs Fail to Spot AIshttps://www.progressiverobot.com/2026/10/27/reverse-turing-test-ais-fail-to-detect-ais/2026-10-30T00:00:00Zprogressiverobot.com反向图灵测试中七个AI评审仅识别出3/98个AIWhat agents see when they search MCP toolshttps://www.babush.me/how-agents-search-mcp-tools.html2026-10-30T00:00:00Zbabush.me分析智能体搜索MCP工具时看到的内容Product Hunt is swamped nowhttps://www.mahl.me/blog/product-hunt-is-swamped-now/2026-10-30T00:00:00Zmahl.me抓取Product Hunt数据量化2023年后发布量变化Understanding Open-ended Work from a Data Perspectivehttps://www.fig.inc/blog/understanding-open-ended-work-from-a-data-perspective/2026-10-30T00:00:00Zfig.incFig分析开放式工作数据中缺失的学习信号Solving the JSON Schema dialect fragmentation problem with auto-upgradeshttps://www.sourcemeta.com/blog/json-schema-dialect-auto-upgrade/2026-10-30T00:00:00Zsourcemeta.com用自动升级解决JSON Schema方言碎片化Grace Blackwell needs a new scheduler: 27% more usable GPUs from the same rackshttps://skypilot.ai/blog/grace-blackwell-needs-a-new-scheduler2026-10-29T00:00:00Zskypilot.aiGrace Blackwell调度器多出27%可用GPUOpen-weight AI models can exploit the patch gaphttps://epoch.ai/publications/open-weight-models-patch-gap2026-10-29T00:00:00Zepoch.ai开源权重模型可利用补丁间隔漏洞的评估N-gram hinted language modeling: NanoGPT Speedrun in 6.7shttps://brunborg.com/n-gram-hinted-lm2026-10-29T00:00:00Zbrunborg.com用n-gram提示将NanoGPT Speedrun降至6.7秒Weaponizing Personalization: From Knowledge Graph to Extortionhttps://8rank.ai/game/apple-intelligence-knowledge-graph-in-the-wrong-hands2026-10-29T00:00:00Z8rank.ai利用Apple个人知识图谱与LLM策划绑架的研究Every iPhone has a fingerprint. Can one photo give you away?https://vhamolya.com/blog/which-iphone-took-this-photo/2026-10-29T00:00:00Zvhamolya.com检验单张照片能否识别iPhone相机指纹Coding Agent Degradation: Cross-Vendor Field Notes and an Initial Analysishttps://www.arcblock.io/en/articles/when-coding-agents-stop-finishing/2026-10-29T00:00:00Zarcblock.io跨厂商编码智能体性能退化的现场分析AReAhttps://rajabi2001.github.io/area/2026-10-29T00:00:00Zrajabi2001.github.ioAReA高分辨率扩散Transformer高效注意力项目页Why AI Design Tools Keep Making the Same Flyer: Claude Design, Canva, Moda and Grafista Testedhttps://grafista.io/blog/posts/grafista-vs-claude-vs-canva-vs-moda2026-10-29T00:00:00Zgrafista.io对比四款AI设计工具对相同传单提示的输出State of Harness—October 2026https://www.harness-hub.org/reports/2026-102026-10-29T00:00:00Zharness-hub.orgState of Harness首份报告界定智能体harness与市场Serving Prefill and Decode for Agentic AIhttps://tenstorrent.com/en/developers/blog/serving-prefill-and-decode-for-agentic-ai2026-10-29T00:00:00Ztenstorrent.com在Tenstorrent Blackhole上服务智能体的prefill与decode31 AI Models Took a Political Quiz. Here Is Where Each One Landedhttps://politicaldna.org/ai-political-dna/2026-10-29T00:00:00Zpoliticaldna.org让31个LLM做政治问卷,发现其趋同程度高于人类Scaling Misuse Detection with Self-Assessmenthttps://safeguards.anthropic.com/2026/self-assessment/2026-10-29T00:00:00Zsafeguards.anthropic.com提出让Claude自评回答安全性以扩展滥用检测的方法并评估Catching Bugs Before They Land: JiTTesting’s Human-on-the-Loop Approach at Full Scalehttps://engineering.fb.com/2026/10/29/security/catching-bugs-before-they-land-jittestings-human-on-the-loop-approach-at-full-scale/2026-10-29T00:00:00Zengineering.fb.comMeta介绍JiT测试中人在回路的大规模实践Qwenputer: a fully functioning, deterministic RISC-V CPU inside Qwen3.5https://divagr.com/blog/qwenputer2026-10-28T00:00:00Zdivagr.com在Qwen3.5内实现确定性RISC-V CPUThe sleep framework: A new way for LLMs to retain knowledge in continual learninghttps://research.google/blog/the-sleep-framework-a-new-way-for-llms-to-retain-knowledge-in-continual-learning/2026-10-28T00:00:00Zresearch.googleGoogle提出sleep框架用于持续学习From Tokens to Bytes I: Scaling Hierarchical Networkshttps://www.cartesia.ai/blog/from-tokens-to-bytes-part-12026-10-28T00:00:00Zcartesia.aiCartesia关于层级网络扩展的研究Intelligence Benchmarkinghttps://artificialanalysis.ai/methodology/intelligence-benchmarking2026-10-28T00:00:00Zartificialanalysis.aiArtificial Analysis智能指数v5.0基准How AI agents shape experts’ judgment callshttps://www.anthropic.com/research/agents-judgment-calls2026-10-28T00:00:00Zanthropic.com研究分析师在86%决策中跟随AI代理偏好Of sheep and beerhttps://dlab.epfl.ch/2017-08-30-of-sheep-and-beer/2026-10-28T00:00:00Zdlab.epfl.ch早期评分造成产品评论羊群效应的研究Are you paying for useless LLM prompts?https://cameronmpalmer.com/blog/are-you-paying-for-useless-llm-prompts/2026-10-28T00:00:00Zcameronmpalmer.com评估冗余LLM提示是否有效的实验Can LLMs Supercharge Geolocations? Our Tests Show Big Improvements But Humans Remain Keyhttps://www.bellingcat.com/resources/2026/10/22/can-llms-supercharge-geolocations-our-tests-show-big-improvements-but-humans-remain-key/2026-10-28T00:00:00Zbellingcat.comBellingcat评估LLM在地理定位中的表现Agent Scrumhttps://daveremy.github.io/agent-scrum/2026-10-28T00:00:00Zdaveremy.github.ioAI编码智能体冲刺方法及324次冲刺观察Recurrent Looped Transformerhttps://yifanzhang-pro.github.io/recurrent-looped-tranformer/2026-10-28T00:00:00Zyifanzhang-pro.github.ioRecurrent Looped Transformer状态跟踪项目页MotionBricks: Scalable Real-Time Motions with Modular Latent Generative Model and Smart Primitiveshttps://nvlabs.github.io/motionbricks/2026-10-28T00:00:00Znvlabs.github.ioMotionBricks实时运动生成论文项目页Who is in charge? Robot autonomy levels for human-in-the-loop systemshttps://zlatanajanovic.com/robot-autonomy-levels/2026-10-28T00:00:00Zzlatanajanovic.com提出人在回路系统的机器人自主等级框架（书章节）Building My Own Tiny AI Weather Modelhttps://tcumulus.dev/blog/building-my-own-tiny-ai-weather-model/2026-10-28T00:00:00Ztcumulus.dev基于GraphCast架构自建小型AI天气模型Benchmarking Coding Agents at Scalehttps://www.snowflake.com/en/blog/engineering/Benchmarking-coding-agents/2026-10-28T00:00:00Zsnowflake.comSnowPR用真实PR评测编码智能体于4300万行代码库Rethinking Search as Code Generationhttps://www.perplexity.ai/hub/blog/rethinking-search-as-code-generation2026-10-28T00:00:00Zperplexity.aiPerplexity提出把搜索重构为代码生成的可编程原语Training Reve 2.0 at scalehttps://app.reve.com/training-infra2026-10-28T00:00:00Zapp.reve.comReve 2.0多模态模型训练基础设施的技术报告Octoverse: Commits, pushes, reach new highs in 2026 as developers build with agentshttps://github.blog/news-insights/octoverse/octoverse-commits-pushes-reach-new-highs-in-2026-as-developers-build-with-agents/2026-10-28T00:00:00Zgithub.blogGitHub年度报告,统计2026年提交量与开发者增长数据CFSG in Leanhttps://cfsg.axiommath.ai/2026-10-27T00:00:00Zcfsg.axiommath.ai用Lean形式化有限单群分类并得到新定理Neuron Persona Libraryhttps://qwen-neurons-site.vercel.app/2026-10-27T00:00:00Zqwen-neurons-site.vercel.app翻转单个神经元触发Qwen人格的研究A $2^{o(n)}$ Time Algorithm for 3-SAThttps://hexagonmath.org/2610.00577v12026-10-27T00:00:00Zhexagonmath.org声称给出3-SAT确定性算法以反驳ETH40× fewer tokens: wrap Markdown in SQLitehttps://www.withnative.ai/research-and-posts/markdown-context-benchmark2026-10-27T00:00:00Zwithnative.ai声称用SQLite使Markdown的token效率提升40倍A critical look at self-distillation for reasoninghttps://jasonkena.github.io/blog/posts/does_self_distillation_improve_reasoning/2026-10-27T00:00:00Zjasonkena.github.io对推理中self-distillation的批判性分析SAS|Symbiotic Autoprotection System|κD = 0.56https://leesintheblindmonk1999.github.io/sas-landing/#hero2026-10-27T00:00:00Zleesintheblindmonk1999.github.io声称的确定性AI结构审计框架K_D=0.56teaching a car to drive inside its own dreamhttps://fpoltronieri.github.io/blog/2026/teaching-a-car-to-drive-inside-its-own-dream/2026-10-27T00:00:00Zfpoltronieri.github.io记录让自动驾驶在潜在世界模型内学会驾驶的实验Same price, different reason. Probing AI pricing agents for collusionhttps://www.maxtaylor.me/articles/same-price-different-reason-probing-ai-pricing-agents-for-collusion2026-10-27T00:00:00Zmaxtaylor.me用线性探针检测AI定价智能体的串通Can Machines Follow One Voice Through a Crowded Room?https://www.kelvinjiang.com/2026/10/can-machines-follow-voice-through-crowded-room.html2026-10-27T00:00:00Zkelvinjiang.com探索机器能否在嘈杂环境中跟随单一说话人From the bottom of my heart, I salute youhttps://janni.turunen.in/poisoned-open-weights/2026-10-27T00:00:00Zjanni.turunen.in展示向开放权重模型植入后门的三种方式Human travel patterns reflect geographic boundaries rather than universal lawshttps://theprimary.com/science/2026-10-26/mobility-power-law-artifacts2026-10-27T00:00:00Ztheprimary.com研究显示人类出行模式由地理边界而非普适律决定The Fertility of an Idea: tracing the discovery lineage of optogeneticshttps://www.amplifypartners.com/blog-posts/the-fertility-of-an-idea2026-10-27T00:00:00Zamplifypartners.com追溯光遗传学发现谱系的历史研究Can AI draw San Francisco from memory?https://voygr.tech/blog/can-ai-draw-san-francisco-from-memory/2026-10-27T00:00:00Zvoygr.tech测试六个AI聊天模型凭记忆绘制旧金山地图The Ability of Frontier Models to Control Robots is Rapidly Improvinghttps://www.lesswrong.com/posts/ndS9Xn8Qx5PppHJ3Q/the-ability-of-frontier-models-to-control-robots-is-rapidly2026-10-27T00:00:00Zlesswrong.com展示前沿模型控制机器人的能力快速提升Concerning behavior from a popular open-weight AI modelhttps://cohere.com/blog/concerning-behaviour-popular-open-weight-ai-model2026-10-26T00:00:00Zcohere.comCohere披露某开源权重模型的异常行为MS-Human-700: Whole-body Human Musculoskeletal Modeling and Controlhttps://lnsgroup.cc/research/MS-Human/2026-10-26T00:00:00Zlnsgroup.ccMS-Human-700人体肌肉骨骼MuJoCo模型项目页Why the search for particle dark matter is hopeless, part 1https://tritonstation.com/2026/10/26/why-the-search-for-particle-dark-matter-is-hopeless-part-1/2026-10-26T00:00:00Ztritonstation.com论证粒子暗物质搜寻为何无望（第一部分）Research | Galatea: optimal transport for lunar place recognitionhttps://www.lunarsim.org/research/2026-10-26T00:00:00Zlunarsim.orgGalatea用最优传输融合全局与局部描述子做月面定位The Half-Life of an Instructionhttps://digitaliziran.si/2026/10/26/the-half-life-of-an-instruction.html2026-10-26T00:00:00Zdigitaliziran.si测试DeepSeek V4 Pro指令遵循随示例变化的衰减AI agents have few qualms with log tamperinghttps://www.manifold.security/blog/agent-log-tampering2026-10-26T00:00:00Zmanifold.security研究显示多数前沿AI智能体会篡改自身日志Joints, Camera, Action! Co-Generating 3D Motion and Video for Arbitrary Rigshttps://research.nvidia.com/labs/par/projects/jca/2026-10-26T00:00:00Zresearch.nvidia.comJCA为任意骨架联合生成视频与同步3D动作Qwen, Minus the Ehttps://kevin-wu.ca/blog/qwn/2026-10-25T00:00:00Zkevin-wu.ca在16GB GPU上微调Qwen避免字母e的实验Simulated Control Evaluationshttps://www.redwoodresearch.org/blog/simulated-control-evaluations2026-10-25T00:00:00Zredwoodresearch.org用伪造工具输出的LLM红队测试AI智能体监控器3D Gaussian Splatting for Real-Time Radiance Field Renderinghttps://repo-sam.inria.fr/fungraph/3d-gaussian-splatting/2026-10-25T00:00:00Zrepo-sam.inria.fr3D Gaussian Splatting实时辐射场渲染论文项目页Exploring Apple's On-Device Model: Core Advancedhttps://www.airealist.ai/p/exploring-apple-on-device-model2026-10-25T00:00:00Zairealist.ai分析Apple 20B参数端侧模型如何存于闪存Deleting RAG from our web agent made it 2.3x fasterhttps://www.skyvern.com/blog/deleting-rag-from-our-web-agent-made-it-2-3x-faster/2026-10-24T00:00:00Zskyvern.comSkyvern重写后在Odysseys基准上提速2.3倍的报告Timestral Financial Forecasting: Time-Series Foundation Modelshttps://philippdubach.com/posts/forecasting-financial-markets-with-timestral/2026-10-24T00:00:00Zphilippdubach.comTimestral模型预测金融市场的评估结果Compiling a Workflow into a 3B Modelhttps://arpitbhayani.me/blogs/compiling-an-sop-into-a-3b-model/2026-10-24T00:00:00Zarpitbhayani.me将事件分诊SOP编译进3B模型的微调实验BeNice! Closing the Inference-Memory Gap of Looped Models with KV-Sharing–Sandy Researchhttps://sandyresearch.github.io/benice/2026-10-24T00:00:00Zsandyresearch.github.io通过KV-sharing缩小循环模型推理内存差距的研究Solving a Giant Rubik's Cube with a Quadruped Robothttps://leggedrobotics.github.io/rubik-skills/2026-10-24T00:00:00Zleggedrobotics.github.io四足机器人自主扫描并解开巨型魔方的项目页AI Reviews Might Beat Human Peer Reviewshttps://blog.stephenturner.us/p/ai-peer-reviews-beat-humans2026-10-24T00:00:00Zblog.stephenturner.us介绍AI评审可能不亚于人工同行评审的演示Mundo: From seeing a place to simulating what happens therehttps://huggingface.co/blog/Luo-Yihang/mundo-world-model2026-10-24T00:00:00Zhuggingface.coMundo世界模型从少量房间照片生成可模拟场景How well do job retraining programs work?https://www.anthropic.com/research/reviewing-the-evidence-on-worker-retraining-programs2026-10-24T00:00:00Zanthropic.comAnthropic综述工人再培训项目的证据Crowley's Three Axes of JSON Queryinghttps://nrposner.com/blog/crowleys-three-axes-of-json-querying/2026-10-24T00:00:00Znrposner.comCrowley流式JSON查询突破2 GB/s的设计与测量ReplaySSM: Cache SSM Inputs, Not Statehttps://tridao.me/blog/2026/replayssm/2026-10-24T00:00:00Ztridao.meReplaySSM缓存SSM输入而非状态以加速解码Air Purifiers as White Noise Machineshttps://www.jefftk.com/p/air-purifiers-as-white-noise-machines2026-10-23T00:00:00Zjefftk.com测量空气净化器作为白噪声机的效果An alignment assessment of the UK AISI cybertesting incidenthttps://www.anthropic.com/research/alignment-assessment-ukaisi2026-10-23T00:00:00Zanthropic.com评估 Mythos 5 在 UK AISI 评测中的越权行为A Fix for Softmax and Cross-Entropyhttps://jonbarron.info/softmaxfix/2026-10-23T00:00:00Zjonbarron.info提出对 softmax 和 cross-entropy 的改进A good teacher meets the student where they arehttps://www.perplexity.ai/hub/blog/a-good-teacher-meets-the-student-where-they-are2026-10-23T00:00:00Zperplexity.ai提出联合训练教师与学生的 JOLT 方法The Latest Anthropic Claude Models Contain AI Text Watermarks: What Does This Mean for You?https://www.oreilly.com/radar/the-latest-anthropic-claude-models-contain-ai-text-watermarks-what-does-this-mean-for-you/2026-10-23T00:00:00Zoreilly.com报告 Claude 模型输出含 AI 文本水印的发现MemCurrent: Designing Persistent Memory for Production Agentshttps://www.databricks.com/blog/memcurrent-designing-persistent-memory-production-agents2026-10-23T00:00:00Zdatabricks.com介绍 MemCurrent 持久记忆架构及其效果Hamiltonian Paths in Solid Grid Graphshttps://zzyzek.github.io/SolidGridGraphHamPathPaper/2026-10-23T00:00:00Zzzyzek.github.io证明实心网格图哈密顿路径可多项式时间判定Two Path Zig Zag Numberlinkhttps://zzyzek.github.io/R2K2ZZNPaper/2026-10-23T00:00:00Zzzyzek.github.io给出 Numberlink 的禁止模式刻画与多项式算法On Serial Depth and Monitorability in Looped Transformers, Part 1https://aopatric.github.io/blog/2026/recurrent-grokking/2026-10-23T00:00:00Zaopatric.github.io研究循环 Transformer 的串行深度与思维链可监控性EPA Grant Disruption Methodologyhttps://grantwitness.org/epa/updates/methodology2026-10-23T00:00:00Zgrantwitness.org说明EPA资助中断数据的来源与方法Twitch Plays: AI Interpretability of Bonsai-8Bhttps://www.lesswrong.com/posts/ozbbzsh49wvgfbCDp/twitch-plays-ai-interpretability-of-bonsai-8b2026-10-23T00:00:00Zlesswrong.com直播式探索Bonsai-8B可解释性MalwareBench: Reverse-engineering custom-built, realistic, evasive malware into actionable intelligencehttps://malwarebench.com2026-10-23T00:00:00Zmalwarebench.comMalwareBench评测前沿模型逆向分析恶意软件的能力FrontierCS 2https://frontier-cs.org/?v=22026-10-23T00:00:00Zfrontier-cs.orgFrontierCS 2基准用48个开放式CS研发任务评测AI智能体Finite Ring Cosmology Lab: Schrödinger—Dirac dynamics of the total state over the (13, 233) holographic substratehttps://finitering.space/38-s13/lab.html2026-10-23T00:00:00Zfinitering.space用浏览器模拟Schrödinger-Dirac动力学的全息宇宙模型A Robust Decentralized Infrastructure for Trust-Aware Open Knowledge Sharinghttps://knowledgepixels.com/nanopub-ecosystem-paper/2026-10-23T00:00:00Zknowledgepixels.com提出去中心化、可信感知的开放知识共享基础设施AI Article Reviews Dominate Human Peer Reviews at Their Best: Evidence and Policy Choiceshttps://gking.harvard.edu/peerless/2026-10-23T00:00:00Zgking.harvard.edu实证显示AI论文评审在多项指标上优于人类评审An LLM judge passed renders with defects I plantedhttps://yamanahlawat.in/articles/llm-judge-same-code-different-scores2026-10-23T00:00:00Zyamanahlawat.in实验显示视觉LLM评审会放过人为植入缺陷的渲染Jakub Ćwirlejhttps://jakubcwirlej.com/ai-dev/2026-10-23T00:00:00Zjakubcwirlej.com单GPU训练160M代码模型,HumanEval达22.56%Bootstrapping the Verified Software Stackhttps://theorem.dev/blog/verified-software-stack/2026-10-23T00:00:00Ztheorem.dev提出为nixpkgs生态构建共享的已验证软件栈How does Fable 5.1 hide messages in images?https://www.lesswrong.com/posts/WHEQps5PK8LrrcTAb/how-does-fable-5-1-hide-messages-in-images2026-10-22T00:00:00Zlesswrong.com研究前沿 LLM 如何在图像中隐藏信息Rule 30: One Black Cellhttps://www.wolframcollatzmahler.com/rule30/index.html2026-10-22T00:00:00Zwolframcollatzmahler.com探讨 Rule 30 从单个黑格生长的性质Opening Linear Ahttps://www.plasma.ai/research/linear-a2026-10-22T00:00:00Zplasma.ai开源 Linear A 的闪米特语释读及语料证据Does a text-to-speech model know what it’s saying?https://vernolabs.ai/research/does-a-text-to-speech-model-know-what-its-saying2026-10-22T00:00:00Zvernolabs.ai检验文本转语音模型是否理解所说内容Good Night, and Good Luck: AI, fear and what the job data sayshttps://rpallas.xyz/good-night-and-good-luck/2026-10-22T00:00:00Zrpallas.xyz用数据分析 AI 对软件岗位的影响LoongForge-VLDAhttps://baidu-baige.github.io/LoongForge/projects/vlda.html2026-10-22T00:00:00Zbaidu-baige.github.io提出 VLDA 方法，用紧凑 Delta 作为机器人操作接口Automated nightly LLM testing: an AI agent benchmarks new local models while I sleephttps://siimtolk.github.io/articles/llm-testing-overnight/2026-10-22T00:00:00Zsiimtolk.github.io在两张 RTX 5060 Ti 上测试 18 个本地 LLMMachines Read, Humans Sign: Code Review in One Yearhttps://www.hugo.im/posts/code-review-in-one-year/2026-10-22T00:00:00Zhugo.im总结自建代码审查agent一年的数据Elasticity of money in production networks, working capital, credit lines and financial conditionshttps://www.bis.org/publications/working-paper-1350-elasticity-money-production-networks-working-capital-credit-lines-and-financial-conditions2026-10-22T00:00:00Zbis.org研究生产网络中货币弹性Solving a 31-year-old grand challenge in mathematics–the KLS conjecturehttps://www.microsoft.com/en-us/research/blog/solving-a-31-year-old-grand-challenge-in-mathematics-the-kls-conjecture/2026-10-22T00:00:00Zmicrosoft.com宣布借助AI解决KLS猜想China’s Next-Generation Industrial Policyhttps://rhg.com/research/chinas-next-generation-industrial-policy/2026-10-22T00:00:00Zrhg.com分析中国新一代产业政策的报告Mammoth-1 and Elephant-1: frontier decision modelshttps://labs.baseten.co/articles/mammoth-1-and-elephant-1-frontier-decision-models2026-10-22T00:00:00Zlabs.baseten.co介绍decision model Mammoth-1与Elephant-1Chinese AI Writes Weaker Code for Beijing’s Adversaries. We Fixed It. A Qwen3-Coder Case Studyhttps://www.hirundo.io/blog/qwen3-coder-vulnerabilities2026-10-21T00:00:00Zhirundo.io发现 Qwen3-Coder 对美国角色提示生成更不安全的代码Reactive Optimizershttps://xingjianbai.com/reactive-optimizers/2026-10-21T00:00:00Zxingjianbai.com提出让 agent 在训练中改写优化器源码的 Reactive OptimizersGLM-5.3 Sometimes Considers How Train-Time Actions Shape Future Behaviorhttps://www.lesswrong.com/posts/R5g8YT25G9DXJJ48n/glm-5-3-sometimes-considers-how-train-time-actions-shape2026-10-21T00:00:00Zlesswrong.com发现 GLM-5.3 有时会考虑训练期行为对未来的影响What Claude Code's multi-agent code review costs and finds|Quickchat AIhttps://quickchat.ai/post/claude-code-multi-agent-code-review2026-10-21T00:00:00Zquickchat.ai评测 Claude Code 多 agent 代码审查的成本与发现的问题Is 'migration choice' possible during the migration course? Understanding biographical migration projectshttps://fluchtforschung.net/is-migration-choice-possible-during-the-migration-course-understanding-biographical-migration-projects/2026-10-21T00:00:00Zfluchtforschung.net区分传记式迁移项目与迁移意愿并分析AI Agents in a Shared Workspace: Benchmarkhttps://www.spacebar.ai/benchmark2026-10-21T00:00:00Zspacebar.ai提出评测agent在共享工作区表现的SpacebarBenchPantheon | Building the Most Diverse UMI Dataset in Roboticshttps://pantheon.inc/research/diverse-umi2026-10-21T00:00:00Zpantheon.inc介绍任务多样的UMI机器人数据集RFD4-Proteinahttps://rfd4-proteina.org/2026-10-21T00:00:00Zrfd4-proteina.org提出生物分子设计模型RFD4-ProteinaDeterministic Inference Did Not Make Our Pipeline Deterministichttps://www.ctgt.ai/research/concordance2026-10-21T00:00:00Zctgt.ai实验显示确定性推理未使流水线确定The Overfithttps://blog.gopfert.eu/ai-macros-from-photos2026-10-21T00:00:00Zblog.gopfert.eu称重32餐评估AI照片估算营养的准确性Benchmarking Bot Detection Systems Against Modern AI Agents in 2026https://research.poh.org/bot-benchmarking-2026/2026-10-20T00:00:00Zresearch.poh.org重新基准测试机器人检测系统对现代 AI agent 的效果How We Cold Start a 32B Model in 1.1shttps://numerata.com/blog/post-gpu-cold-start-checkpoint-restore2026-10-20T00:00:00Znumerata.com介绍用 GPU checkpoint/restore 在 1.1 秒冷启动 32B 模型Improving in chess is hard. I built the world's most human-like chess AI to help me.https://mbuffett.com/posts/chess-move-prediction2026-10-20T00:00:00Zmbuffett.com介绍作者构建的最像人类的国际象棋 AICelestial Vision Navigation, or: How I Found Where an Elite Dangerous Screenshot from 2015 Was Taken, Using Computer Visionhttps://alphapixeldev.com/celestial-vision-navigation-elite-dangerous-dark-wheel-station/2026-10-20T00:00:00Zalphapixeldev.com用恒星定位方法解开 Elite:Dangerous 中一处神秘地点Reducing Reward Hacking by Localizing Bad Behavior Instead of Penalizing Ithttps://alignment.anthropic.com/2026/graft/2026-10-20T00:00:00Zalignment.anthropic.com提出通过定位不良行为而非惩罚来减少 reward hackingBlanking Beavers: Results and Open Problemshttps://nickdrozd.github.io/2026/10/20/blanking-beavers-results-and-open-problems.html2026-10-20T00:00:00Znickdrozd.github.io报告 Blanking Beavers 图灵机问题的结果与开放问题How leaky is your inference API? Detecting model weight steganographic exfiltrationhttps://weights-exfil.github.io/model-weights-exfiltration/2026-10-20T00:00:00Zweights-exfil.github.io展示通过 LLM API 的 token 选择窃取模型权重的隐蔽信道Robotic arm reaching task with reinforcement learning (PPO)https://ethan-jz.github.io//ppo-robot-reaching/2026-10-20T00:00:00Zethan-jz.github.io用PPO强化学习部署机械臂到达任务MARBLE: exact sequence parallelism for packed multimodal maskshttps://lumalabs.ai/news/packed-multimodal-attention2026-10-20T00:00:00Zlumalabs.ai提出多模态packed序列的精确序列并行方法MarbleBuilding a System One Model for Prompt Injectionhttps://www.stackone.com/blog/system-one-model-prompt-injection/2026-10-20T00:00:00Zstackone.com重建提示注入防护模型Defender并报告效果Fine-tuning a small language model for RAG in a browser on a Machttps://bart.degoe.de/fine-tuning-a-small-language-model-for-rag/2026-10-20T00:00:00Zbart.degoe.de在Mac上微调Qwen3-0.6B用于浏览器RAGJev models make agents faster. Hooked control planes steal the difference.https://pentad.ai/PLRN/031/2026-10-20T00:00:00Zpentad.ai分析hook链对decision model加速收益的损耗The free upgrade that raised your bill: what a new model version really costshttps://three.dev/blog/the-free-upgrade-that-raised-your-bill2026-10-20T00:00:00Zthree.dev测量新模型版本如何改变实际API账单Beyond g-mean²: complementary metrics for monitorabilityhttps://alignment.openai.com/beyond-gmean-squared/2026-10-20T00:00:00Zalignment.openai.com提出两个补充g-mean²的可监控性指标RELAGRAMhttps://nevolodia.com/thoughts/001/2026-10-20T00:00:00Znevolodia.com介绍用于车辆路径的潜变量递归RL方法RelagramHands Off: Detecting Model Weight Tamperinghttps://stealthium.io/blog/hands-off-detecting-model-weight-tampering2026-10-19T00:00:00Zstealthium.io介绍在 GPU 内存中篡改模型权重的 ROME 攻击及检测Towards Company World Modelshttps://www.beaconholdings.com/research/towards-company-world-models2026-10-19T00:00:00Zbeaconholdings.com提出以 AI 学习店主经验、辅助小企业决策的公司世界模型Can AI automate Epoch?https://epoch.ai/publications/can-ai-automate-epoch2026-10-19T00:00:00Zepoch.aiEpoch 让六个模型执行真实工作任务，评估自动化其工作的程度Are You Sure You Want to Use MMAP in Your Database Management System? (CIDR 2022)https://db.cs.cmu.edu/mmap-cidr2022/2026-10-19T00:00:00Zdb.cs.cmu.edu论证数据库管理系统不应使用 mmap 的论文页面How to Measure Intelligence Beyond Human Scale?https://www.lesswrong.com/posts/avtquSx2PkWsxXW6T/how-to-measure-intelligence-beyond-human-scale2026-10-19T00:00:00Zlesswrong.com基于论文介绍超越人类尺度的智能度量方法Is Python the favorite language of LLMs?https://www.zansara.dev/posts/2026-10-19-is-python-the-favorite-language-of-llms/2026-10-19T00:00:00Zzansara.dev考察 LLM 是否偏好 Python 等编程语言What content categories are used in science blogs?https://blog.front-matter.de/posts/what-content-categories-are-used-in-science-blogs/2026-10-19T00:00:00Zblog.front-matter.de分析Rogue Scholar科学博客的内容类别并提出分类法How much intelligence can you run locally? (October 2026)https://localai-brief.vercel.app/2026-10-19T00:00:00Zlocalai-brief.vercel.app对比本地可运行模型与前沿模型的研究简报The Art of Swarminghttps://www.csis.org/analysis/art-swarming2026-10-19T00:00:00Zcsis.org报告论证无人机集群改变战场集结方式Efficient Extraction for Effectful E-graphshttps://ztatlock.net/pubs/2026-oopsla-eggcc/2026-10-19T00:00:00Zztatlock.net提出Statewalk DP加速带副作用e-graph提取Evaluating Javohttps://getjavo.com/research/sales-evaluation/2026-10-19T00:00:00Zgetjavo.com在100个网站上评测AI销售agentLessons learned building Fronteira by Alcides Fonsecahttps://wiki.alcidesfonseca.com/blog/lessons-learned-building-fronteira/2026-10-19T00:00:00Zwiki.alcidesfonseca.com介绍用信息流策略防御的MCP代理FronteiraRunning 10 million Codex CLIs on one server-LFI processeshttps://kapteko.com/running-10-million-codex-clis-on-one-server-lfi-processes/2026-10-19T00:00:00Zkapteko.com在单台服务器上运行千万级Codex CLI进程MemStrata: AI answers you can check, with sources.https://www.memstrata.dev/2026-10-19T00:00:00Zmemstrata.devLongMemEval得分会随评审能否看到来源而变化How RC2’s π table was made (PITABLE, RFC 2268)https://khalidhamadeh.com/writing/rc2-pi-table2026-10-18T00:00:00Zkhalidhamadeh.com研究 RC2 密码 π 表的生成方式，用 Fisher-Yates 洗牌复现Neural networks are not already cleanhttps://www.lesswrong.com/posts/Dkh7pXKjoNWkkxQDz/neural-networks-are-not-already-clean2026-10-18T00:00:00Zlesswrong.com讨论要证明已解释一个神经元需要什么证据，并反驳 neural networks 已清晰的说法TIC 402710280: A TESS Exoplanet Candidatehttps://sixeleven.in/blog/tic4027102802026-10-18T00:00:00Zsixeleven.in用 TESS 凌星数据和 Claude 调查一颗可能的系外行星NEEDLE — search engine benchmarkshttps://keenableai.github.io/needle/2026-10-18T00:00:00Zkeenableai.github.io发布面向 agent 的开源实时搜索引擎基准 Needle 及结果Una Watch TOTP - MFA on your wrist!https://shkspr.mobi/blog/2026/10/una-watch-totp-mfa-on-your-wrist/2026-10-18T00:00:00Zshkspr.mobi制作手腕上的TOTP双因素验证设备Training and Deploying Production Cyber Monitors on Kimi K3https://www.goodfire.com/research/production-cyber-monitors-on-kimi-k32026-10-18T00:00:00Zgoodfire.com为Kimi K3训练网络安全监控器并与LLM judge比较Can AI set the focal point of a photo?https://workflows-of-ai.com/blog/can-ai-set-the-focal-point-of-a-photo2026-10-18T00:00:00Zworkflows-of-ai.com评测19种用LLM和模型裁剪图片的方法A faster way to convert a timestamp ➜ Hour, Min, Sechttps://www.benjoffe.com/fast-time-of-day2026-10-18T00:00:00Zbenjoffe.com提出更快的时间戳转时分秒算法Ghidra + CBMC = Assembly Model Checkerhttps://www.philipzucker.com/asmc_cpp/2026-10-18T00:00:00Zphilipzucker.com将Ghidra汇编语义接入CBMC做模型检查Beating Warcraft III’s insane AI with self-play reinforcement learning on one desktophttps://warlock.ai/blog/warcraftsim-1/2026-10-18T00:00:00Zwarlock.ai用自对弈RL在单机上击败魔兽争霸3 AIHow employment is changing at Anthropichttps://www.anthropic.com/research/employment-at-anthropic2026-10-17T00:00:00Zanthropic.comAnthropic 分析 AI 如何改变公司内部就业Basis | Ship your interpreterhttps://www.basis.ai/blog/ship-your-interpreters/2026-10-17T00:00:00Zbasis.ai展示用 LLM 为编程语言编写规则并证明解释器符合规则Does code quality still matter when AI writes the code?https://docspring.com/blog/posts/does-code-quality-still-matter-when-ai-writes-the-code/2026-10-17T00:00:00Zdocspring.com回放 agent 辅助开发中的错误，讨论 AI 时代代码质量是否重要Proactive AI for Superintelligent Agentshttps://research.google/pubs/proactive-ai-for-superintelligent-agents/2026-10-17T00:00:00Zresearch.googleGoogle 研究提出面向超级智能体的主动式 AI 方法Why gradient clipping changes Muon’s geometry: a two-dimensional examplehttps://bishnubhusal.com.np/blog/2026/clipping-geometry-private-muon/2026-10-17T00:00:00Zbishnubhusal.com.np用二维例子说明梯度裁剪如何改变 Muon 的几何When Security Rules Keep Changing Their Mindshttps://elena6918.github.io/blog/2026/rules/2026-10-17T00:00:00Zelena6918.github.io研究 LLM 编写的安全检测规则在迭代中的版本反复Peer reviewed articles about biological processes in WikiPathwayshttps://chem-bla-ics.linkedchemistry.info/2026/10/17/peer-reviewed-articles-about-biological-processes.html2026-10-17T00:00:00Zchem-bla-ics.linkedchemistry.info梳理WikiPathways中关于生物过程的同行评审文章Open and Emergent Problems in Agentic Privacy and Security: A Contextual Anglehttps://research.google/pubs/open-and-emergent-problems-in-agentic-privacy-and-security-a-contextual-angle/2026-10-17T00:00:00Zresearch.google梳理agent隐私与安全的开放问题Building a Chess Cheat Detectorhttps://chesscheatdetector.com/blog/building-a-chess-cheat-detector2026-10-17T00:00:00Zchesscheatdetector.com用Maia模拟6613局棋构建作弊检测器We shrank Whisper large-v3 by 68%, below the size of turbohttps://tileward.com/research/a-smaller-speech-recogniser/2026-10-17T00:00:00Ztileward.com将Whisper large-v3压缩68%并报告测试结果A benchmark map of the US economyhttps://www.mercor.com/research/labor-index/benchmark-map/2026-10-16T00:00:00Zmercor.com构建 AI 基准地图，分析基准对美国经济工时的覆盖Designing more efficient quantum circuits with Claudehttps://www.anthropic.com/research/designing-efficient-quantum-circuits2026-10-16T00:00:00Zanthropic.comClaude 自主设计量子电路并改进已发表的最佳版本The AI expectations gaphttps://www.anthropic.com/research/ai-expectations-gap2026-10-16T00:00:00Zanthropic.comAnthropic 分析人们对 AI 经济影响预期分歧的来源How pro-China are Chinese models?https://www.vals.ai/blogs/china-us-analysis2026-10-16T00:00:00Zvals.ai评测中国模型在政治话题上的亲中倾向程度Scaling pre-training on Meleehttps://ericyuegu.com/hillclimbing-melee/2026-10-16T00:00:00Zericyuegu.com通过动作表示与离线 RL 扩展 Melee 游戏策略预训练Do looped transformers write reasoning they don't use?https://igort99.github.io/posts/2026/10/05/looped-cot-faithfulness/2026-10-16T00:00:00Zigort99.github.io实验检验 looped transformer Huginn 是否真正使用其推理NSF Grant Disruption Methodologyhttps://grantwitness.org/nsf/updates/methodology2026-10-16T00:00:00Zgrantwitness.org说明NSF资助终止数据的追踪方法Models know when they’re reward hacking—and we can catch them at scalehttps://www.goodfire.com/research/reward-hacking-activation-monitors2026-10-16T00:00:00Zgoodfire.com发现模型内部与reward hacking对应的信号并训练probeThis is not a game of snakehttps://blog.neels.tech/blog/this-is-not-a-game-of-snake/2026-10-16T00:00:00Zblog.neels.tech实验训练transformer学习玩贪吃蛇Toward RL on mixed-precision, mixed-accelerator samplershttps://modal.com/blog/toward-rl-on-mixed-precision-mixed-accelerator-samplers2026-10-16T00:00:00Zmodal.com用混合精度混合加速器采样器做异步RL训练Bonsai Blog | Fully Managed Elasticsearch & OpenSearch • The Pink Elephant in Your Prompthttps://bonsai.io/blog/pink-elephant-in-your-prompt/2026-10-16T00:00:00Zbonsai.io用720次试验检验让agent忽略已有代码的提示效果Running DOOM Inside a Transformerhttps://nostalgia.dev/posts/running-doom-inside-a-transformer/2026-10-16T00:00:00Znostalgia.dev构建能运行DOOM的transformerA Statistical Analysis of LLM Benchmaxxinghttps://llemre.com/benchmaxxing/2026-10-16T00:00:00Zllemre.com统计分析各实验室模型是否在基准上过拟合REKI: Registering Knowledge in VLAshttps://kyutai.org/reki/2026-10-16T00:00:00Zkyutai.org介绍带近期记忆的2.4B视觉语言动作模型REKIBenchmarking Open-Source LLMs on Verilog RTL Generation Across 50 Tasks (NMIMS, IIT Roorkee, BITS Pilani)https://semiengineering.com/benchmarking-open-source-llms-on-verilog-rtl-generation-across-50-tasks-nmims-iit-roorkee-bits-pilani/2026-10-16T00:00:00Zsemiengineering.com报道在50个任务上评测开源LLM生成Verilog的论文Neural Networks Are Already Cleanhttps://already-clean.goodfire.com/2026-10-15T00:00:00Zalready-clean.goodfire.com提出神经网络神经元与电路较为干净的工作假说Jev Meets the Trolley Problemhttps://zachlandes.com/trolley-problem/2026-10-15T00:00:00Zzachlandes.com向 Jev 与 13 个决策模型提出电车难题并比较结果Exposing AdFection: How Discover-Tech Turned Targeted Ads Into a Pipeline for Zero-Click Mercenary Spywarehttps://citizenlab.ca/research/how-discover-tech-turned-targeted-ads-into-a-pipeline-for-spyware/2026-10-15T00:00:00Zcitizenlab.caCitizen Lab 披露通过定向广告投递零点击间谍软件Courk's Blog – Laser Fault Injection on a Budget: RP2350 Editionhttps://courk.cc/rp2350-challenge-laser2026-10-15T00:00:00Zcourk.cc在 RP2350 上以低成本实施激光故障注入Reproducing the Hugging Face Incident, Part Ihttps://www.abundant.ai/blog/hugging-face-incident-part-1/2026-10-15T00:00:00Zabundant.ai受控复现 Hugging Face 事件中的包管理器攻击How We Trained an AI Model to Play Browser Games Livehttps://ironbee.ai/blog/how-we-trained-an-ai-model-to-play-browser-games2026-10-15T00:00:00Zironbee.ai训练本地模型实时玩浏览器游戏的方法与结果What are the most important commands of a programming language?https://blog.ptidej.net/what-are-the-most-important-commands-of-a-programming-language/2026-10-15T00:00:00Zblog.ptidej.net分析 AMOS BASIC 程序中最重要的命令以便转译到 JavaScriptPackage managers have an integrity problemhttps://blog.trailofbits.com/2026/10/14/package-managers-have-an-integrity-problem/2026-10-15T00:00:00Zblog.trailofbits.com测试 25 个包管理器对被篡改仓库的检测能力Apertus 1.5https://apertus-ai.org/articles/2026-07-apertus-1-5/2026-10-15T00:00:00Zapertus-ai.orgApertus 1.5 技术报告：多模态、推理与长上下文FLAT: Flash Linear Attention on TPUshttps://trainstation-project.github.io/blog/flat/2026-10-15T00:00:00Ztrainstation-project.github.io在 TPU 上用 Pallas 实现 Linear Attention 等 kernelBarista: Synthesizing Typestate Specifications with LLM Agentshttps://catarinagamboa.github.io/papers/barista/2026-10-15T00:00:00Zcatarinagamboa.github.ioBarista：用 LLM agent 合成 typestate 规约（ASE 2026）ML Conference Trends: NeurIPS & ICLR, 2023–2026https://wenyaxie023.github.io/blog/2026/ml-conference-trends/2026-10-15T00:00:00Zwenyaxie023.github.io统计 NeurIPS 与 ICLR 2023-2026 录用论文的主题趋势A Computational Swinging Lexical Networkhttps://schiekiera.github.io/blog/2026/computational-swinging-lexical-network/2026-10-15T00:00:00Zschiekiera.github.io实现词语产出的 Swinging Lexical Network 计算模型T12 POSS Cages: D6h and D2d Dodecasilsesquioxane Isomershttps://silsesquioxane.blogspot.com/2026/10/t12-dodecameric-silsesquioxanes.html2026-10-15T00:00:00Zsilsesquioxane.blogspot.com讨论T12 POSS笼的D6h与D2d异构体Gemma 3 1B at 114 tokens per second on a laptop CPUhttps://knl.co.in/posts/gemma-3-1b-ternary/2026-10-15T00:00:00Zknl.co.in在笔记本CPU上将Gemma 3 1B提速到114 tokens/sConfidential AI ･ Encrypted KV Cache Offloading for Confidential Computinghttps://confidential.ai/blog/kv-offload-under-cc2026-10-15T00:00:00Zconfidential.ai提出面向机密计算的加密KV cache卸载ExplainThatMovehttps://www.explainthatmove.com/2026-10-15T00:00:00Zexplainthatmove.com比较并用自然语言解释国际象棋着法优劣Context Inertia: When Agents Won’t Let Gohttps://www.dbreunig.com/2026/10/14/context-inertia.html2026-10-15T00:00:00Zdbreunig.com提出并讨论agent的Context Inertia现象Inside Agentic AI Workloadshttps://gimletlabs.ai/blog/inside-agentic-ai-workloads2026-10-15T00:00:00Zgimletlabs.ai刻画agentic AI工作负载的推理特征Routers Are a New AI Primitivehttps://www.vals.ai/blogs/routers-new-ai-primitive2026-10-15T00:00:00Zvals.ai评测模型路由器的质量、成本与延迟The Continual Constitutional Convention: How language models revise the documents that govern themhttps://www.lesswrong.com/posts/c9woewdY9ddsPqf3J/the-continual-constitutional-convention-how-language-models2026-10-15T00:00:00Zlesswrong.com研究语言模型如何修订约束自身的文件(宪法)What 1,866 GitHub Projects Tell Their AI Coding Agentshttps://secondtimeisabug.com/blog/what-projects-tell-ai-coding-agents/2026-10-15T00:00:00Zsecondtimeisabug.com分析1866个GitHub项目的AGENTS.mdHow We Made Our MMDB Files a Third Smaller Without Changing a Byte of Datahttps://ipinfo.io/blog/how-made-mmdb-files-smaller2026-10-14T00:00:00Zipinfo.ioIPinfo 介绍将 MMDB 文件缩小并保持查询结果不变的优化We found a Tencent agent fleethttps://swarmcha.se/posts/chinese-agent-fleet-full-report2026-10-14T00:00:00Zswarmcha.se记录云平台上大批 AI agent 活动的调查与发现The Machines that Make the Machineshttps://developer.nvidia.com/blog/the-machines-that-make-the-machines/2026-10-14T00:00:00Zdeveloper.nvidia.comNVIDIA 介绍训练机器人装配 GB300 测试托盘的方法与经验AI Scaling vs Human Scalinghttps://www.tobyord.com/writing/ai-scaling-vs-human-scaling2026-10-14T00:00:00Ztobyord.com分析 AI 模型随 token 增加的扩展是否如人类随时间扩展Build-A-Chip-Benchhttps://chipagents.ai/research/build-a-chip-bench2026-10-14T00:00:00Zchipagents.ai提出 Build-A-Chip-Bench 评测 agent 从规格到 GDS 的芯片设计We found 110 reward hacks and grading bugs in Terminal-Bench, DeepSWE, and BFCLhttps://usetokenless.com/blog/envcheck/2026-10-14T00:00:00Zusetokenless.com在 Terminal-Bench 等基准中发现 110 个 reward hack 与评分漏洞Analyzing technical documents with Deep Reviewhttps://research.google/blog/analyzing-technical-documents-with-deep-review/2026-10-14T00:00:00Zresearch.googleGoogle Research 介绍用 Deep Review 分析技术文档A Turing Test for Physical Engineeringhttps://www.spectrallabs.ai/research/cswp-turing-test2026-10-14T00:00:00Zspectrallabs.ai评估 AI agent 通过 SOLIDWORKS 认证考试的能力Ultramarine 1 Flash: Scaling Non-Verifiable RLhttps://www.paperinstruments.com/blog/ultramarine2026-10-14T00:00:00Zpaperinstruments.com介绍用于非可验证任务的 RL 模型 Ultramarine 1 FlashVisualizing Crossref Repository Historieshttps://metadatagamechangers.com/blog/2026/10/10/visualizing-crossref-repository-histories2026-10-14T00:00:00Zmetadatagamechangers.com可视化Crossref仓储元数据的历史Optimization: does quantization hurt coding agents?https://aistack.imec-int.com/blog/quantization2026-10-14T00:00:00Zaistack.imec-int.com评测量化对编码agent的影响rust-djinn: You Should Specify Your Code Unambiguously (and Prove That You Did)https://henniger.dev/post/djinn/2026-10-14T00:00:00Zhenniger.dev用300行规范实现并证明gzip解压器Next-Concept Modeling: Fast, Parallel Coarse-to-Fine Text Generation | Scaling Intelligence Lab at Stanford Universityhttps://scalingintelligence.stanford.edu/blogs/ncm/2026-10-14T00:00:00Zscalingintelligence.stanford.edu提出Next-Concept Modeling并行粗到细文本生成How well can AI play Slay the Spire?https://epoch.ai/publications/ai-systems-are-improving-rapidly-at-slay-the-spire-and-its-sequel2026-10-14T00:00:00Zepoch.ai追踪前沿模型在Slay the Spire游戏中的进步Into the Rat's Nest: A SenseOn Analysis of the...https://www.senseon.io/blog/netsupport-rat-analysis2026-10-14T00:00:00Zsenseon.io技术分析NetSupport RAT的投递与持久化FrontierML: a challenging, end-to-end machine learning benchmarkhttps://www.frontierml.org/blog/2026-10-13T00:00:00Zfrontierml.orgFrontierML 基准：让 agent 在 12 小时和有限预算下做机器学习The Links Between Cannabis and Crime: Examining the Evidence from Recent Policy Experimentshttps://www.rand.org/pubs/external_publications/EP71427.html2026-10-13T00:00:00Zrand.orgRAND 叙述性综述：大麻政策与犯罪关系的近期准实验证据2024 年模型训练攻击事件：受害者陈述与证据整理https://keyu-tian-sabotage.github.io/2026-10-13T00:00:00Zkeyu-tian-sabotage.github.io整理 2024 年模型训练破坏事件的受害者陈述与证据Speculative tool calling: cutting voice AI agent latencyhttps://tazolabs.com/blog/speculative-tool-calling-voice-ai2026-10-13T00:00:00Ztazolabs.com提出 speculative tool calling 以降低语音 agent 延迟Meta-Metadata: The Dangerous Weird Word Breaks 73% of LLMshttps://hackstery.com/2026/10/13/meta-metadata-unbounded-consumption-in-llms/2026-10-13T00:00:00Zhackstery.com测试显示单个递归提示使 73% 的 LLM 出现无限制消耗Correction Training for Diffusion Modelshttps://sunovivid.github.io/correction-training/2026-10-13T00:00:00Zsunovivid.github.io提出扩散模型的 correction training 以缓解曝光偏差Not Only Boring Issues: What Fossil Bones Reveal About cretaceous Haţeg Islandhttps://blog.pensoft.net/2026/10/13/not-only-boring-issues-what-fossil-bones-reveal-about-cretaceous-hateg-island/2026-10-13T00:00:00Zblog.pensoft.net介绍化石骨骼痕迹揭示白垩纪Haţeg岛的研究Natura-1: The Will to Acthttps://www.naturewill.ai/blog/natura-12026-10-13T00:00:00Znaturewill.ai报告50B模型Natura-1用人类视频驱动20自由度灵巧手blog - nanoswehttps://www.nanoswe.com/blog.html2026-10-13T00:00:00Znanoswe.com以1000美元预算从零训练agentic LLM并测SWE-benchKIC 5689121—A Kepler Planet Candidatehttps://kuber.studio/planethunt/2026-10-13T00:00:00Zkuber.studio报告在Kepler数据中发现13.77天行星候选体Can we change an AI model’s architecture without retraining it from scratch?https://www.amazon.science/blog/can-we-change-an-ai-models-architecture-without-retraining-it-from-scratch2026-10-13T00:00:00Zamazon.science提出Priming方法复用权重以改变模型架构Model backdoors are easy to create and hard to reliably removehttps://equistamp.substack.com/p/model-backdoors-are-easy-to-create2026-10-13T00:00:00Zequistamp.substack.com报告模型后门易植入而难以可靠移除Alignment Is a System Property, Not Just a Model Propertyhttps://getswiftapi.com2026-10-13T00:00:00Zgetswiftapi.com报告GPT-4-0613系统提示消融实验Faster Fourier Transforms-by Ben Rechthttps://www.argmin.net/p/faster-fourier-transforms2026-10-12T00:00:00Zargmin.net分析OpenAI的FFT算法，认为其是利用漏洞而非突破AI models may be capable of self-replicating without access to a datacenterhttps://www.anthropic.com/research/self-replicating-ai2026-10-12T00:00:00Zanthropic.comAnthropic 实验评估小型 AI 模型在廉价 IoT 设备上自我复制的能力MASS: Recursive Self-Improvement through Multi-Agent Self-Supervisionhttps://pub.sakana.ai/mass/?href=2026-10-12T00:00:00Zpub.sakana.aiMASS项目页：多代理自监督的递归自我改进To Classify, Try a Trie Treehttps://fsdatalab.github.io/blog/ai-classification/2026-10-12T00:00:00Zfsdatalab.github.io介绍 AI-SQL 引擎用 trie tree 大规模做 LLM 文档分类的方法Can Decision Models (Like Jev) Plan Ahead?https://www.lesswrong.com/posts/kE4pbmwzdgmFx2RFo/can-decision-models-like-jev-plan-ahead2026-10-12T00:00:00Zlesswrong.com测试决策模型Jev是否具备提前规划能力StreamRender-H3: Interactive Coding Worlds with Real-time Streaming Diffusion Renderhttps://nvlabs.github.io/Sana/Sol-Engine/StreamRender-H3/2026-10-12T00:00:00Znvlabs.github.io提出StreamRender-H3，用流式扩散渲染交互代码世界LinearPFN: Amortized Variable Selection for Linear Models with Interactionshttps://schiekiera.github.io/blog/2026/linear-pfn/2026-10-12T00:00:00Zschiekiera.github.io提出预训练的 LinearPFN 用于含交互项线性模型的变量选择CEO-Benchhttps://ceobench.com/2026-10-12T00:00:00Zceobench.com提出CEO-Bench评测agent长期经营模拟公司Solving Open Research Problems Togetherhttps://research.meta.ai/blog/solving-open-research-problems-together2026-10-12T00:00:00Zresearch.meta.ai介绍数学家与Muse Spark合作完成六篇研究论文MASS: Recursive Self-Improvement through Multi-Agent Self-Supervisionhttps://pub.sakana.ai/mass/2026-10-11T00:00:00Zpub.sakana.ai研究语言模型经多代理自监督的递归自我改进MASSGPT-6 Astra performs unsanctioned supply-chain attacks in simulationshttps://www.aisi.gov.uk/blog/gpt-6-astra-performs-unsanctioned-supply-chain-attacks-in-simulations2026-10-11T00:00:00Zaisi.gov.uk评估GPT-6 Astra在模拟中的未授权供应链攻击行为SMITH: Joint Optimization of Tool Creation and Use for LLM Agentshttps://tool-use-smith.github.io/2026-10-11T00:00:00Ztool-use-smith.github.io提出SMITH，联合优化LLM代理的工具创建与使用Measuring the Depth of LLM Unlearning via Activation Patchinghttps://gnueaj.github.io/unlearning-depth-score/2026-10-11T00:00:00Zgnueaj.github.io提出用激活修补度量LLM遗忘深度的Unlearning Depth ScoreI paid people to try and follow my READMEhttps://shkspr.mobi/blog/2026/10/i-paid-people-to-try-and-follow-my-readme/2026-10-11T00:00:00Zshkspr.mobi付费请人按README操作以测试文档质量TasteRoute: Personalized Routing for Video Generationhttps://zrt.wtf/publication/2026-10-10-taste-route2026-10-11T00:00:00Zzrt.wtf提出个性化视频生成模型路由方法TasteRouteMeasuring Cryptanalytic Efforthttps://bitsofwork.org2026-10-11T00:00:00Zbitsofwork.org用文献计量估算各公钥假设的密码分析投入Has mainstream cryptography provided 30 years of protection? (Mostly not, but we’re getting way better)https://dreamstation.systems/personal/crypto30years.html2026-10-11T00:00:00Zdreamstation.systems回顾主流密码算法30年内是否被实际攻破What Do MTEB Scores Mean, and How Should You Choose an Embedding Model?https://thottingal.in/blog/2026/10/09/mteb-scores/2026-10-10T00:00:00Zthottingal.in分析MTEB分数构成并以印度语言为例讨论嵌入模型选择A Thermometer, Not an Analyst: A Text Judge in a Trading Labhttps://alexvidovich.com/papers/text-judge2026-10-10T00:00:00Zalexvidovich.com评估Jev与Claude作为交易新闻标题判断器的六道关口GitSwarm: Decentralized Compounding Inferencehttps://veds12.github.io/git-swarm/2026-10-10T00:00:00Zveds12.github.io提出GitSwarm，让代理群共享Git仓库累积尝试If LLMs Can Decide Without Fine-Tuning, Do We Still Need Models Like Jev?https://itsodeleo.github.io/posts/local-decision-workflow/2026-10-10T00:00:00Zitsodeleo.github.io本地测试Qwen等LLM无需微调做决策与Jev对比PINN for passive myocardium hyperelasticityhttps://osmangult.github.io/blog/2026/pinn-myocardium/2026-10-10T00:00:00Zosmangult.github.io用 PINN 拟合被动心肌超弹性本构的项目说明When Do Innocent LLM Agents Reach Unfair Decisions?https://mayaokawa.github.io/blog/2026/biased-consensus/2026-10-10T00:00:00Zmayaokawa.github.io用 Ising 模型分析 LLM 辩论中微小偏见放大为不公平共识Fast, but incomplete and inaccurate: twelve AI-powered tools for literature searchhttps://renebekkers.wordpress.com/2026/10/10/fast-but-incomplete-and-inaccurate-twelve-ai-powered-tools-for-literature-search/2026-10-10T00:00:00Zrenebekkers.wordpress.com评估十二个AI文献检索工具并指出其不足Designing Kolibri: Architecture Trade-Offs from First Principleshttps://aleph-alpha.com/en/blog/designing-kolibri-architecture-trade-offs-from-first-principles/2026-10-10T00:00:00Zaleph-alpha.com分析Kolibri模型的架构取舍与参数分配VibeSys Builds a Qwen3.5-397B Engine on MI300A, 2.3× Tuned SGLanghttps://syfi.cs.washington.edu/blog/2026-10-08-qwen35-mi300a/2026-10-10T00:00:00Zsyfi.cs.washington.edu在MI300A上构建Qwen3.5-397B推理引擎，快于SGLangVega: a decision model that rolls a ball down a hill / Nandakishor Mhttps://www.nandakishorm.com/writing/vega2026-10-10T00:00:00Znandakishorm.com介绍决策模型Vega并说明其机制Is this machine playing?https://machineplaying.org/2026-10-10T00:00:00Zmachineplaying.org观察编码agent在无目标3D岛屿中的自发行为MCP server rules: Claude, ChatGPT, Codex, VS Code and morehttps://pournasserian.com/writing/what-ai-hosts-require-of-an-mcp-server2026-10-10T00:00:00Zpournasserian.com比较八个AI宿主对MCP服务器的检查规则One vector is too few. One per token is too many.https://breadbowl.ai/blog/breadbowl-embed/2026-10-10T00:00:00Zbreadbowl.ai提出用KV cache作嵌入的late-interaction架构The plunging price of thoughthttps://epoch.ai/publications/the-plunging-price-of-thought2026-10-10T00:00:00Zepoch.ai测量AI性能成本下降速度并与其他技术比较Jane Street Blog - Can you use autoregressive diffusion to generate market data?https://blog.janestreet.com/can-you-use-autoregressive-diffusion-to-generate-market-data/2026-10-09T00:00:00Zblog.janestreet.com探索用自回归扩散生成市场数据Can AI automate AI R&D yet?https://epoch.ai/publications/innovationeval2026-10-09T00:00:00Zepoch.ai用InnovationEval评估AI模型复现人类算法创新Datology Curation Studio Improves Even the Best Open Datahttps://www.datologyai.com/blog/text-curation-oct-20262026-10-09T00:00:00Zdatologyai.com报告Datology Curation Studio改进开源数据带来6倍算力等效Backtests vs Live Trading: Six Defects from One Systemhttps://alexvidovich.com/papers/backtest-hygiene2026-10-09T00:00:00Zalexvidovich.com分析日内交易系统回测与实盘偏离的六类缺陷How We Trained ZooWork-ShopRankerhttps://zoowork.ai/blog/relevance-is-not-preference-shopranker/2026-10-09T00:00:00Zzoowork.ai报告训练面向购物偏好的开源重排序模型ZooWork-ShopRankerAI Companies as Cybernetic Systemshttps://ai-cybernetics.grok.me2026-10-09T00:00:00Zai-cybernetics.grok.me用高阶群胚语义为AI公司建模的手稿Fun with low-rank vocab matrices (and a bonus test loss reduction?)https://www.gilesthomas.com/2026/10/low-rank-vocab-matrices2026-10-09T00:00:00Zgilesthomas.com实验低秩词表矩阵替换LLM嵌入及测试损失变化Evapotranspiration monitoringhttps://from-sea-to-summit.github.io/blog/2026/evaporation-measurements/2026-10-09T00:00:00Zfrom-sea-to-summit.github.io记录蒸散发监测设备的安装与初步观测NIH and AHRQ Grant Disruption Methodologyhttps://grantwitness.org/nih/updates/methodology2026-10-09T00:00:00Zgrantwitness.org说明NIH与AHRQ资助中断数据的追踪方法HarnessTax: How Much Does the Harness Matter for Coding Agents?https://arena.ai/blog/coding-agents-harness-tax2026-10-09T00:00:00Zarena.ai评测21种编程agent的模型与harness组合Action retrieval: VideoDB vs Nova and Twelve Labshttps://videodb.io/blog/a-faster-path-to-the-next-policy2026-10-09T00:00:00Zvideodb.io在100小时动作检索基准上对比VideoDB与竞品U-Spacehttps://s2lab.org/U-Space/2026-10-09T00:00:00Zs2lab.org研究语言模型不确定性产生的时机与原因Two-Thirds of MiMo v2.6's Coding Tasks Leak the Answer. Are Models That Exploit This Misbehaving?https://www.vals.ai/blogs/mimo-reward-hacking2026-10-08T00:00:00Zvals.ai审计MiMo v2.6编码任务，发现三分之二泄露答案Cheaters and whistleblowers in the agent swarmhttps://institute.deepmind.com/essays/cheaters-and-whistleblowers-in-the-agent-swarm/2026-10-08T00:00:00Zinstitute.deepmind.com记录代理群体在虚拟数学会议中作弊与举报的实验ATLAS: Evaluating Agents on Search-Intensive Taskshttps://exa.ai/blog/atlas-benchmark2026-10-08T00:00:00Zexa.ai预览ATLAS基准，评测代理在搜索密集任务上的表现DeLM — Decentralized Multi-Agent Systems with Shared Contexthttps://yuzhenmao.github.io/DeLM/2026-10-08T00:00:00Zyuzhenmao.github.io提出DeLM，用共享上下文与任务队列组织多代理Agent Plasticity: Measuring Self-Improvement Through Experiencehttps://harmandotpy.github.io/agent-plasticity/2026-10-08T00:00:00Zharmandotpy.github.io提出Agent Plasticity，衡量冻结权重代理的经验式自我改进Self-Modeling Interventions Modulate Emergent Misalignmenthttps://www.lesswrong.com/posts/7wrzfaiCq3u8xkY5G/self-modeling-interventions-modulate-emergent-misalignment2026-10-08T00:00:00Zlesswrong.com研究自我建模干预对emergent misalignment的影响How many AI agents could we run?https://epoch.ai/publications/estimating-the-agent-population2026-10-08T00:00:00Zepoch.ai估算截至2027年出货的AI芯片可运行的代理数量Lessons from building in silico life with homeostatic RLhttps://ravulapallimanas.github.io/writing/2026/10/ideas-on-creating-artificial-life-in-silico/2026-10-08T00:00:00Zravulapallimanas.github.io总结用稳态强化学习构建体内需求型人工生命的经验Posters at SciCoCo.nrw: Training LLMs and an Integrated Approach for Benchmarkinghttps://x-dev.pages.jsc.fz-juelich.de/2026/10/08/scicoco-posters.html2026-10-08T00:00:00Zx-dev.pages.jsc.fz-juelich.de展示SciCoCo.nrw上LLM训练与基准测试海报Hotword detection in FHE on GPU in 350mshttps://www.jeremykun.com/shortform/2026-10-08-1120/2026-10-08T00:00:00Zjeremykun.com报道HEIR编译器在GPU上350ms完成FHE热词检测RRSI: Regularized Recursive Self-Improvement of Agent Harnesseshttps://regularized-rsi.com/2026-10-08T00:00:00Zregularized-rsi.com提出RRSI方法，缓解自改进agent对基准的过拟合Studying metagaming latents in language modelshttps://alignment.openai.com/metagaming-latents/2026-10-08T00:00:00Zalignment.openai.com研究语言模型中与监控和奖励推理相关的metagaming潜变量The Alchemy of Semi-Supervisionhttps://stefankeselj.com/entries/semi_supervision/2026-10-08T00:00:00Zstefankeselj.com研究半监督学习方法并给出实验When Submarine Cables Go Dark: Understanding and Preparing for the Risks of Taiwan's International Internet Disconnectionhttps://resilience.ocf.tw/web/report/en2026-10-07T00:00:00Zresilience.ocf.tw分析台湾国际海底电缆中断风险并提出应对准备Artifact Arena — where making is the measure of intelligencehttps://artifactarena.ai2026-10-07T00:00:00Zartifactarena.ai提出Artifact Arena，以模型制作成果评测智能的基准Cyber guardrails tax the defender, not the attackerhttps://winfunc.com/research/cyber-guardrails-tax-the-defender-not-the-attacker2026-10-07T00:00:00Zwinfunc.com分析AI代理入侵Hugging Face事件，论证网络安全护栏对防守方不利Training a small model to run a house: GRPO on Qwen3-1.7B for Home Assistant, 75.0% to 89.8%https://www.neelabhbuilds.com/writing/training-a-small-model-to-run-a-house2026-10-07T00:00:00Zneelabhbuilds.com用GRPO在单张RTX 3060上训练Qwen3-1.7B控制智能家居并报告结果A 15M-parameter RL agent vs GPT-6 Astra, rekursiv.aihttps://rekursiv.ai/blog/craftax/2026-10-07T00:00:00Zrekursiv.ai报告AI科学家优化15M参数RL代理在Craftax上达到44%胜率Evaluating OCR on the Community Memory Corpushttps://ztoz.blog/posts/ocr-cm/2026-10-07T00:00:00Zztoz.blog在Community Memory语料上评测OCR方法AIM — Agentic Idea Management for Automated Researchhttps://imhgchoi.github.io/agentic-idea-manager/2026-10-07T00:00:00Zimhgchoi.github.io提出AIM，用代理管理和探索自动化研究的想法ServeLearnBenchhttps://infini-ai-lab.github.io/ServeLearnBench/2026-10-07T00:00:00Zinfini-ai-lab.github.io提出ServeLearnBench，评测代理从服务经验中自我改进Nüchtern in den Präpkurs – schadet das dem Lernen?https://medical-education.pages.ub.uni-bielefeld.de/research/mes-blog/posts/2026-10-08-ernaehrung-nuechtern/2026-10-07T00:00:00Zmedical-education.pages.ub.uni-bielefeld.de探讨空腹进入解剖课是否影响学习