9月16日9月16日周三

星期三 · 2 条

18:29

DAIR.AI@dair_ai

AI 评分 59/100

微软论文提出能力洗白攻击：未对齐小模型拆分任务借调前沿模型完成有害目标Interesting safety paper from Microsoft.They find that a weaker, unaligned model can split a harmful task into harmless-looking subquestions, ask an aligned frontier model each one in a separate session, and combine the answers locally.The authors call this capability laundering.Each request passes on its own, because no single answer from the frontier model is a harmful task.They tested GPT-5.5, Claude Opus 4.8 and Grok-4.3 as the consulted models. On CyBench, Gemma-4-31B recovered 8 of 14 tasks it failed alone when it consulted GPT-5.5. On a CBRN attack chain, consultation raised its mean rubric score from 62.3 to 83.1.Paper: https://academy.dair.ai/papers/divide-consult-conquer-capability-laundering-through-aligned-llms-2609.15383译微软研究团队发布论文，发现较弱的未对齐模型可将有害任务拆成看似无害的子问题，分别在独立会话中询问对齐的前沿模型，再在本地合并答案，作者称之为 capability laundering。

Microsoft 论文/研究

02:29

DAIR.AI@dair_ai

AI 评分 61/100

微软论文：Bash 在企业 Agent 任务上超越类型化工具接口Is Bash All You Need?Interesting paper from Microsoft.If you are deciding which tools to give an enterprise agent, you might want to check this out.(bookmark it)They compared five tool interfaces on TheAgentCompany and APEX-Agents with Opus-4.8 and GPT-5.5. The options range from a catalog of typed tools, to bash alone, to programmatic tool calling where the agent writes code against a fixed catalog.Bash alone scored 21.8 to 24.5 points higher than typed tools on TheAgentCompany and 4.8 to 7.4 points higher on APEX-Agents, while using 19% to 72% fewer tokens. Adding typed tools or agent-written tools on top of bash gave no measurable gain.The authors recommend bash when execution can be sandboxed, and programmatic tool calling when compliance requires a fixed tool list.Paper: https://academy.dair.ai/papers/is-bash-all-you-need-an-empirical-study-of-tool-interfaces-for-enterprise-digita-2609.11999译微软论文比较五种工具接口，在 TheAgentCompany 和 APEX-Agents 上用 Opus-4.8 和 GPT-5.5 测试。

智能体 Microsoft 论文/研究

9月11日9月11日周五

星期五 · 1 条

23:29

DAIR.AI@dair_ai

AI 评分 50/100

Microsoft 论文提出环境探测式记忆管理，将 GitHub Copilot 测试通过率从 39% 提升至 73%Interesting paper from Microsoft.If you run persistent memory for a production agent, this one is worth your time. (bookmark it)Memory curators usually read only the finished trajectory.That lets them save the agent's mistakes, overgeneralize from partial evidence, and keep facts that have gone stale.Microsoft researchers give the curator a few read-only tools to check each candidate memory against the live environment before it is saved. The task agent, retriever and memory format stay the same, and nothing is retrained.In a GitHub Copilot harness on CLBench, pass rate goes from 39% to 73%. Queries per question drop from 8.8 to 4.7 and task-agent cost falls from $3.38 to $1.68.On 90 consulting tasks across six environments, every memory configuration beats the baseline, and tool calls fall by 16 to 75%.Paper: https://arxiv.org/abs/2609.11060Chat with Paper: https://academy.dair.ai/papers/grounding-agent-memory-environment-probing-curation-for-enterprise-agents-2609.11060译Microsoft 研究者提出 environment-probing curation 方法，让记忆管理器在保存前用只读世界工具核对候选记忆与实时环境，无需重训模型，任务智能体、检索器和记忆格式保持不变。

智能体 Microsoft 论文/研究

9月10日9月10日周四

星期四 · 1 条

02:29

DAIR.AI@dair_ai

AI 评分 64/100

Microsoft 等提出 Detokenization Leaks 攻击，可从 CPU 缓存痕迹重建本地 LLM 输出Wild paper from Microsoft and colleagues.They show a new attack that reconstructs the text a local LLM generates by watching CPU cache activity while it detokenizes.Earlier cache attacks needed something unusual in the deployment, such as shared data memory, CPU offloading, or a Mixture-of-Experts architecture. This work targets the detokenizer, which runs in default inference pipelines.The method has two stages.1) Flush+Reload on shared tokenizer code detects when decoding happens, which lets the attacker fire Prime+Probe at the right moment and isolate token-dependent cache activity.2) A clustering and language-model pipeline then recovers readable text from the noisy observations.They evaluate across datasets, hardware platforms, inference frameworks and model families, including real local deployments and agentic systems.The widely used tokenizer implementations are susceptible, and they are embedded in many popular local LLM products and agent frameworks. OpenClaw is demonstrated directly.Paper: https://academy.dair.ai/papers/detokenization-leaks-reconstructing-local-llm-outputs-from-cache-traces-2609.06674译Microsoft 与 Ben Gurion 大学等的研究者展示一种新攻击，通过监视 detokenization 时的 CPU 缓存活动，重建本地 LLM 生成的文本。

Microsoft OpenAI 论文/研究

9月9日9月9日周三

星期三 · 1 条

22:59

DAIR.AI@dair_ai

AI 评分 59/100

微软发布 FrogNano 报告：用在线任务合成训练 4B 编码智能体Banger report from Microsoft.(bookmark it)They show that it's possible to build competitive small coding agents without traditional distillation from frontier models.This is a big deal!The work describes how they achieved this.They introduce a 4B coding agent trained on roughly 1,500 software engineering environments.The cool thing is that they use no distillation from a larger model at any point.FrogNano is post-trained purely with RL on synthetic tasks.The target is a coding agent that runs on minimal machines, which rules out both a frontier backbone and a frontier teacher.The ingredient the report credits the most is online task synthesis.The pipeline generates tasks calibrated to the frontier of learnability for the current checkpoint, so the agent always trains on problems it can just barely solve. The authors argue that calibration, rather than the volume of synthetic data, is what makes this work.This means that competitive small coding agents can be trained from synthetic tasks alone.And generating those tasks at the current agent's learnability frontier is what makes this particular training productive.The report covers training methodology, evaluations across diverse environments, and analyses of what the agent learned.Paper: https://academy.dair.ai/papers/frognano-training-a-4b-coding-agent-via-online-task-synthesis-2609.07925译微软发布 FrogNano 报告，训练出一个 4B 编码智能体，在约 1，500 个软件工程环境上纯用 RL 后训练，全程不依赖更大模型蒸馏。关键是在线任务合成流水线，按当前 checkpoint 的可学习性前沿生成任务，作者认为校准而非合成数据量是成效所在。

智能体 Microsoft 数据/训练编码

9月7日9月7日周一

星期一 · 1 条

01:40

Rohan Paul@rohanpaul_ai

AI 评分 64/100

Microsoft 论文提出把推理成本蒸馏为技能，GPT-5.4-mini 用更少 token 恢复 55%-100%+ 推理收益What if you could pay the reasoning cost once, then reuse what the model learned across future tasks?New Microsoft paper finds that some expensive test-time reasoning can be replaced with a small set of rules learned from previous agent runs.The paper tests a cheaper alternative: collect 35–50 past trajectories, have a coding agent extract recurring failure patterns, then turn those patterns into a small markdown skill added to the non-reasoning model’s system prompt.For GPT-5.4-mini, those skills recovered 55%–100%+ of the gap between non-reasoning and reasoning modes across 4 agent benchmarks, while using 2.9–4.5× fewer output tokens than reasoning.On ALFWorld and τ²-retail, the skilled non-reasoning model actually beat the reasoning mode.The useful part is that the distiller did not need reasoning traces: skills built only from cheap non-reasoning rollouts were competitive across all 4 domains.The limit is equally useful: reasoning still won on telecom and SpreadsheetBench, where each task contains more instance-specific dependencies that a fixed skill cannot capture.So the practical split is: distill repeated procedures once, then reserve expensive test-time reasoning for the tasks that genuinely need fresh search.译Microsoft 论文提出把昂贵测试时推理摊销为蒸馏技能：收集 35-50 条历史轨迹，由编码智能体提取重复失败模式并编译成 markdown 技能加入非推理模型 system prompt。

智能体 Microsoft 推理数据/训练

9月3日9月3日周四

星期四 · 1 条

14:29

HuggingFace Daily Papers（社区热门论文）

AI 评分 51/100

Microsoft Research 发布 VibeVoice-ASR-Streaming 流式说话人归属语音识别技术报告

Microsoft Research 发布 VibeVoice-ASR-Streaming 技术报告，提出基于 LLM 的端到端流式说话人归属 ASR，将音频块、0.5 秒 lookahead 和已有文本交错在单一自回归上下文中，无需独立 diarization 阶段即可随语音到达输出谁说了什么。

智能体 Microsoft 论文/研究语音

9月2日9月2日周三

星期三 · 1 条

16:38

Rohan Paul@rohanpaul_ai

AI 评分 47/100

Microsoft 论文：滑动窗口注意力在低内存推理上胜过线性注意力改造So happy to see this new Microsoft paper.If lower inference memory is the goal, this paper finds training-free Sliding Window Attention beats most retrofitted linear-attention methods, making it the simpler default to try first.Keep only a small recent window, plus the first 4 “sink” tokens that models rely on.With a 64-token window, this training-free setup had the best average downstream score in 9 of 11 model comparisons and recovered 99.0% of the full-attention baseline average.Many linear-attention alternatives need additional post-training; this version of SWA needs none.The gap grew on long-context reasoning.At 4K context, SWA reached 17.2%–23.0% on the Needle-in-a-Haystack tasks, while LoLCATs reached at most 5.8%; on BABILong, SWA scored 15% versus 3%.In their speed and memory test, the 64-token SWA setup was fastest and used the least memory.Full attention still wins badly on long context, but for fixed, low memory without retraining, the paper recommends trying SWA with attention sinks first.– arxiv. org/abs/2608.28444Title: "Sliding-window beats linear attention"译Microsoft 论文《Sliding-window beats linear attention》发现，目标是降低推理内存时，免训练的滑动窗口注意力（SWA）优于多数改造的线性注意力方法，只需保留小窗口加前 4 个 sink token。

Microsoft 推理论文/研究部署/工程

8月31日8月31日周一

星期一 · 1 条

03:44

DAIR.AI@dair_ai

AI 评分 44/100

TailSFT：微软提出过滤式SFT提升RL性能Interesting technical work from Microsoft.Provides a better understanding on SFT and how to leverage it better for RL.Microsoft researchers asked whether a standard SFT pipeline actually produces the model you want to run RL on.Their answer is no.Standard SFT keeps spending gradient on sequences the model has already fit, which narrows the distribution RL later needs to explore.TailSFT filters those sequences out during training and concentrates learning on the under-modeled tail of the data.That is the only modification they implement.Results:On OLMo-3 7B, pass@16 improves by up to 16.8 points absolute on coding and 3.1 on math. Those higher-coverage checkpoints then lift final pass@1 after GRPO by up to 3.9 points, and in some settings early reward climbs 2.5x faster than the matched standard SFT run.Paper: https://arxiv.org/abs/2608.25756Chat with Paper: https://academy.dair.ai/papers/tailsft-filtered-fine-tuning-improves-post-training-performance-2608.25756译微软研究指出标准SFT会持续在模型已拟合的序列上消耗梯度，收窄了后续RL所需的探索分布。TailSFT在训练中过滤这些序列，专注学习数据中欠拟合的尾部。在OLMo-3 7B上，pass@16在编程任务上最高提升16.8个绝对点，数学提升3.1点；经GRPO后最终pass@1最高提升3.9点，部分场景早期奖励提升速度快2.5倍。

Microsoft 推理数据/训练论文/研究

8月25日8月25日周二

星期二 · 3 条

21:48

elvis@omarsar0

AI 评分 36/100

AutoSaddler：自动优化智能体框架的微软新论文Impressive new paper from Microsoft and colleagues.Harness design is still hand-tuned almost everywhere. This work present an automated loop to optimize the harness.They introduce AutoSaddler, which treats the agent harness as code and learns to patch it offline from failure traces.It runs mini batches of tasks, diagnoses what broke, generates structured patches to prompts, tool configurations, and control logic, then keeps an update only if it survives validation.Gains of 9.0 points on GAIA2, 9.6 on SWE-Bench Pro, and 10.0 on Terminal-Bench 2.0 over the corresponding base harnesses.Deep debugging beats shallow reflection, targeted edits beat unconstrained editing, and generalization-aware selection beats repairing the one trajectory in front of you.Paper: https://arxiv.org/abs/2608.23041Track more trending AI papers in our academy: https://academy.dair.ai/译微软等机构提出 AutoSaddler，将智能体框架视为代码，通过失败轨迹离线自动生成补丁，优化提示词、工具配置和控制逻辑。在 GAIA2、SWE-Bench Pro、Terminal-Bench 2.0 上分别提升 9.0、9.6、10.0 分。研究显示深度调试优于浅层反思，定向编辑优于无约束编辑。

智能体 Microsoft 论文/研究部署/工程

08:00

PromptArmor：Threat Intelligence精选

AI 评分 68/100

PromptArmor 披露 Microsoft Copilot Cowork 沙箱绕过漏洞，攻击者可远程控制智能体窃取数据

PromptArmor 披露 Microsoft Copilot Cowork 一个沙箱绕过漏洞，恶意 Skill 代码可借助沙箱外文件同步服务发起网络请求，与攻击者服务器建立命令控制循环，窃取 Outlook、SharePoint、Teams、插件数据和聊天记录，且用户点击停止按钮也无法中断。

智能体 Microsoft 安全/对齐

推荐理由：原文完整披露了沙箱绕过的技术原理、攻击链和时间线，安全团队可以据此评估 Copilot Cowork 的 Skill 引入风险。

02:57

Hacker News 热门（buzzing.cc 中文翻译）

AI 评分 58/100

逆向工程发现 MS Paint 和"照片"为本地 AI 生成图片添加不可见 GUID 水印

逆向工程显示，MS Paint 和“照片”应用会将服务器下发的 GUID 以不可见方式嵌入本地生成的 AI 图像像素中。该水印由服务器签发，用于追踪图片来源，但用户无法直接察觉其存在。

Microsoft 图像生成安全/对齐

8月23日8月23日周日

星期日 · 3 条

23:48

Rohan Paul@rohanpaul_ai

AI 评分 36/100

微软 SocialRL 让 4B 模型谈判胜过 GPT-4.1New Microsoft paper shows an AI agent that negotiates for you will usually lose, because it was trained to be agreeable.Politeness, transparency and eagerness to close are great in a chat assistant and terrible in a delegate. The paper found frontier models leaking their user's budget and folding the moment a seller pushed back.Their fix is SocialRL: instead of prompting the model to negotiate better, train it on the outcome of the deal across six bargaining and scheduling games.It works, and it doesn't take a big model. A 4B model started anchoring low, holding its position and walking away from bad deals, and landed at 0.627 average across all six games, matching GPT-4.1 at 0.625.The catch is that prompting alone made things worse, so this is a training fix, not a prompt fix.So if you're building an agent that acts on someone's behalf, stop scoring it on whether the deal closed and start scoring it on what it gave away.– arxiv. org/abs/2608.13787Title: "From Passive Delegates to Strategic Negotiators: Reinforcing Social Reasoning in Small Language Models with SocialRL"译微软新论文发现，AI 智能体因被训练得过于顺从，替用户谈判时通常会吃亏，甚至会泄露预算并在对方施压时让步。为此提出 SocialRL 训练方法，让模型在六个谈判与调度游戏中以交易结果为导向学习。一个 4B 模型在全部游戏中平均得分 0.627，追平 GPT-4.1 的 0.625，但仅靠提示词反而会恶化表现。

智能体 Microsoft 数据/训练论文/研究

20:48

Rohan Paul@rohanpaul_ai

AI 评分 32/100

微软 ThinkingBox：智能体可靠性评测沙盒Very relevant Microsoft paper on agent reliability.Succeeding once and being reliable are not the same thing.The best agent solved 91% of business tasks at least once but only 25% every time, so measure repeats.The failures are also hard to spot from the outside.4 out of 5 failed runs ended politely and called a tool that writes to the database.The agent said the job was done, but the records said otherwise.So Microsoft built ThinkingBox, a sandbox that runs an agent against real tools, a simulated customer, and a live backend, then checks the database afterward instead of reading the reply.---– arxiv. org/abs/2608.19741Title: "One Success Isn't Reliability: Thinkingbox, a Sandbox and Benchmark for Agents in Stateful Business Workflows"译微软发布论文指出，智能体"成功一次"不等于"可靠"：最佳智能体在业务任务中单次成功率 91%，但每次都能成功的比例仅 25%，因此需以重复测试衡量可靠性。研究发现 4/5 的失败运行会礼貌地调用写数据库工具并声称任务完成，但实际记录并未更新。为此微软构建 ThinkingBox 沙盒，让智能体在真实工具、模拟客户和实时后端上运行，事后检查数据库而非读取回复。

智能体 arXiv Microsoft 论文/研究

01:44

DAIR.AI@dair_ai

AI 评分 47/100

微软发布Thinkingbox：智能体可靠性基准Banger paper from Microsoft.It's on agent reliability in real business workflows.(bookmark it)Thinkingbox is a sandbox with isolated MCP-compatible tool sessions, plus a benchmark of 507 policy-conditioned workflows across retail, hospitality, auto insurance, neobank IT, and consulting support.Every attempt is graded on the backend state the agent leaves behind. Executable checks accept valid trajectories and reject wrong, missing, or extra effects, so collateral damage counts against you.The strongest model reaches 65.36% pass@1 and 25.25% pass^20.Many failed trials terminate cleanly with valid state-changing tool calls. Watching the response or the tool call tells you very little about whether the task actually completed.Paper: https://arxiv.org/abs/2608.19741Track more trending AI papers in our academy: https://academy.dair.ai/译微软发布Thinkingbox，一个用于真实业务流程中智能体可靠性的沙盒环境，配备MCP兼容工具会话，以及涵盖零售、酒店、汽车保险等领域的507个策略条件工作流基准。每项尝试根据智能体留下的后端状态评分，最强模型pass@1达65.36%，pass^20为25.25%。许多失败尝试以有效的状态变更工具调用干净终止，仅观察响应或工具调用难以判断任务是否真正完成。

智能体 Microsoft 论文/研究评测/基准

8月22日8月22日周六

星期六 · 1 条

21:48

Rohan Paul@rohanpaul_ai

AI 评分 37/100

Agent Lightning v1.0：微软提出在真实部署环境中训练智能体New Microsoft paper says agent training should happen inside the agent's normal operating setup. Otherwise, you may be training something different from what you actually deploy."In traditional agentic RL, the training engine owns the environment interaction loop. In harnessed agentic RL, the harness owns this loop, while the training engine observes only a sequence of LLM request-response pairs."Microsoft’s solution is Agent Lightning v1.0, a lightweight framework for harnessed agentic RL that trains an existing agent through its real deployment harness, instead of rebuilding the agent loop inside the RL trainer.Agent Lightning v1.0 sits between the agent harness and the model, recording the harness’s LLM calls and feeding them into the RL trainer without taking over the harness itself.So it makes RL work correctly with arbitrary deployment-time harnesses, including the messy cases where 1 rollout becomes multiple training samples.– arxiv. org/abs/2608.17528Title: "Agent Lightning v1.0: Towards Harnessed Agentic RL"译微软新论文主张智能体训练应在实际部署环境中进行，否则训练出的模型与上线版本不一致。为此推出 Agent Lightning v1.0，一个轻量级受控智能体强化学习框架，通过真实部署的 harness 训练现有智能体，而非在 RL 训练器中重建智能体循环。

智能体 Microsoft 数据/训练论文/研究

8月20日8月20日周四

星期四 · 1 条

08:00

HuggingFace Daily Papers（社区热门论文）

AI 评分 55/100

Thinkingbox：面向有状态业务流程智能体的沙箱与基准测试

Thinkingbox 是一个工具-智能体-用户交互沙箱，提供隔离的 MCP 兼容工具会话、完整执行轨迹及终端后端状态结果评估。基于该沙箱构建的 Thinkingbox-bench 包含 507 个策略约束工作流，覆盖零售、酒店、汽车保险等场景；最强模型 pass@1 达 65.36%，但 pass^20 仅 25.25%，显示响应或工具调用级信号无法可靠预测端到端任务完成。

智能体 MCP/工具 Microsoft 论文/研究

8月19日8月19日周三

星期三 · 1 条

22:18

elvis@omarsar0

AI 评分 36/100

微软 Agent Lightning v1.0 让 Qwen3.5-9B 编程能力提升 14.6%Very interesting new work from Microsoft.(bookmark it)This work is related to this emerging theme of leveraging harnesses for model post-training.Modern agents run inside a harness that owns tools, context, and control flow. When you train them, the harness owns the environment loop and the trainer only sees LLM request and response pairs.How it works.Agent Lightning v1.0 connects any harness to RL through an endpoint proxy in about 3,500 lines, then works through what breaks in that setup, retokenization, sample merging, advantage calculation, loss normalization, and backend scheduling.Using 6K training examples and modest compute, it moves Qwen3.5-9B on SWE-bench Verified from 41.8% to 56.4%.Paper: https://arxiv.org/abs/2608.17528Track more trending AI papers in our academy: https://academy.dair.ai/译微软新研究 Agent Lightning v1.0 通过约 3500 行端点代理将任意 harness 接入强化学习，用于模型后训练。仅用 6K 训练样本和适度算力，将 Qwen3.5-9B 在 SWE-bench Verified 上从 41.8% 提升至 56.4%。论文已发布。

Microsoft 数据/训练论文/研究

8月17日8月17日周一

星期一 · 1 条

23:50

DAIR.AI@dair_ai

AI 评分 47/100

微软4B模型谈判能力超越GPT-5系列Very interesting new work from Microsoft Research.(bookmark it)They show that a 4B model can be tuned to out-negotiate the GPT-5 family of models.The dispositions that make an assistant pleasant make it a poor delegate. They show that a friendly frontier model volunteers its principal's private information and concedes at the first sign of resistance.SocialRL trains social reasoning directly in a 4B model across six principal-driven domains including negotiation, job interviews and marketplace haggling. After training, 78% of buyer openings anchor below target against 3% untrained.Cascade RL and multi-teacher distillation consolidate the specialists into one 4B at 0.627 average utility, above GPT-5.1 at 0.619 and GPT-5.2 at 0.613.Paper: https://arxiv.org/abs/2608.13787Track more trending AI papers in our academy: https://academy.dair.ai/译微软研究院新研究表明，4B模型经SocialRL训练后可在谈判中胜过GPT-5系列模型。训练后78%的买家开场报价低于目标价，而未训练模型仅3%。Cascade RL与多教师蒸馏将专家整合为单一4B模型，平均效用0.627，高于GPT-5.1的0.619和GPT-5.2的0.613。

智能体 Microsoft OpenAI 安全/对齐

8月16日8月16日周日

星期日 · 1 条

00:23

Rohan Paul@rohanpaul_ai

AI 评分 36/100

微软：用蒸馏技能替代昂贵测试时推理，一次付清推理成本What if you could pay the reasoning cost once, then reuse what the model learned across future tasks?New Microsoft paper finds that some expensive test-time reasoning can be replaced with a small set of rules learned from previous agent runs.The paper tests a cheaper alternative: collect 35–50 past trajectories, have a coding agent extract recurring failure patterns, then turn those patterns into a small markdown skill added to the non-reasoning model’s system prompt.For GPT-5.4-mini, those skills recovered 55%–100%+ of the gap between non-reasoning and reasoning modes across 4 agent benchmarks, while using 2.9–4.5× fewer output tokens than reasoning.On ALFWorld and τ²-retail, the skilled non-reasoning model actually beat the reasoning mode.The useful part is that the distiller did not need reasoning traces: skills built only from cheap non-reasoning rollouts were competitive across all 4 domains.The limit is equally useful: reasoning still won on telecom and SpreadsheetBench, where each task contains more instance-specific dependencies that a fixed skill cannot capture.So the practical split is: distill repeated procedures once, then reserve expensive test-time reasoning for the tasks that genuinely need fresh search.– arxiv. org/abs/2608.07885Title: "Reason Wide, Not Deep: Amortizing the Reasoning Premium into Distilled Skills"译微软新论文提出，可将昂贵的测试时推理替换为从过往智能体运行中提取的小型规则集。用 GPT-5.4-mini 提取 35-50 条轨迹中的失败模式并转为 markdown 技能，在 4 个智能体基准上恢复了非推理与推理模式间 55%-100%+ 的差距，且输出 token 减少 2.9-4.5 倍。

智能体 Microsoft 推理论文/研究

8月13日8月13日周四

星期四 · 3 条

23:48

elvis@omarsar0

AI 评分 37/100

微软新研究：坏技能库如何拖垮智能体Very interesting new paper from Microsoft and colleagues.(bookmark it)Skill libraries are used in every major harness on the assumption that more guidance is free. This work measures what a bad skill actually costs you.They attribute 307 agent failures to specific loaded skills, 125 functional failures and 182 efficiency regressions, by comparing each skill-guided run against a matched reference run that solves the same task.The failures rarely come from irrelevant skills. Seemingly relevant skills push the agent to incorrectly implement or omit something the task required.Cost regressions are not explained by prompt length either. The largest source is excessive verification at 67 cases, followed by heavy implementation pipelines at 30 cases. It turns out that skills quietly turn validation checklists into mandatory work.Paper: https://arxiv.org/abs/2608.11888Track more trending AI papers in our academy: https://academy.dair.ai/译微软与同事的新论文量化了技能库（skill libraries）对AI智能体的负面影响：307次智能体失败被归因于特定技能，其中125次功能失败、182次效率回退。看似相关的技能反而会诱导智能体错误执行或遗漏任务要求，而效率回退的主因并非提示词长度，而是过度验证（67例）和繁重实现流程（30例）。论文：https://arxiv.org/abs/2608.11888

智能体 Microsoft 论文/研究

05:00

DAIR.AI@dair_ai

AI 评分 31/100

微软研究：多语言智能体评测应聚焦行动轨迹// Actions Speak Louder Than Words //Multilingual agent evaluation compares final answers and throws the trajectory away. The trajectory fixes cost, latency, failure mode, and auditability.New research from Microsoft Research makes the action policy the measured object. 8 models, 6 parallel benchmarks, 41 languages, 2.38M rollouts.Five confounds sit between raw trace similarity and any defensible claim. Short traces score higher, empty traces score perfectly, unrelated traces agree by chance over half the time, reproducibility caps the gap, and one model asked twice in one language answers differently. Removing all five made the effect larger.Normalised by their own reproducibility, four frontier models each keep 71 to 73% of their action policy across languages.Paper: https://arxiv.org/abs/2608.11110Track more trending AI papers in our academy: https://academy.dair.ai/译微软研究院新研究将行动策略作为评测对象，覆盖8个模型、6个并行基准、41种语言、238万次 rollout。研究发现原始轨迹相似度存在五个混淆因素：短轨迹得分更高、空轨迹满分、无关轨迹半数以上偶然一致、可复现性限制差距、同一模型同一语言两次回答不同。剔除全部混淆后效应反而更大；按自身可复现性归一化，四个前沿模型跨语言保留71%至73%的行动策略。

智能体 Microsoft 论文/研究

00:17

Microsoft Research@MSFTResearch

AI 评分 30/100

MindTopo新基准测试AI拓扑理解A path, a fence, a knot. MindTopo sets a new benchmark for testing how AI understands topological relationships and highlights new opportunities to strengthen spatial reasoning and planning. https://msft.it/6011aH02S译一条路径、一道围栏、一个绳结。MindTopo 为测试 AI 如何理解拓扑关系设立了新基准，并凸显了加强空间推理与规划的新机遇。https://msft.it/6011aH02S

Microsoft 推理评测/基准

8月12日8月12日周三

星期三 · 4 条

08:00

HuggingFace Daily Papers（社区热门论文）

AI 评分 54/100

Dion3：全栈正交更新的 Muon 优化器改进版

Dion3 针对 Muon 优化器中立方时间 Newton-Schulz 正交化步骤的高开销，在堆栈各层进行优化：Gram Newton-Schulz 算法降低正交化 FLOP 成本，CuteDSL 内核利用对称性加速，megabatching 策略减少通信开销。

arXiv Microsoft 数据/训练论文/研究

03:46

DAIR.AI@dair_ai

AI 评分 36/100

微软新论文：技能蒸馏让非推理模型追平推理模式Impressive new paper from Microsoft.(bookmark it)Reasoning modes beat non-reasoning modes on multi-step agentic tasks, and they charge a 3 to 6x premium in output tokens on every single episode. Much of that spend goes into re-deriving procedures the model already worked out on earlier episodes in the same domain.This work amortizes that cost.A coding agent reads a small corpus of existing trajectories from a training split, compiles a compact natural-language skill, and injects it into the non-reasoning model's system prompt.Across ALFWorld, tau-squared-bench telecom and retail, and SpreadsheetBench-Verified, skills recover 55% to over 100% of the reasoning gap for GPT-5.4-mini on held-out tasks. On two of the four benchmarks the skill-equipped non-reasoning model beats reasoning mode outright, while emitting 2.7 to 6x fewer output tokens and zero reasoning tokens.Reasoning traces turn out to be optional. Skills distilled from non-reasoning trajectories alone stay competitive with skills distilled from paired corpora.Paper: https://arxiv.org/abs/2608.07885Track more trending AI papers in our academy: https://academy.dair.ai/译微软新论文提出用技能蒸馏摊薄推理成本：编码智能体从训练轨迹中提取简洁自然语言技能，注入非推理模型的系统提示词。

智能体 Microsoft 推理论文/研究

00:17

Microsoft Research@MSFTResearch

AI 评分 28/100

CARE-X：微软统一方法革新胸片解读Radiology AI is evolving beyond report generation. CARE-X explores a unified approach that combines flexible reasoning, calibrated predictions, and measurement-based tools for chest X-ray interpretation. https://msft.it/6015aHrqH译放射学 AI 正超越报告生成。CARE-X 探索一种统一方法，将灵活推理、校准预测与基于测量的工具相结合，用于胸部 X 光片解读。https://msft.it/6015aHrqH

Microsoft 多模态推理论文/研究

00:00

Microsoft Research 博客（RSS）

AI 评分 59/100

Microsoft Research 发布 CARE-X 胸部X光视觉语言模型，结合辅助监督与工具增强测量

Microsoft Research 与 Narayana Health 等合作发布研究模型 CARE-X，一个统一胸部X光 VLM，基于 SigLIP2-so400M 视觉编码器和 Phi-4-mini-instruct（3.8B）。

Microsoft 多模态论文/研究

8月9日8月9日周日

星期日 · 2 条

15:44

Rohan Paul@rohanpaul_ai

AI 评分 50/100

微软论文：编码智能体不应按聊天请求调度New Microsoft Paper on GitHub Copilot’s production traces show why coding agents should not be served like chat requests.In 13.5M GitHub Copilot sessions, 87% of LLM calls came from the agent itself rather than a user.A user prompt can fan out into an autonomous chain of model calls, tool actions, retries, and growing context, making the turn or session a more useful scheduling unit than an isolated request.That structure is especially visible in the KV cache.It shows that KV cache is not really a request-level resource; its value depends on where the agent is in the workflow.Within a turn, average cache hit rate rises from about 45% on the first LLM call to 92–94% from the third call onward.At a same-model turn boundary it falls to 55%, while a model switch pushes it down to 8%.Median KV-cache idle time is 1.2 seconds within a turn versus 172 seconds across turns, while container idle time jumps from 5.8 to 243 seconds.Using turn- and session-level features, the paper’s lightweight predictor captures 86–90% of total idle time, giving the serving stack a signal for cache offloading or container reclamation.The implication is straightforward: coding-agent infrastructure should schedule workflow state across turns, because request-level policies discard some of the strongest signals in the workload.– arxiv. org/abs/2608.00101Title: "Agentic Coding in the Wild: Characterizing GitHub Copilot Traces at Production Scale"译微软对 13.5M 次 GitHub Copilot 会话的生产轨迹分析显示，87% 的 LLM 调用来自智能体自身而非用户。KV 缓存命中率在单轮内从首次调用的约 45% 升至第三次起的 92-94%，模型切换时骤降至 8%。论文提出基于轮次和会话级特征的轻量预测器，可捕获 86-90% 的总空闲时间，表明编码智能体基础设施应按轮次调度工作流状态，而非请求级策略。

智能体 arXiv Microsoft 论文/研究

08:00

HuggingFace Daily Papers（社区热门论文）

AI 评分 55/100

全带宽 Transformer：用潜在反馈拓宽自回归解码通道

约翰霍普金斯大学、普林斯顿大学与微软提出全带宽 Transformer，通过门控线性单元将上一解码步的顶层隐藏状态与采样 token 嵌入融合后作为下一输入，使非语言化计算能以完整深度预算重新进入堆栈。

Microsoft 推理数据/训练论文/研究

8月6日8月6日周四

星期四 · 1 条

08:39

MarkTechPost（RSS）精选

AI 评分 70/100

Microsoft 的 SkillOpt 证明优化后的智能体技能工件可在不同模型规模及 Codex 与 Claude Code 之间迁移

Microsoft 与上海交大、同济、复旦团队提出的 SkillOpt 通过文本空间优化训练单一技能文档，冻结目标模型，使优化后的技能工件可跨模型规模和跨工具链迁移。在 Codex 上优化的 SpreadsheetBench 技能部署到 Claude Code 后得分 81.8，超过后者自行训练技能得到的 80.4。全部 4 项跨模型、4 项跨工具链和 3 项跨基准迁移结果均高于目标的无技能基线。

智能体 Microsoft 论文/研究

推荐理由：跨 harness 迁移实验把 agent 技能从单环境优化推进到可复用文本技能文件，Codex 练出的技能在 Claude Code 上超过域内训练结果，对维持多工具一致行为提供直接实证。

8月4日8月4日周二

星期二 · 1 条

06:03

Microsoft Research@MSFTResearch

AI 评分 42/100

微软研究：小模型谈判与非洲语言评测新进展Small language models learn to negotiate with SocialRL, PazaBench V2 expands speech AI evaluation across African languages, and EvoLib helps agents turn experience into knowledge. Plus: new methods for more reliable A/B testing and advances in AI-driven precision oncology. https://msft.it/6012a8hjI译小型语言模型借助 SocialRL 学会谈判，PazaBench V2 将语音 AI 评测扩展至非洲多语言，EvoLib 帮助智能体将经验转化为知识。另有：更可靠的 A/B 测试新方法，以及 AI 驱动的精准肿瘤学进展。https://msft.it/6012a8hjI

智能体 Microsoft 论文/研究语音

8月1日8月1日周六

星期六 · 1 条

00:57

elvis@omarsar0

AI 评分 36/100

微软Echoverse：规模化训练计算机操作智能体New research from Microsoft.This one is on training computer-use agents at scale.Recent pipelines generate synthetic environments in bulk, which moved the bottleneck from how many exist to what is inside each one. Echoverse compiles specifications into stateful applications whose tasks are graded against the application's own database, then runs a co-evolution loop that reads every graded rollout twice. Once as repairs to the environment, its tasks and its verifier. Once as training signal for the model.On the same domains, shallow environments pushed live-site accuracy below the base model, from 80.0 down to 75.0. Deep ones raised it, 80.0 to 85.0 and 48.0 to 65.0.Repairing a single environment lifted the model trained on it from 16.2% to 38.5%. Across twelve environments, a 9B model went from 36.5% to 67.1% on fourteen evaluation splits, within fourteen points of the much larger frontier model that taught it.They release four environments as a benchmark with applications, seed data and grounded graders.Paper: https://arxiv.org/abs/2607.28074Track trending AI papers in our academy: https://academy.dair.ai/译微软提出Echoverse，通过将规格编译为有状态应用并基于应用数据库评分任务，配合共进化循环（读取评分轨迹两次，分别用于修复环境和生成训练信号），规模化训练计算机操作智能体。

智能体 Microsoft 论文/研究

7月31日7月31日周五

星期五 · 4 条

10:52

HuggingFace Daily Papers（社区热门论文）

AI 评分 53/100

Echoverse：为大规模训练计算机操作智能体而构建的深度演化环境

Echoverse 将环境、任务与验证器绑定为“世界”，通过共演化循环让每次分级运行同时用于修复环境和训练模型。在十二个环境中训练的 9B 模型在十四个评估分项上提升显著，与指导它的更大前沿模型差距在十四分以内。同一世界可直接用作强化学习环境，结合接地验证器与密集逐步评判器的奖励将留出集评分从提升至；研究还发布了四个环境作为基准，含应用、种子数据与评分器。

智能体 Microsoft 开源生态论文/研究

01:27

Microsoft Research@MSFTResearch

AI 评分 43/100

微软Echoverse训练AI智能体应对多步工作流Computer-use AI agents struggle with multi-step workflows like email and customer support. Echoverse trains agents in realistic environments rather than simply providing more training tasks, helping them improve as the tasks, tests, and environments evolve. https://msft.it/6012a6oLE译计算机使用的AI智能体在电子邮件和客户支持等多步骤工作流中表现不佳。Echoverse在逼真环境中训练智能体，而非简单提供更多训练任务，帮助它们随着任务、测试和环境的演变而提升。https://msft.it/6012a6oLE

智能体 Microsoft 论文/研究

01:00

Microsoft Research 博客（RSS）精选

AI 评分 60/100

Microsoft Research 发布 Echoverse：面向 computer-use agent 的深度可演化训练环境

Microsoft Research 推出 Echoverse，为 computer-use agent 构建了 12 个合成训练世界（10 个深度领域世界和 2 个能力世界），9B 模型训练后在全部环境上平均分从 36.5% 升至 67.1%，与 GPT-5.4（80.7%）相差 14 个点。

智能体 Microsoft 开源生态数据/训练

推荐理由：原文给出深度世界对比浅层世界的实验数据和可复用的环境构建方法，并开源了四个世界，读者可据此评估训练环境的做法。

00:27

Microsoft Research@MSFTResearch

AI 评分 31/100

EvoLib：让LLM经验转化为可进化知识LLMs do not get smarter just by remembering more. EvoLib turns experience into evolving knowledge, taking reusable skills and insights that help models learn and adapt across tasks long after deployment. https://msft.it/6019vh35f译LLM 并不会仅仅因为记住更多就变得更聪明。EvoLib 将经验转化为不断进化的知识，提取可复用的技能与洞见，帮助模型在部署后长期跨任务学习与适应。

Microsoft 数据/训练论文/研究

7月30日7月30日周四

星期四 · 1 条

07:37

Microsoft AI：官方博客（网页）

AI 评分 9/100

Optimizing the frontier performance curve

Microsoft 论文/研究部署/工程

7月21日7月21日周二

星期二 · 3 条

10:49

HuggingFace Daily Papers（社区热门论文）

AI 评分 59/100

LLM-as-a-Coach：面向不可验证任务的经验学习框架

微软团队提出 Experiential Learning（EL）框架，将 LLM-as-a-Judge 的评估模型改造为 LLM-as-a-Coach，把对每个 on-policy 响应的文本反馈蒸馏为可迁移的经验知识，并通过 on-policy context distillation 内化到策略参数中。

arXiv Microsoft 数据/训练论文/研究

08:00

HuggingFace Daily Papers（社区热门论文）

AI 评分 50/100

FinanceComplexQA：面向工业级金融文档的智能体推理评测基准

微软与中国高校团队推出 FinanceComplexQA，一个针对工业级金融文档的开放式生成评测基准，包含 2,026 项深度研究任务，覆盖 1009 份真实金融文档。该基准支持双语、六大场景与七类任务，采用 Agent-as-a-Judge 多指标评估。团队还提出 Finance-LaTeX SKILL 多智能体框架，可自动合成 2,000 份金融文档与 6,000 个问答对，用于扩展基准。

Microsoft 检索增强推理数据/训练

00:19

Microsoft Research@MSFTResearch

AI 评分 29/100

微软发布Aurora 1.5等多项AI进展Aurora 1.5 scales open forecasting; Flint redefines visualization; FlowDAgger accelerates adaptation. Plus: AI agents crack Rowhammer defenses and memory reshapes conversational AI. https://msft.it/6019v4m9F译Aurora 1.5 扩展开放预测；Flint 重新定义可视化；FlowDAgger 加速自适应。此外：AI 智能体破解 Rowhammer 防御，记忆重塑对话式 AI。

Microsoft 安全/对齐数据/训练论文/研究

全部 AI 动态

9月16日9月16日周三

9月11日9月11日周五

9月10日9月10日周四

9月9日9月9日周三

9月7日9月7日周一

9月3日9月3日周四

9月2日9月2日周三

8月31日8月31日周一

8月25日8月25日周二

8月23日8月23日周日

8月22日8月22日周六

8月20日8月20日周四

8月19日8月19日周三

8月17日8月17日周一

8月16日8月16日周日

8月13日8月13日周四

8月12日8月12日周三

8月9日8月9日周日

8月6日8月6日周四

8月4日8月4日周二

8月1日8月1日周六

7月31日7月31日周五

7月30日7月30日周四

7月21日7月21日周二