Microsoft Research Focus – Edition #97

Microsoft Research Focus – Edition #97 — Researchers across the Microsoft research community

Researchers across the Microsoft research community

AI-assisted summary of Microsoft Research Focus – Edition #97 — original source. Check the original for full context.

Source published:

Key ideas

  • FolkPilot introduces a dynamic backtracking framework for AI agents, matching state-of-the-art performance while reducing processing resource demands by up to 59% across six benchmarks. (0:00)
  • Diagnostic tests with the Cabra framework reveal that while coding agents use external tools to mask code modification weaknesses, comprehension of unfamiliar code remains their primary failure mode. (0:58)
  • Teletune allows computer-use agents to infer underlying human intent and extract reusable workflows directly from raw, unlabeled click and keystroke activity logs. (1:42)
  • Probabilistic Bias Correction (PBC) won a global meteorological modeling competition and is currently deployed in Kenya, providing crucial 2- to 6-week subseasonal monsoon forecasts for smallholder farmers. (4:48)
  • Microsoft Research announced the public preview of BioEmu2 on GitHub, a foundation model designed to simulate the dynamic conformational fluctuations and equilibrium states of proteins for drug discovery. (5:42)

Detailed notes

## Background Published on October 29, 2026, by researchers across the Microsoft research community on the Microsoft Research Podcast platform, this audio brief is hosted by Sowbhagya Suresh. According to the episode description, Edition #97 highlights the latest research spanning the community, addressing challenges facing coding agents, enterprise and workplace AI adoption patterns, extended-range weather forecasting, and molecular biology.

## TL;DR Host Sowbhagya Suresh examines cutting-edge AI breakthroughs and empirical field studies emerging from Microsoft Research. The update covers agent frameworks like FolkPilot for efficient backtracking, Cabra for isolating code comprehension bottlenecks, and Teletune for learning computer workflows from user logs. It also explores workplace dynamics through enterprise Copilot usage telemetry and the real-world deployment of Shiksha Copilot among teachers in India. Finally, the briefing highlights scientific AI advances, including Probabilistic Bias Correction for extended-range monsoon forecasting in Kenya and BioEmu2 for modeling dynamic biomolecular conformational states.

## Atmosphere & style The presentation feels brisk, polished, and distinctly journalistic, delivering high-density technical findings with clarity and precision. Suresh maintains a steady, authoritative, yet engaging host cadence, avoiding hype in favor of clear experimental outcomes, benchmarks, and real-world implications. The audio has the crisp, focused rhythm of an institutional research digest, balancing foundational software engineering insights with grounded, human-centric deployment stories.

## Bulletin - `00:00` FolkPilot introduces a dynamic backtracking framework for AI agents, matching state-of-the-art performance while reducing processing resource demands by up to 59% across six benchmarks. - `00:58` Diagnostic tests with the Cabra framework reveal that while coding agents use external tools to mask code modification weaknesses, comprehension of unfamiliar code remains their primary failure mode. - `01:42` Teletune allows computer-use agents to infer underlying human intent and extract reusable workflows directly from raw, unlabeled click and keystroke activity logs. - `04:48` Probabilistic Bias Correction (PBC) won a global meteorological modeling competition and is currently deployed in Kenya, providing crucial 2- to 6-week subseasonal monsoon forecasts for smallholder farmers. - `05:42` Microsoft Research announced the public preview of BioEmu2 on GitHub, a foundation model designed to simulate the dynamic conformational fluctuations and equilibrium states of proteins for drug discovery.

## Key takeaways - `00:00` Deploying AI agents on multi-step workflows often leads to reliability failures where error recovery requires heavy compute with uncertain outcomes. - `00:00` FolkPilot addresses agent recovery costs by learning when it is computationally worthwhile to backtrack and explore alternative paths rather than restarting from scratch. - `00:00` Across six standard benchmarks, FolkPilot matched leading foundation models while slashing compute demands by up to 59%. - `00:58` The Cabra framework isolates coding agent competencies by synthesizing controlled coding challenges from scratch across varying complexity levels. - `00:58` Coding agents frequently overcome code modification and editing hurdles by leveraging external tools like linters and debuggers, but fail when forced to interpret unfamiliar codebases. - `01:42` User activity logs are an abundant source of procedural knowledge for computer-use agents, but raw click and keystroke telemetry rarely records underlying user intent. - `01:42` Teletune parses raw interaction streams, infers high-level user goals, and extracts structured, reusable workflows across desktop and browser environments. - `02:32` Aggregated enterprise telemetry shows document editing is the single most common Copilot chat activity, accounting for over 20% of interactions. - `02:32` Excluding the top 20 aggregate Copilot activities reveals pronounced operational divergence and domain-specific workflows across different industry sectors. - `03:36` A deployment study of Shiksha Copilot across low-resource schools in India evaluated how AI assistance aids administrative bureaucracy and bilingual curriculum creation. - `03:36` Over 1,000 participating educators reported reduced workplace stress while actively editing and repurposing AI-generated outputs rather than accepting them passively. - `04:48` Predictive accuracy in the 2- to 6-week subseasonal weather window has historically lagged due to systemic drift in physical forecasting models. - `04:48` Probabilistic Bias Correction (PBC) uses machine learning to learn historical forecasting error distributions and adjust future extended-range model projections. - `04:48` PBC won first place in a global extended-range forecasting competition and currently provides operational monsoon predictions in Kenya. - `05:42` BioEmu2, developed by Microsoft Research's AI for Science team, simulates equilibrium distributions and conformational protein dynamics rather than static 3D structures. - `05:42` Simulating dynamic biomolecular interactions in BioEmu2 assists pharmacological research in identifying mechanisms to modulate beneficial or harmful protein behaviors.

## Detailed walkthrough ### `00:00` FolkPilot: Efficient Self-Improving AI Agents

Host Sowbhagya Suresh introduces edition #97 of *Microsoft Research Focus*, highlighting new research across the Microsoft community on topics ranging from coding agents and workplace AI adoption to extended-range weather forecasting and protein biology. The edition opens with the challenge of deploying AI agents for complex, multi-step tasks that demand sequential decisions and interactions with external tools. While these agents can handle intricate workflows, their reliability falters when things go wrong; recovering from errors often requires substantial compute resources without any guarantee of achieving a better outcome.

To address this dilemma, researchers developed FolkPilot, a self-improving agent framework that learns when it is computationally worthwhile to backtrack and revisit earlier decisions. Rather than blindly restarting or brute-forcing alternatives, FolkPilot evaluates previous choices to identify different paths and dynamically adapts its problem-solving strategy as new outcomes emerge. Evaluated across six benchmarks alongside leading AI models, the framework matched the performance of existing state-of-the-art approaches while slashing AI processing demands by up to 59%. Emphasizing the operational significance of these efficiency gains, Suresh notes: “These results suggest that AI agents can solve complex problems more reliably and efficiently, making them more practical and cost-effective to deploy at scale.”

- Evaluated across six distinct benchmarks against leading foundation AI models. - Achieves up to a 59% reduction in AI processing resource demands while maintaining competitive task performance.

---

`00:58` Cabra: Diagnosing Coding Agent Bottlenecks

Coding agents are designed to execute user requests by reasoning over code, modifying files, and running scripts. However, building reliable systems requires a clear understanding of why and where they fail. Current evaluation benchmarks make it difficult to disentangle the exact factors that drive agent success or failure because codebases and task difficulties vary unpredictably. To create controlled diagnostic environments, Microsoft researchers developed Cabra, a framework that synthesizes coding challenges from scratch, allowing investigators to systematically dial up complexity and isolate the specific capabilities that challenge modern models.

Through tests encompassing thousands of generated challenges, the researchers uncovered an unexpected pattern: agents frequently overcome technical coding hurdles simply by utilizing external tools, such as linters, interpreters, or debuggers. However, tool usage does not solve comprehension issues. When tasked with navigating and interpreting unfamiliar codebases, agent performance drops significantly. Suresh highlights the primary takeaway from the Cabra study: “The findings suggest that understanding existing code rather than modifying it remains one of the biggest obstacles to developing more effective coding agents.”

- Generates thousands of synthetic, controlled coding challenges from scratch to benchmark agent competencies. - Reveals that tool usage effectively masks code generation and editing weaknesses, but fails to compensate for comprehension gaps in unfamiliar code.

---

`01:42` Teletune: Teaching Computer-Use Agents via Activity Logs

Computer-use agents represent a class of AI systems engineered to interact with graphical interfaces by clicking, typing, filling out digital forms, processing administrative requests, and navigating websites. To execute these workflows reliably across diverse software environments, agents must learn human procedural knowledge. User activity logs provide a natural, abundant data source for observing these operations, but they present significant modeling challenges: raw activity logs rarely record the underlying user intent behind discrete clicks and keystrokes, and distinct tasks are frequently mixed together within the same session.

To bridge this gap, researchers introduced Teletune, a framework designed to extract reusable procedural skills and structured workflows directly from raw user logs. Teletune automatically parses interaction streams and infers the high-level task goals governing specific sequences of user actions. Across empirical evaluations spanning standard desktop software and browser-based workflows, Teletune consistently outperformed existing approaches. Highlighting its deployment potential, Suresh explains: “By learning from existing activity, Teletune offers a scalable way to help AI agents adapt to unfamiliar software and carry out complex multi-step workflows.”

```mermaid flowchart TD accTitle: Teletune Agent Workflow Architecture accDescr: Diagram illustrating how Teletune processes raw user logs into inferred goals, structured workflows, and computer agent execution.

subgraph DataIngestion["Input Data"] A["Raw User Activity Logs"] B["Mixed Clicks & Keystrokes"] A --> B end

subgraph TeletuneEngine["Teletune Framework"] C["Intent & Goal Inference"] D["Skill Segmentation"] E["Reusable Workflow Extraction"] B --> C C --> D D --> E end

subgraph AgentExecution["Autonomous Execution"] F["Computer-Use Agent"] G["Multi-Step Browser & Software Tasks"] E --> F F --> G end ```

- Infers implicit user intent and task goals from unlabeled keystroke and click telemetry. - Outperforms existing baselines on benchmarks covering browser navigation and native desktop application tasks.

---

`02:32` Analyzing Copilot Usage: Beyond Surface-Level AI Adoption

While organizations across different industries might appear to adopt generative AI in homogeneous ways on the surface, a deeper look into enterprise telemetry reveals substantial divergence. Microsoft researchers analyzed aggregated Copilot chat activity spanning a diverse spectrum of customer categories. At the macro level, document editing stood out as the single most frequent classified activity, representing more than 20% of all interactions.

However, when researchers filtered out the top 20 most common activities, the uniform surface cracked: distinct usage patterns and specialized workflows between customer categories became substantially more pronounced. High-frequency administrative and writing prompts tend to mask the specialized, domain-specific tasks that differentiate industries. Suresh emphasizes that enterprise leaders must look past generic adoption metrics when designing AI strategies: “Don't assume you have the same AI needs as others in your industry.” She elaborates on this strategic principle, stating: “Finding the right opportunities means looking beyond user statistics to understand what employees actually do, how they work, and what outcomes matter most.”

- Document editing represents over 20% of all classified Copilot chat interactions across enterprise categories. - Stripping away the top 20 aggregate use cases exposes pronounced, industry-specific operational differences.

---

`03:36` Shiksha Copilot: Transforming Teacher Workflows in the Global South

In low-resource educational settings, administrative burdens and bilingual lesson planning often consume valuable instructional time. A new deployment study examined Shiksha Copilot, a human-in-the-loop AI teaching assistant designed to support educators in classrooms across the Global South. Conducted in low-resource schools in India, the study evaluated how digital assistance affects daily preparatory routines. Researchers found that Shiksha Copilot significantly curtailed the time teachers spent managing administrative bureaucracy and creating bilingual lesson plans, freeing up time and energy for interactive classroom activities.

The field deployment included more than 1,000 participating teachers, many of whom reported noticeable reductions in workplace stress and voiced strong positive sentiment regarding AI’s role in their professional routine. Crucially, educators did not simply accept automated outputs as finished products; instead, they maintained active editorial oversight over the generated materials. Suresh observes: “These teachers actively engage with Shiksha Co-Pilot, using their professional expertise to adapt and repurpose materials to soothe their own classrooms.”

- Deployed across low-resource schools in India with more than 1,000 active teacher participants. - Significantly cut hours dedicated to administrative paperwork and bilingual curriculum development. - Reported tangible reductions in teacher stress levels while preserving human-in-the-loop pedagogical agency.

---

`04:48` Probabilistic Bias Correction (PBC) for Extended-Range Weather

Standard 10-day numerical weather forecasts have reached a high level of dependability, but extending predictive accuracy into the multi-week horizon—specifically the two- to six-week subseasonal window—remains a major scientific hurdle. This gap presents serious real-world challenges because critical preventative decisions regarding agricultural planting, wildfire management, and extreme disaster preparedness are made weeks ahead of time.

To overcome the systemic drift and errors inherent in physical models at this range, Microsoft researchers developed Probabilistic Bias Correction (PBC). PBC is a machine learning framework that analyzes a weather model's historical forecasting discrepancies and learns to compensate for its past mistakes. Demonstrating its technical effectiveness, PBC won top honors in a global competition focused on advancing extended-range forecasting. The framework is actively deployed in Kenya, helping deliver critical monsoon forecasts to local farming communities. Suresh frames the broader promise of the tool: “This development could help make weather predictions more useful across a broad range of applications.”

- Targets the difficult subseasonal prediction window beyond standard 10-day forecasts. - Learns historical error distributions from numerical weather models to probabilistically adjust future projections. - Awarded first place in a global competition on extended-range meteorological modeling. - Operational in Kenya to support smallholder farmers relying on accurate monsoon timing.

---

`05:42` BioEmu2: Modeling Biomolecular Dynamics

Understanding the structure, dynamic movements, and biological functions of proteins is critical to advancing molecular biology and pharmacological design. Microsoft Research’s AI for Science team announced the public preview of BioEmu2, a foundation model specifically engineered to simulate biomolecular systems. BioEmu2 builds upon the Biomolecular Emulator architecture, employing deep learning to capture the equilibrium distributions and conformational dynamics of proteins rather than merely providing static three-dimensional snapshots.

Deciphering these dynamic physical interactions has direct implications for drug discovery and therapeutic intervention. By simulating how biomolecules fluctuate and bind, researchers can pinpoint mechanisms to either boost beneficial functions or neutralize harmful interactions. Suresh summarizes the therapeutic horizon of the technology: “By learning more about how proteins function, scientists could move closer to developing more effective medications that enhance beneficial protein activity or prevent proteins from causing harm.” The model is available in public preview via the Microsoft Research BioEmu2 GitHub repository.

- Developed by the Microsoft Research AI for Science team as a successor to Biomolecular Emulator. - Models equilibrium dynamics, conformational flexibility, and functional properties of biomolecules. - Released in public preview on the Microsoft Research GitHub repository for academic and pharmaceutical exploration.

## Notable quotes > `02:32` Finding the right opportunities means looking beyond user statistics to understand what employees actually do, how they work, and what outcomes matter most. — Sowbhagya Suresh

> `02:32` Don't assume you have the same AI needs as others in your industry. — Sowbhagya Suresh

> `00:58` The findings suggest that understanding existing code rather than modifying it remains one of the biggest obstacles to developing more effective coding agents. — Sowbhagya Suresh

## Mentioned Sowbhagya Suresh, Microsoft Research Focus, Microsoft, Microsoft Research, FolkPilot, Cabra, Teletune, Microsoft Copilot, Shiksha Copilot, India, Probabilistic Bias Correction, Kenya, Microsoft Research AI for Science, BioEmu2, Biomolecular Emulator, GitHub

## Video description <p>Host Sowbhagya Suresh shares some of the latest news and cutting-edge research from across Microsoft—from understanding the challenges facing coding agents to uncovering how people use AI at work and improving extended-range weather forecasting.</p><p><a href="https://www.microsoft.com/en-us/research/podcast/microsoft-research-focus-edition-97/">Show notes</a></p><p><a href="https://www.linkedin.com/newsletters/research-focus-7335362495383040001/">Research Focus newsletter on LinkedIn</a></p><p><a href="https://www.microsoft.com/en-us/research/podcast/">Microsoft Research Podcast</a></p>

[Source](https://media.blubrry.com/microsoftresearch/content.blubrry.com/microsoftresearch/MRP-2026-10-30-322.mp3)

中文简报

微软研究院焦点 – 第 97 期 主持人 Sowbhagya Suresh 深入探讨了来自微软研究院的前沿人工智能突破和实证现场研究。本期内容涵盖多个智能体框架，例如用于高效回溯的 FolkPilot、用于隔离代码理解瓶颈的 Cabra，以及用于从用户日志中学习计算机工作流的 Teletune。报告还通过企业 Copilot 使用遥测数据以及 Shiksha Copilot 在印度教师中的实际部署，探讨了职场动态。最后，简报重点介绍了科学人工智能领域的进展，包括用于肯尼亚扩展期季风预测的概率偏差校正（PBC），以及用于模拟动态生物分子构象状态的 BioEmu2。 FolkPilot 为人工智能智能体引入了一种动态回溯框架，在六个基准测试中达到了最先进的性能，同时将处理资源需求降低了多达 59%。 (0:00) 使用 Cabra 框架进行的诊断测试显示，尽管编程智能体利用外部工具来掩盖代码修改方面的不足，但对陌生代码的理解仍然是它们最主要的技术失效模式。 (0:58) Teletune 使计算机操作智能体能够直接从未经标记的原始点击和按键活动日志中推断潜在的人类意图，并提取出可重用的工作流程。 (1:42) 概率偏差校正（PBC）在一项全球气象建模竞赛中胜出，目前已在肯尼亚部署，为小农户提供至关重要的 2 至 6 周次季节季风预测。 (4:48) 微软研究院宣布在 GitHub 上公开预览 BioEmu2，这是一种基础模型，旨在模拟药物发现中蛋白质的动态构象涨落和平衡态。 (5:42)

Reading summary

Edition #97 of Microsoft Research Focus surveys breakthroughs across autonomous agents, enterprise adoption patterns, and scientific machine learning in meteorology and molecular biology. The highlighted research focuses on reducing computational waste, diagnosing model comprehension limits, and aligning AI capabilities with practical domain-specific workflows. Compute Efficiency and Comprehension in Software Agents (0:00) Deploying AI agents for multi-step tasks often results in high compute costs when systems fail and attempt to recover through brute-force retries. To mitigate this, researchers developed FolkPilot, a self-improving framework that assesses whether backtracking to earlier decision points is computationally worthwhile. Tested across six benchmarks against leading foundation models, FolkPilot matched state-of-the-art performance while reducing processing resource demands by up to 59%. Separately, the Cabra framework investigates coding agent failure modes by synthesizing controlled challenges from scratch. Testing across thousands of generated problems revealed that while external tools like debuggers and linters enable agents to mask code-editing deficits, tools fail to compensate for comprehension gaps. Host Sowbhagya Suresh noted that understanding unfamiliar codebases, rather than modifying code, remains one of the primary obstacles to building capable coding agents. Extracting Interface Workflows from User Telemetry (1:42) Computer-use agents designed to interact with graphical interfaces—such as filling forms and navigating browsers—require human procedural knowledge to operate across varied environments. However, raw user activity logs typically combine disparate tasks into single sessions and lack explicit records of user intent. To address this challenge, researchers created Teletune, a framework that infers high-level task goals from unlabeled click and keystroke telemetry to segment and extract reusable workflows. In empirical evaluations spanning native desktop applications and browser tasks, Teletune consistently outperformed baseline methods, offering a scalable method for training agents to handle multi-step workflows in unfamiliar software. Enterprise Telemetry and Classroom AI Deployment (2:32) An analysis of aggregated enterprise Copilot chat data revealed that while document editing was the most frequent classified activity (exceeding 20% of interactions), broad statistics obscure real operational differences. Once researchers filtered out the top 20 aggregate use cases, distinct industry-specific workflows emerged. Suresh emphasized that organizations should look beyond high-level adoption metrics to evaluate actual work practices and domain-specific needs. In practical field deployments, the human-in-the-loop assistant Shiksha Copilot was tested with over 1,000 teachers in low-resource schools in India. The tool curtailed hours dedicated to administrative bureaucracy and bilingual curriculum development, leading to reported reductions in workplace stress. Rather than uncritically accepting automated outputs, participating teachers maintained editorial agency, actively adapting and refining generated lesson materials for their classrooms. Subseasonal Weather Forecasting and Biomolecular Dynamics (4:48) While numerical weather forecasts are dependable up to 10 days, predicting conditions in the two- to six-week subseasonal window remains difficult due to systemic model drift. Microsoft researchers created Probabilistic Bias Correction (PBC), a machine learning framework that evaluates historical discrepancies in numerical weather models to probabilistically adjust future projections. PBC won first place in a global extended-range modeling competition and is currently operational in Kenya to provide monsoon timing forecasts to smallholder farmers. In computational biology, Microsoft Research’s AI for Science team released a public preview of BioEmu2 on GitHub. Building on the Biomolecular Emulator architecture, BioEmu2 uses deep learning to model the equilibrium dynamics and conformational fluctuations of biomolecules rather than static 3D structures. Simulating these physical interactions aims to aid pharmacological discovery by revealing mechanisms to enhance beneficial protein functions or inhibit harmful binding events. Across these diverse applications, the findings indicate that deploying AI at scale depends on diagnosing failure modes, learning from real-world human interactions, and accounting for domain-specific contexts rather than relying on generic model capabilities.

精炼阅读摘要

微软研究院在第97期《Microsoft Research Focus》中集中展示了涵盖自主智能体效能优化、行业实际落地洞察以及前沿科学计算的最新成果。研究表明，推动人工智能实用化不仅需要降低多步决策的算力冗余与攻克底层认知瓶颈，还需结合深层业务场景与物理规律，以实现从专业工作流辅助到气象预报和蛋白质动态模拟的规模化应用。 编程智能体的决策回溯与代码理解瓶颈 (0:00) 针对人工智能智能体在执行复杂多步任务时常因单点错误导致整体崩溃且恢复算力高昂的问题，微软研究人员开发了具备自改进机制的智能体框架 FolkPilot。该框架能够动态评估此前各项决策的计算价值，自主判断在何时回溯并重新探索其他路径，从而取代了盲目的全盘重启或暴力求解。在六项基准测试中，FolkPilot 在保持与顶尖模型同等水准的前提下，将 AI 处理所需的计算资源消耗降低了高达 59%，显著提升了规模化部署的经济可行性。 与此同时，为精确定位编程智能体的失效机理，研究人员设计了受控合成诊断框架 Cabra，通过从零生成数千个梯度递增的代码挑战来隔离特定技能。实验发现了一个普遍现象：智能体经常依赖代码检查器、解释器和调试器等外部工具来掩盖其在代码生成与修改方面的不足；然而一旦面对陌生的代码库，智能体的理解能力便会出现断崖式下跌。主持人 Suresh 指出，这项发现证明了理解已有代码而非修改代码，依然是阻碍更有效编程智能体发展的核心瓶颈。 基于活动日志解析的计算机操作智能体 (1:42) 计算机使用智能体旨在通过模拟点击、键入与界面导航等操作来代替人类完成数字化工作流。然而，从日常用户活动日志中提炼这些程序性知识面临重大建模难题：未经标注的底层击键和鼠标遥测数据通常难以体现用户的真实意图，且多个不相关的任务往往混杂在同一个会话中。 为此，微软推出了 Teletune 框架，能够从未经标注的原始操作流中自动推断出高阶任务目标，进而分割出可复用的程序技能与结构化工作流。在针对标准桌面软件和网页端任务的多项实证测试中，Teletune 的任务执行表现均超越了现有基准方法。Suresh 评价称，Teletune 为智能体学习人类操作经验、适应未知软件环境并执行复杂多步操作提供了一种高度可扩展的解决方案。 真实落地中的行业差异与人机协同实践 (2:32) 在企业落地层面，研究人员对跨行业的 Copilot 聚合聊天日志分析揭示了表面同质化之下的深层差异。从宏观统计来看，文档编辑是出现频率最高的单项活动，在全部已分类交互中占比超过 20%。然而，当过滤掉排名前 20 的高频常规任务后，不同行业类别之间的专业化工作流差异开始明显分化。Suresh 提醒企业管理者，高频的写作与行政提示词往往掩盖了具有行业特异性的核心需求，企业制定 AI 战略时绝不能盲目套用通用指标，而应深入观察员工的实际业务流程与关键成果。 在教育场景中，一项在印度低资源学校针对 1000 多名教师的实地部署研究评估了教学助手 Shiksha Copilot 的实际效果。该工具显著缩短了教师在繁琐行政事务和双语教案设计上消耗的时间，大幅缓解了其工作压力。值得注意的是，教师并未完全依赖自动生成的内容，而是保持了积极的人在回路介入，运用自身的专业教学经验对生成材料进行二次修改与适配，在保留教学主导权的同时提升了课堂互动质量。 科学 AI：次季节气象预报与蛋白质构象模拟 (4:48) 传统数值天气预报在10天以内的准确度已相对成熟，但要将预测时效延伸至2至6周的“次季节”窗口仍存在巨大的科学障碍，这直接影响到农业耕作、野火防治与极端自然灾害的预防部署。为此，微软研发了概率偏差校正模型（PBC），该机器学习模型通过系统性分析物理气象模型的历史误差分布，学会对未来的数值预测进行概率维度的动态修正。PBC 在一项全球延伸期天气预报竞赛中摘得桂冠，目前已在肯尼亚投入实际运营，为当地依赖季风周期的农业社区提供关键气象支持。 在分子生物学领域，微软 AI for Science 团队发布了生物大分子基础模型 BioEmu2 的公开预览版。不同于以往仅生成静态三维晶体结构的模型，BioEmu2 基于 Biomolecular Emulator 架构，利用深度学习来模拟蛋白质在微观环境中的平衡动态分布与构象柔性。Suresh 指出，通过更深入地掌握蛋白质如何运动与结合，科研人员能够更精准地设计药物分子以增强有益生物活性或阻断有害分子相互作用，为靶向疗法研发开辟了新途径。 本期研究展示了微软在推进人工智能从通用大模型向兼具高能效、高鲁棒性以及垂直专业深度的技术架构演进。无论是在编程与系统操作中通过反思和日志解析攻克认知与算力瓶颈，还是在企业、教育、气象及生物医疗领域实现针对性突破，研究表明 AI 的真正价值取决于模型对特定物理规律、真实意图及专业领域先验知识的结合能力。

Detailed notes — zh

## Background 本期内容来自微软研究院播客频道（Microsoft Research Podcast），由微软研究院社区的研究人员于2026年10月29日发布在RSS与网络平台上。节目单简介显示，主持人Sowbhagya Suresh分享了微软最新的前沿科研进展，涵盖代码智能体面临的挑战、员工在职场中对AI的真实使用情况，以及扩展期天气预报等突破性研究。 ## TL;DR 本期节目集中梳理了微软在AI Agent架构、人机协作及科学AI领域的最新科研突破。在Agent技术方面，微软推出了降低59%计算资源开销的自我改进框架ForkPilot、用于深度诊断编程Agent代码理解短板的Cabra框架，以及从混乱操作日志中学习工作流的计算机操作框架Teletune。同时，企业级Copilot数据揭示了通用任务掩盖下的行业专属AI需求，而Shiksha Copilot则在印度低资源学校成功落地，赋能基层教师编写双语教案。最后，节目还介绍了斩获全球算法挑战赛头名并在肯尼亚落地应用的次季节天气预报校准框架PBC，以及面向蛋白质动力学模拟开源的基础模型BioEmu 2。 ## Atmosphere & style 整期节目呈现出一种严谨、前沿且快节奏的科技简报风格，主持人Sowbhagya Suresh以清晰沉稳、条理分明的语调娓娓道来。叙事逻辑严密，在理论模型创新与现实世界落地应用之间建立了流畅的连接，既有对底层算力优化和代码理解瓶颈的硬核拆解，又蕴含着利用AI改善全球南方教育公平与农业生计的人文关怀。听众仿佛置身于一场高效充实的微软实验室前沿学术巡礼，整体感受既理性务实，又充满对技术可能性的探索热情。 ## Bulletin - `00:00` 微软研究院推出 Agent 自我改进框架 ForkPilot，可自主判断何时回溯决策，在保持顶尖性能的同时将计算资源开销降低高达 59%。 - `00:00` 通过 Cabra 框架的自动生成基准测试，研究人员发现编程 Agent 的核心瓶颈并非写代码或使用外部工具，而是在面对陌生代码库时的代码理解与推理能力。 - `01:45` 全新 Teletune 框架能够直接从杂乱、无意图标注的用户日志中提取任务目标与可复用技能模块，大幅提升 Agent 在桌面和浏览器环境的操作泛化能力。 - `03:45` 在气象与科学智能领域，概率偏差校正框架 PBC 突破了 10 天以上次季节天气预报难题并在肯尼亚落地，而蛋白质动力学基础模型 BioEmu 2 也已公开预览并开源。 ## Key takeaways - `00:00` 复杂多步骤任务中，AI Agent 在执行出错时动态调整往往伴随巨大的计算开销与不确定性。 - `00:00` ForkPilot 框架通过自主学习识别具有回溯价值的决策节点并自适应调整策略，在 6 项主流基准上实现高达 59% 的资源节省。 - `00:00` Cabra 基准测试生成框架采用阶梯式合成挑战的方法，摆脱了静态测试集无法精确定位 Agent 失败变量的限制。 - `00:00` 对数千道代码挑战的分析表明，限制编程 Agent 效能的关键是理解陌生代码，而非编写新代码或调用工具。 - `01:45` 计算机操作 Agent 面临的核心痛点是人类软件操作日志缺乏动作意图标注，且多任务高度混杂。 - `01:45` Teletune 框架从原始操作日志中解析用户意图并抽象出标准化技能块，显著提升 Agent 跨陌生软件的适应性。 - `01:45` 企业 Copilot 交互数据显示“文档编辑”占据总活动的 20% 以上，但这一高频通用场景掩盖了各行业真实的特异性需求。 - `01:45` 剥离前 20 种高频通用活动后，不同行业员工使用 AI 的深层差异显现，企业决策应穿透表面数据关注核心成果。 - `03:45` Shiksha Copilot 采用人机协同模式辅助印度千余名低资源学校教师编写双语教案，有效减轻职业与行政压力。 - `03:45` 实地研究表明基层教师并非被动采纳 AI 输出，而是结合本土情境对教案进行二次重塑以契合学生需求。 - `03:45` 两周至数周的次季节预报是传统气象模型的盲区，对农业生产与防灾减灾等中远期决策至关重要。 - `03:45` 微软机器学习框架 PBC 借助历史物理模型误差学习纠偏，在全球预测赛夺冠并部署于肯尼亚监测季风降雨。 - `03:45` 微软 AI for Science 推出基础模型 BioEmu 2，专注于解析蛋白质三维结构与构象动力学，以加速靶向药物研发。 ## Detailed walkthrough ### `00:00` AI Agent 自我优化框架与编程 Agent 的瓶颈突破 随着人工智能系统承担起越来越复杂的任务，AI Agent（智能体）通常需要经历多步骤推理、连续决策以及频繁调用外部工具的过程。然而，当执行链路出现错误时，Agent 若要进行动态调整，往往需要消耗极其庞大的计算资源，且无法确保最终能得到更优的结果。为了解决这一可靠性与资源开销之间的矛盾，微软研究院推出了名为 ForkPilot 的自我改进框架。该框架能够自主学习并判断在何时“值得回溯先前的决策以探索全新路径”，并基于新产生的结果自适应调整后续策略。在涵盖主流 AI 模型的 6 项基准测试中，ForkPilot 不仅在性能上比肩现有的顶尖方法，更将 AI 处理资源的需求大幅降低了多达 59%，从而为大规模、低成本且高可靠性地部署复杂任务 Agent 铺平了道路。 与此同时，针对专门用于编写、推理和运行代码的编程 Agent（Coding Agents），研究人员正努力厘清其失败背后的根本原因。现有的基准测试通常难以精确定位导致成功或失败的独立变量。为此，微软研究院构建了名为 Cabra 的基准测试生成框架，能够从零开始合成代码挑战，以严格受控的方式逐步提升题目复杂度，从而准确定位对 Agent 而言最棘手的技能维度。通过对数千道自动生成的挑战进行深入分析，研究人员发现：Agent 往往能够熟练利用外部工具克服纯粹的代码编写障碍，但一旦面对陌生的现有代码库，其推理与理解能力就会急剧下降。关于这一点，主持人 Sowbhagya Suresh 指出：“研究结果表明，理解现有代码而非修改代码，仍然是开发更高效编程 Agent 所面临的最大障碍之一。” - **ForkPilot 核心表现**：在 6 项基准测试及主流大模型上验证，能够在保持顶尖任务表现的同时，降低高达 59% 的 AI 计算资源开销。 - **Cabra 框架机制**：摆脱静态测试集的局限，受控合成代码挑战并阶梯式增加复杂度，精准剥离出工具使用能力与代码理解能力的边界。 --- ### `01:45` 计算机操作 Agent 与职场 AI 真实使用图景 除了纯代码与文本环境，能够模拟人类进行点击、打字、填写表单、处理审批以及网页导航的“计算机操作 Agent”（Computer-Use Agents）正成为重要的前沿方向。这类 Agent 若要顺利完成工作，核心在于掌握人类实际操作软件的操作规程。用户操作日志虽然完整记录了点击与键盘输入，但通常缺乏动作背后的明确意图，且多项任务往往相互交织。新推出的 Teletune 框架正是为了打破这一瓶颈：它能直接从混乱的用户日志中推断出动作序列背后的任务意图，并提炼出可复用的技能模块与标准工作流。评测显示，Teletune 在桌面软件及浏览器环境中均大幅超越了现有方案，使 Agent 能够以极高的扩展性快速适应陌生软件并执行复杂的多步骤任务。 ```mermaid flowchart TD accTitle: Teletune 框架运行流程 accDescr: 展示从用户原始操作日志中学习动作意图、提取可复用技能模块并驱动 Agent 执行复杂工作流的过程。 subgraph DataInput["日志采集"] A["人类日常操作日志 (包含点击/键入/杂乱交互)"] end subgraph Parsing["Teletune 意图推断与抽象"] B["推断动作序列背后的具体任务目标"] C["提取通用、可复用的技能操作块"] end subgraph Execution["Agent 泛化执行"] D["在陌生软件与浏览器中执行多步骤工作流"] end A --> B A --> C B --> D C --> D ``` 在宏观的 AI 普及应用层面，员工使用 AI 的真实模式展现出了耐人寻味的复杂性。表面上看，不同行业的打工人似乎在以高度相似的方式使用生成式 AI。微软研究团队对广泛客户群体的 Copilot Chat 聚合活动数据进行了深度挖掘，数据显示“文档编辑”（Editing Documents）在所有分类活动中稳居首位，占比超过 20%。然而，当研究人员在分析中剥离掉最常见的这前 20 种高频通用活动后，各行业客户群体之间的行为差异便瞬间显露出来。通用功能的掩盖效应极易让管理者低估自身业务的特殊性。对于企业的决策者而言，这一洞察具有极为明确的导向意义，正如主持人 Sowbhagya Suresh 所强调：“寻找正确的机会意味着不仅要看表面用户统计数据，更要深入理解员工实际在做什么、他们如何开展工作，以及哪些成果最为关键。” - **Teletune 的优势**：解决了日志无意图标注、任务混杂的难题，实现了从被动行为记录到自主程序化技能的迁移。 - **Copilot 行为模式发现**：文档编辑占据全部 AI 交互的 20% 以上，但高频通识任务隐藏了不同业务场景对 AI 的专属性需求。 --- ### `03:45` 全球南方教育赋能、次季节天气预报与蛋白质动力学建模 在印度等低资源学校的实地调研中，一款名为 Shiksha Copilot 的人机协同（Human-in-the-loop）AI 工具展现了技术赋能“全球南方”（Global South）教育一线工作者的巨大潜力。该工具显著缩短了教师在双语教案编写和应对官僚行政流程上耗费的时间，使他们能够将更多精力投入到设计更具互动性与吸引力的课堂活动中。参与该项目的 1000 多名基层教师普遍反馈自身职业压力明显减轻，并对 AI 助教角色持积极态度。研究观察到，教师们并非被动接受 AI 生成的教案，而是积极发挥自身的专业教学经验，对生成材料进行因地制宜的二次剪裁与本土化重塑，从而完美贴合自身班级学生的实际需求。 在气象科学领域，十天以内的短期天气预报目前已具有极高的准确度，但两周至数周的“次季节预报窗口”（Extended-range Forecasting Window）在传统物理模型中依然是一大严峻难题。这一窗口期对于农业播种收割、森林野火防控以及极端灾害准备而言恰恰是做出关键决策的黄金时间。微软研发的全新机器学习纠错框架 PBC（Probabilistic Bias Correction，概率偏差校正）通过主动学习气象物理模型过去所犯的历史错误，有效填补了这一预测断层。PBC 不仅在一项全球扩展期预报挑战赛中夺得头魁，目前还已正式部署于肯尼亚，用于大幅提升对当地农民生计至关重要的季风降雨预测。 此外，由微软研究院科学智能团队（AI for Science）研发的全新基础模型 BioEmu 2 现已面向公众开放预览（Public Preview）。该模型脱胎于早先的生物分子仿真器（Biomolecular Emulator），专注于解析蛋白质等生物大分子的三维结构、构象动力学和生物学功能。主持人 Sowbhagya Suresh 阐述了这一模型对医药科学的深远价值：“通过更深入地了解蛋白质的功能运作，科学家们能够更进一步研发出更有效的药物，以增强有益的蛋白质活性或阻止有害蛋白质造成损伤。”该模型的相关代码与权重目前已上线微软研究院的 GitHub 仓库供科研人员探索。 - **Shiksha Copilot 落地实践**：覆盖 1000 多名印度低资源学校教师，以人机协同模式攻克双语教案备课与行政负担，重塑教师课堂投入度。 - **PBC 气象预报突破**：基于误差学习机制突破 10 天以上次季节气象预报瓶颈，全球算法大赛夺冠并落地肯尼亚支持农业决策。 - **BioEmu 2 进展**：微软 AI for Science 团队推出生物大分子结构与动力学基础模型，助力靶向药物分子机理研究，代码已在 GitHub 开源。 ## Notable quotes > `00:00` 研究结果表明，理解现有代码而非修改代码，仍然是开发更高效编程 Agent 所面临的最大障碍之一。 — Sowbhagya Suresh > `01:45` 寻找正确的机会意味着不仅要看表面用户统计数据，更要深入理解员工实际在做什么、他们如何开展工作，以及哪些成果最为关键。 — Sowbhagya Suresh ## Mentioned Sowbhagya Suresh, 微软 (Microsoft), 微软研究院 (Microsoft Research), ForkPilot, Cabra, Teletune, Copilot Chat, Shiksha Copilot, PBC (Probabilistic Bias Correction), BioEmu 2, GitHub ## Video description <p>Host Sowbhagya Suresh shares some of the latest news and cutting-edge research from across Microsoft—from understanding the challenges facing coding agents to uncovering how people use AI at work and improving extended-range weather forecasting.</p><p><a href="https://www.microsoft.com/en-us/research/podcast/microsoft-research-focus-edition-97/">Show notes</a></p><p><a href="https://www.linkedin.com/newsletters/research-focus-7335362495383040001/">Research Focus newsletter on LinkedIn</a></p><p><a href="https://www.microsoft.com/en-us/research/podcast/">Microsoft Research Podcast</a></p> [Source](https://media.blubrry.com/microsoftresearch/content.blubrry.com/microsoftresearch/MRP-2026-10-30-322.mp3)

Source quotes

[0:00] Sowbhagya Suresh: 研究结果表明，理解现有代码而非修改代码，仍然是开发更高效编程 Agent 所面临的最大障碍之一。 [1:45] Sowbhagya Suresh: 寻找正确的机会意味着不仅要看表面用户统计数据，更要深入理解员工实际在做什么、他们如何开展工作，以及哪些成果最为关键。

Source quotes

[2:32] Sowbhagya Suresh: Finding the right opportunities means looking beyond user statistics to understand what employees actually do, how they work, and what outcomes matter most. [2:32] Sowbhagya Suresh: Don't assume you have the same AI needs as others in your industry. [0:58] Sowbhagya Suresh: The findings suggest that understanding existing code rather than modifying it remains one of the biggest obstacles to developing more effective coding agents.

Original transcript

[0:00] Hi, I'm Saurabhagya Suresh. Welcome to Microsoft Research Focus. You've probably read our newsletter and now you can get highlights from across the research community at Microsoft here on YouTube and major podcast platforms. Whether you read, watch or listen to it, Research Focus is your source for research news at Microsoft. Coming up in this edition, what everyday AI use reveals about the way we work, how a new AI model could help scientists better understand proteins, to improving weather forecasts weeks in advance and more. So here's what's new. More and more, AI agents are handling tasks that are complex. Tasks that require multiple steps, decisions, and interactions with external tools. But how can we count on AI agents to do the work if their ability to adjust when things go wrong may require a lot of resources with no guarantees of a better result? A new self-improving framework called FolkPilot learns when it's worth revisiting earlier decisions to find a different approach and adapts its strategy based on new outcomes. Tested across six benchmarks and leading AI models, Folkpilot measures up to the best existing methods while reducing AI processing demands by up to 59%. These results suggest that AI agents can solve complex problems more reliably and efficiently, making them more practical and cost-effective to deploy at scale. Coding agents execute user queries by reasoning about, editing, and running code. Understanding why these agents fail is critical if we want to build more capable and reliable coding agents. Unfortunately, existing benchmarks often make it difficult to identify which factors contribute to success or failure. Cabra is a framework for generating coding challenges from scratch. It lets researchers increase complexity in a controlled way and pinpoints the skills that are most difficult for agents. Across thousands of generated challenges, the researchers found that agents often overcame coding problems by using tools. Yet the agents still struggled when they needed to understand unfamiliar code bases. The findings suggest that understanding existing code rather than modifying it remains one of the biggest obstacles to developing more effective coding agents. Computer use agents are AI systems that are designed to complete tasks that require clicking and typing. They can fill in forms, process requests, and navigate websites. To do this effectively, they need to learn the procedures that people use to get work done. User activity logs can offer a way to capture this knowledge. but activity logs typically don't record the intent behind each action and often combine tasks. The new framework Teletune learns reusable skills and workflows from these logs. It also infers the task goals associated with the sequences of actions. In evaluations across software and browser tasks, Teletune outperformed existing approaches. By learning from existing activity, Teletune offers a scalable way to help AI agents adapt to unfamiliar software and carry out complex multi-step workflows. Workers across industries may appear to be using AI in surprisingly similar ways, but a closer look reveals a different story. Microsoft researchers analyzed aggregated copilot chat activity across a broad range of customer categories and found that editing documents led all classified activities accounting for more than 20%. But when researchers set aside the 20 most common activities, differences between customer categories became substantially more pronounced. The most common uses of AI may be hiding important differences in how organizations use it for more specialized work. For business leaders, the takeaway is clear. Don't assume you have the same AI needs as others in your industry. Finding the right opportunities means looking beyond user statistics to understand what employees actually do, how they work, and what outcomes matter most. A new study shows how Shiksha Co-Pilot, a human-in-the-loop AI tool, could transform teacher workflows in Global South classrooms. The study in low-resource schools in India found the tool reduced the time teachers spent on bilingual lesson planning and bureaucracy. That gave them more time to craft more engaging classroom activities. More than a thousand teachers participated and many reported reduced stress levels and a positive impression of how AI can help them. These teachers actively engage with Shiksha Co-Pilot, using their professional expertise to adapt and repurpose materials to soothe their own classrooms. 10-day weather forecasts have become remarkably dependable. But what about the weeks that follow? The two decisions [5:00] window remains a major challenge for forecasting models, which is a problem. That's when a wide range of critical decisions get made related to farming, managing wildfires, and preparing for extreme weather. A new machine learning framework called Probabilistic Bias Correction, or PBC, is helping close that gap. BBC learns from a weather model's past mistakes and helps correct them. And the early results are promising. BBC took top honours in a global competition focused on improving extended range forecasting. It's also being deployed in Kenya, where it's helping improve critical monsoon forecasts for farmers. This development could help make weather predictions more useful across a broad range of applications. BioEMU2, a new foundation model for understanding biomolecular structure, dynamics, and function, is now available in public preview. Developed by the Microsoft Research AI for Science team, BioEMU2 builds on Biomolecular Emulator, a deep learning model designed to help scientists better understand how proteins work. That understanding could have important implications for medicine. By learning more about how proteins function, scientists could move closer to developing more effective medications that enhance beneficial protein activity or prevent proteins from causing harm. Go to the BioEMU2 Microsoft Research GitHub repository to explore the model and learn more. And that's your research focus. Want to learn more and stay connected? Explore the latest research, discover the people behind it, and keep tabs on the idea shaping what's next by following Microsoft Research on major social platforms. I'm Saurabhagya Suresh. Thanks so much for tuning in. We'll see you in the next Microsoft Research Focus.