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

SkillOpt：https://github.com/microsoft/SkillOpt 是一个文本空间优化器，由来自微软、上海交通大学、同济大学和复旦大学的研究团队开发。

SkillOpt 在目标模型保持冻结的情况下训练单个自然语言技能文档。优化器模型读取评分回合并提出有限的添加/删除/替换编辑。只有当评分严格提高时，保留的选择划分才会接受编辑。导出的产物是一个文件，best_skill.md。

迁移表格报告三列。Baseline 是目标模型的无技能分数。Direct 是在该目标域内训练的 SkillOpt。Transferred 应用在其他地方训练的技能，没有进一步优化。

有用的比较不是 transferred 与 direct 的对比，而是域内增益在迁移后能保留多少。

技能在 GPT-5.4 上训练，并部署到较小的变体上。

两行值得关注。GPT-5.4-mini 上的 SpreadsheetBench 保留了 82% 的域内增益。这接近于免费复用。GPT-5.4-nano 上的 LiveMath 行更奇怪：迁移的技能得分为 28.8，而域内 SkillOpt 结果为 27.2。论文将其解读为某些学习的程序与目标模型无关的证据。

GPT-5.4-nano 上的 SpreadsheetBench 行较弱，仅为 16%。保留效果并不均匀，论文也没有声称均匀。其声明的界限更窄：没有任何一行低于目标模型的无技能基线。

注意范围。所有四行都在同一 GPT 系列内。跨系列迁移，例如 GPT 到 Qwen：https://qwen.ai/blog?id=qwen3.5，尚未测试。

这是部署中最重要的部分。所有行均使用 GPT-5.5。

第一行是重点。在 Codex：https://openai.com/index/introducing-codex/ 内优化的技能将 Claude Code：https://www.anthropic.com/claude-code 从 22.1 提升到 81.8。略高于 Claude Code 通过从头训练自身技能达到的 80.4。

这两种工具架构暴露出不同的工具和文件 API 以及不同的命令界面。能够在这种变化下仍然生存的技能，并不是编码命令配方。研究论文：https://arxiv.org/pdf/2605.23904 将 SpreadsheetBench 的可移植性归因于工作簿级的过程：先结构化检查、识别公式的验证、以及静态值物化。这些内容不受运行 Python 的 CLI 类型影响。

LiveMath 展示了相反的情况。Codex → Claude Code 只保留了领域内提升的 10%。这种不对称值得关注。流程型技能——如何检查、验证和格式化——似乎是可移植的类别。依赖推理的技能则更依赖于其训练环境。

这里没有“直接”列。没有报告 Omni-MATH 上的领域内 SkillOpt 运行，因此比较对象仅为无技能情况。三种模型规模的增益都是正的，但很小。研究论文得出的结论是，技能在测试实例和答案格式规范变化之后，仍能保留可复用的数学过程。

研究论文中明确说明了机制。三种执行模式：直接聊天、Codex、Claude Code——都使用相同的 best_skill.md 文件格式。正是这种共享契约，使跨工具架构实验成为可能。

Codex 架构将当前技能渲染到每个任务的 SKILL.md 文件中，与任务文件一起，然后读取紧凑的执行痕迹。Claude Code 架构通过 claude CLI 反映相同的工作空间契约。两种工具架构都没有使用定制的技能格式。

该成果的形式也支持可移植性。最终技能在六个基准测试中运行时长度在 379 到 1,995 个 token 之间，中位数约 920。它们由 1 到 4 次通过审核的编辑组合而成。论文的图 4 从每个基准测试中抽取了一条学习规则，且全部为流程型，而非特定实例的规则。SpreadsheetBench 规则逐字原文：检查工作簿结构和公式，然后在整个请求的目标范围内写入评估后的静态值，而不是依赖 Excel 的重新计算。

训练成本一次性支付，离线计算，并且可测量。研究论文报告每个绝对测试点的训练标记数为0.6M到46.4M，取决于基准。SpreadsheetBench为每点0.6M；DocVQA为46.4M。优化器模型仅在训练期间运行，在部署时不会增加推理调用。

如果在一个环境中训练的技能在另一个环境中也能发挥作用，那么一次性成本就可以分摊到多个环境中。Codex → Claude Code SpreadsheetBench的结果就是存在性证明。这也意味着你可以在工具成本最低的地方进行优化，并在产品所在的地方部署。

审计角度是独立且被低估的。部署的产物是一个文本文件，领域从业者可以在几分钟内阅读。对其每一次的修改都是可追踪的：每一步都会记录一个 edit_apply_report.json，其中包含每次编辑的接受和跳过状态。可迁移性加上可检查性，是与仅发布微调权重不同的操作态度。

资源：论文：https://arxiv.org/abs/2605.23904，GitHub：https://github.com/microsoft/SkillOpt，项目页面：https://microsoft.github.io/SkillOpt/，文档：https://github.com/microsoft/SkillOpt/blob/main/docs/index.md，PyPI：https://pypi.org/project/skillopt/，演示视频：https://youtu.be/JUBMDTCiM0M

参考的基线：GEPA：https://arxiv.org/abs/2507.19457，TextGrad：https://arxiv.org/abs/2406.07496，EvoSkill：https://arxiv.org/abs/2603.02766 以及 Trace2Skill：https://arxiv.org/abs/2603.25158

参考的基准：SearchQA：https://arxiv.org/abs/1704.05179, SpreadsheetBench：https://arxiv.org/abs/2406.14991, DocVQA：https://arxiv.org/abs/2007.00398, LiveMathematicianBench：https://arxiv.org/abs/2604.01754 和 ALFWorld：https://openreview.net/forum?id=0IOX0YcCdTn

Asif Razzaq是Marktechpost Media Inc.的首席执行官。作为一位富有远见的企业家和工程师，Asif致力于利用人工智能的潜力造福社会。他最近的工作是推出人工智能媒体平台Marktechpost，该平台以其对机器学习和深度学习新闻的深入覆盖而脱颖而出，内容既技术可靠又易于广大受众理解。该平台每月访问量超过200万，显示了其在受众中的受欢迎程度。

[免费指南] 保护 AI 代理、MCP 服务器和 LLM 应用程序：https://pxllnk.co/lxn88m

SkillOpt：https://github.com/microsoft/SkillOpt is a text-space optimizer developed by a team of researchers from Microsoft, Shanghai Jiao Tong University, Tongji University, and Fudan University.

SkillOpt trains a single natural-language skill document while the target model stays frozen. An optimizer model reads scored rollouts and proposes bounded add/delete/replace edits. A held-out selection split accepts an edit only when the score strictly improves. The exported artifact is one file, best_skill.md .

The transfer tables report three columns. Baseline is the target’s no-skill score. Direct is SkillOpt trained in-domain on that exact target. Transferred applies a skill trained elsewhere, with no further optimization.

The useful comparison is not transferred versus direct. It is how much of the in-domain gain survives the move.

Skills were trained on GPT-5.4 and deployed on smaller variants.

Two rows deserve attention. SpreadsheetBench on GPT-5.4-mini keeps 82% of the in-domain gain. That is close to free reuse. The LiveMath row on GPT-5.4-nano is stranger: the transferred skill scores 28.8 against an in-domain SkillOpt result of 27.2. The paper reads this as evidence that some learned procedures are target-model agnostic.

The GPT-5.4-nano SpreadsheetBench row is the weak one at 16%. Retention is not uniform, and the paper does not claim it is. Its stated bound is narrower: no row falls below the target’s no-skill baseline.

Note the scope. All four rows stay inside one GPT family. Cross-family transfer, such as GPT to Qwen：https://qwen.ai/blog?id=qwen3.5, is not tested.

This is the section that matters most for deployment. All rows use GPT-5.5.

The first row is the headline. A skill optimized inside Codex：https://openai.com/index/introducing-codex/ lifted Claude Code：https://www.anthropic.com/claude-code from 22.1 to 81.8. That slightly exceeds the 80.4 Claude Code reached by training its own skill from scratch.

The two harnesses expose different tool and file APIs and different command surfaces. A skill that survives that shift is not encoding command recipes. The research paper：https://arxiv.org/pdf/2605.23904 attributes SpreadsheetBench’s portability to workbook-level procedures: structure-first inspection, formula-aware verification, and static-value materialization. Those hold regardless of which CLI runs the Python.

LiveMath tells the opposite story. Codex → Claude Code retains only 10% of the in-domain gain. The asymmetry is worth sitting with. Procedural skills — how to inspect, verify, and format — appear to be the portable class. Reasoning-heavy skills appear more tied to their training environment.

There is no Direct column here. No in-domain SkillOpt run on Omni-MATH is reported, so the comparison is against no-skill only. Gains are positive across all three model scales but small. The research paper’s reading is that the skill retained reusable mathematical procedure after both the test instances and the answer-format conventions changed.

The mechanism is stated plainly in the research paper. All three execution modes: direct chat, Codex, Claude Code – consume the same best_skill.md file format. That shared contract is what makes the cross-harness experiment possible in the first place.

The Codex harness renders the current skill to a per-task SKILL.md alongside task files, then reads back a compact execution trace. The Claude Code harness mirrors the same workspace contract through the claude CLI. Neither harness gets a bespoke skill format.

The artifact’s shape supports portability too. Final skills run 379 to 1,995 tokens across the six benchmarks, with a median near 920. They are assembled from 1 to 4 accepted edits. The paper’s Figure 4 samples one learned rule per benchmark, and all are procedural rather than instance-specific. The SpreadsheetBench rule, verbatim: inspect workbook structure and formulas, then write evaluated static values across the full requested target range instead of relying on Excel recalculation.

Training cost is paid once, offline, and measured. The research paper reports 0.6M to 46.4M training tokens per absolute test point, depending on benchmark. SpreadsheetBench sits at 0.6M per point; DocVQA at 46.4M. The optimizer model runs only during training and adds zero inference-time calls at deployment.

If a skill trained in one harness holds up in another, that one-time cost spreads across environments. The Codex → Claude Code SpreadsheetBench result is the existence proof. It also implies you can optimize where tooling is cheapest and deploy where the product lives.

The audit angle is separate and underrated. The deployed artifact is a text file a domain practitioner can read in minutes. Every change to it is traceable: each step records an edit_apply_report.json with per-edit accept and skip status. Portability plus inspectability is a different operational posture than shipping fine-tuned weights.

Resources: Paper：https://arxiv.org/abs/2605.23904, GitHub：https://github.com/microsoft/SkillOpt, Project page：https://microsoft.github.io/SkillOpt/, Docs：https://github.com/microsoft/SkillOpt/blob/main/docs/index.md, PyPI：https://pypi.org/project/skillopt/ and Demo video：https://youtu.be/JUBMDTCiM0M

Baselines referenced: GEPA：https://arxiv.org/abs/2507.19457, TextGrad：https://arxiv.org/abs/2406.07496, EvoSkill：https://arxiv.org/abs/2603.02766 and Trace2Skill：https://arxiv.org/abs/2603.25158

Benchmarks referenced: SearchQA：https://arxiv.org/abs/1704.05179, SpreadsheetBench：https://arxiv.org/abs/2406.14991, DocVQA：https://arxiv.org/abs/2007.00398, LiveMathematicianBench：https://arxiv.org/abs/2604.01754 and ALFWorld：https://openreview.net/forum?id=0IOX0YcCdTn

Asif Razzaq is the CEO of Marktechpost Media Inc.. As a visionary entrepreneur and engineer, Asif is committed to harnessing the potential of Artificial Intelligence for social good. His most recent endeavor is the launch of an Artificial Intelligence Media Platform, Marktechpost, which stands out for its in-depth coverage of machine learning and deep learning news that is both technically sound and easily understandable by a wide audience. The platform boasts of over 2 million monthly views, illustrating its popularity among audiences.

[FREE GUIDE] Securing AI Agents, MCP Servers & LLM Apps ：https://pxllnk.co/lxn88m