\uselogo\uselogo

WikiSkill: Compiling Agent Experience into Persistent Knowledge for Skill EvolutionWikiSkill：将智能体经验编译为持久化知识以实现技能演进

Liyan Tang Affiliation: Google Research Cyrus Rashtchian Affiliation: Google Research Chun-Sung Ferng Affiliation: Google Research Andrew Tomkins Affiliation: Google Research Da-Cheng Juan Affiliation: Google Research Tu Vu Corresponding author: lytang@google.com, ttvu@google.com Affiliation: Google Research Affiliation: Virginia Tech

Abstract摘要

Agent skills package specialized knowledge and workflows into reusable resources that extend AI agent capabilities. Recent work automatically discovers such skills from agent experience, which enables agents to progressively adapt through interaction. However, the insights that guide skill development typically remain scattered across optimization histories, limiting their systematic reuse across iterations. We introduce WikiSkill, a framework that co-evolves agent skills with a persistent knowledge base (wiki). At a high level, WikiSkill separates raw execution experience, accumulated knowledge, and executable skills, while continuously consolidating experience into the wiki, which subsequent skill updates can build on. Across diverse benchmarks and models, WikiSkill consistently outperforms state-of-the-art skill-evolution methods and improves over no-skill baselines in most model-benchmark settings. We find that skill evolution complements model scaling: larger models generally benefit more from evolved skills, while smaller models with skills can outperform substantially larger models without them. We also find that evolved skills transfer effectively across models and model families, and skills evolved by other models can outperform self-evolved skills. Finally, our ablation studies confirm that persistent knowledge accumulation in the wiki is critical for effective skill evolution. These results demonstrate the benefits of systematically accumulating and refining agent experience for developing reusable and transferable skills.智能体技能将专业知识和工作流封装为可重用的资源，从而扩展了 AI 智能体的能力。近期的研究通过智能体经验自动发现此类技能，使智能体能够通过交互逐步适应。然而，指导技能开发的见解通常散落在优化历史中，限制了其在迭代间的系统性重用。我们引入了 WikiSkill，这是一个将智能体技能与持久化知识库（Wiki）协同演进的框架。从宏观层面看，WikiSkill 将原始执行经验、积累的知识和可执行技能分离开来，同时不断将经验整合进 Wiki，供后续的技能更新作为基础。在多项基准测试和模型中，WikiSkill 的表现始终优于最先进的技能演进方法，并在大多数模型-基准测试场景中超过了无技能基线。我们发现，技能演进与模型扩展互补：较大的模型通常能从演进后的技能中获益更多，而拥有技能的较小模型可以超越性能大幅领先但缺乏技能的模型。我们还发现，演进后的技能可以在不同模型和模型系列间有效迁移，且由其他模型演进的技能有时能优于自身演进的技能。最后，消融研究证实，Wiki 中的持久化知识积累对于有效的技能演进至关重要。这些结果证明了系统性地积累和提炼智能体经验对于开发可重用和可迁移技能的益处。

Figure 1: WikiSkill consistently improves over both the no-skill baseline and existing skill-evolution methods. Interestingly, its advantage becomes more pronounced for stronger models. We report average accuracy across the evaluated benchmarks for each model using no skills or skills evolved by EvoSkill, SkillOpt, and WikiSkill (see Table 1 for details).图 1：WikiSkill 在无技能基线和现有技能演进方法的基础上均实现了持续提升。有趣的是，其优势在更强的模型中表现得更为显著。我们报告了每个模型在评估基准上的平均准确率，分别使用无技能、EvoSkill、SkillOpt 和 WikiSkill 演进的技能（详情见表 1）。

1 Introduction1 引言

General-purpose AI agents are increasingly capable of performing complex tasks across domains (Patwardhan et al., 2026; Jackson et al., 2025; Merrill et al., 2026; Phan et al., 2026). However, reliably accomplishing real-world tasks often requires domain-specific expertise (e.g., procedural knowledge and workflows). Agent skills (Zhang et al., 2025; Li et al., 2026; Chen et al., 2026a; Liu et al., 2026) provide a lightweight, open format for capturing such expertise without updating model parameters. At its core, a skill packages instructions, scripts, and other resources into a reusable filesystem-based module (i.e., an organized directory) (Zhang et al., 2025; Li et al., 2026; Anthropic, 2026; Xia et al., 2026b). This design makes specialized knowledge consistent, auditable, and reusable across skill-compatible agents. It also supports progressive disclosure (Jiang et al., 2026), where agents only load relevant content at any given time, which saves context space. More broadly, skills provide a natural mechanism for accumulating knowledge independently of model parameters.通用 AI 智能体在跨领域执行复杂任务方面的能力日益增强（Patwardhan 等人，2026；Jackson 等人，2025；Merrill 等人，2026；Phan 等人，2026）。然而，要可靠地完成现实世界中的任务，通常需要特定领域的专业知识（例如程序性知识和工作流）。智能体技能（Zhang 等人，2025；Li 等人，2026；Chen 等人，2026a；Liu 等人，2026）提供了一种轻量级、开放的格式，无需更新模型参数即可捕获此类专业知识。其核心在于，技能将指令、脚本和其他资源封装为一个基于文件系统的可重用模块（即一个组织良好的目录）（Zhang 等人，2025；Li 等人，2026；Anthropic，2026；Xia 等人，2026b）。这种设计使得专业知识在技能兼容的智能体之间保持一致、可审计且可重用。它还支持渐进式披露（Jiang 等人，2026），即智能体仅在需要时加载相关内容，从而节省上下文空间。更广泛地说，技能提供了一种独立于模型参数积累知识的自然机制。

Developing effective skills, however, remains challenging. Most agent skills are manually authored, which requires anticipating the procedural knowledge and workflows that an agent will need (Li et al., 2026; Liang et al., 2026; Xu and Yan, 2026). This challenge motivates recent work that iteratively develops agent skills by executing agents on training tasks, analyzing successful and failed trajectories, and refining skills based on the resulting experience (Yuksekgonul et al., 2025; Agrawal et al., 2026; Alzubi et al., 2026; Ni et al., 2026; Ouyang et al., 2026; Yang et al., 2026).然而，开发有效的技能仍然具有挑战性。大多数智能体技能是手动编写的，这需要预判智能体所需的程序性知识和工作流（Li 等人，2026；Liang 等人，2026；Xu 和 Yan，2026）。这一挑战推动了近期的研究，即通过在训练任务上执行智能体、分析成功和失败的轨迹，并基于所得经验提炼技能，从而迭代开发智能体技能（Yuksekgonul 等人，2025；Agrawal 等人，2026；Alzubi 等人，2026；Ni 等人，2026；Ouyang 等人，2026；Yang 等人，2026）。

A key design question is how to preserve and organize what an agent learns throughout skill evolution. Prior work addresses this question in different ways. EvoSkill (Alzubi et al., 2026) maintains a cumulative history of prior proposals and their evaluation outcomes; Trace2Skill (Ni et al., 2026) extracts and consolidates lessons across execution trajectories into skill updates; and SkillOpt (Yang et al., 2026) uses rejected-edit feedback and epoch-wise meta guidance. However, these methods do not maintain what has been learned as a separate, evolving knowledge representation. Inspired by Karpathy (2026)’s perspective on LLM Wiki, which advocates compiling experience into persistent, compounding knowledge, we ask: Can agent experience be similarly compiled into persistent knowledge to support long-term skill evolution? We introduce WikiSkill, which adds a structured knowledge layer between raw experience and executable procedures (i.e., skills). This layer allows skill development to build on increasingly well-supported and integrated knowledge across iterations, rather than on knowledge scattered across skill-evolution artifacts.一个关键的设计问题是如何在技能演进过程中保存和组织智能体所学到的东西。以往的研究以不同方式解决了这个问题。EvoSkill（Alzubi 等人，2026）维护了过往提议及其评估结果的累积历史；Trace2Skill（Ni 等人，2026）提取并整合了跨执行轨迹的经验教训，用于技能更新；而 SkillOpt（Yang 等人，2026）则使用了拒绝编辑反馈和周期性的元指导。然而，这些方法并未将所学内容维护为一种独立的、演进的知识表示。受 Karpathy（2026）关于 LLM Wiki 的观点启发（主张将经验编译为持久的、复合的知识），我们提出：智能体经验是否也能被编译为持久化知识，以支持长期的技能演进？我们引入了 WikiSkill，它在原始经验和可执行程序（即技能）之间增加了一个结构化知识层。该层允许技能开发在迭代过程中基于日益完善和整合的知识进行，而不是基于散落在技能演进产物中的知识。

WikiSkill organizes the agent workspace into three layers: a Raw Layer that stores immutable execution traces, a Wiki Layer that maintains structured knowledge, and a Skill Layer that contains evolving procedural knowledge (Figure 2). Each iteration involves four components: an Inference Agent that executes rollouts using the current skills, a Wiki Maintainer that consolidates traces into the wiki, a Skill Proposer that uses the wiki and traces to propose skill updates, and a Gating and Rollback mechanism that retains updates that improve validation performance. While skill updates can be rolled back, the wiki persists so that future updates can build on accumulated knowledge. At a high level, these components form a continual loop in which experience is consolidated into persistent knowledge that supports skill evolution.WikiSkill 将智能体工作空间组织为三个层级：存储不可变执行追踪的“原始层”（Raw Layer）、维护结构化知识的“Wiki 层”（Wiki Layer），以及包含演进中程序性知识的“技能层”（Skill Layer）（图 2）。每次迭代包含四个组件：使用当前技能执行任务的推理智能体（Inference Agent）、将追踪整合进 Wiki 的 Wiki 维护者（Wiki Maintainer）、利用 Wiki 和追踪提出技能更新的技能提议者（Skill Proposer），以及保留提升验证性能的更新的门控与回滚机制（Gating and Rollback）。尽管技能更新可以回滚，但 Wiki 会持续存在，以便后续更新能够建立在积累的知识之上。从宏观层面看，这些组件形成了一个持续循环，将经验整合为支持技能演进的持久化知识。

We evaluate WikiSkill across five benchmarks spanning mathematical reasoning (LiveMathematicanBench (He et al., 2026)), web search (SealQA (Pham et al., 2026)), spreadsheet manipulation (SpreadSheetBench (Ma et al., 2024)), long-context document question answering (OfficeQA (Singhvi et al., 2025)), and interactive embodied tasks (ALFWorld (Shridhar et al., 2021)), using five models from the Qwen (Qwen Team, 2026a; Qwen Team, 2026b), Gemma (Gemma Team, 2026), and Gemini (Google DeepMind, 2026) families. We find that WikiSkill outperforms existing skill-evolution methods and improves over no skills in most settings. Interestingly, skill evolution complements model scaling. Within the Qwen family, WikiSkill improves average performance by 12.3%, 17.5%, and 23.9% for 4B, 9B, and 27B models, respectively, with gains increasing with model scale. At the same time, evolved skills can compensate for substantial model scale: Qwen-3.5-9B with WikiSkill outperforms Qwen-3.6-27B without skills (47.4% vs. 39.4%). We further find that evolved skills transfer effectively across model families and can outperform self-evolved skills. On ALFWorld, for example, Qwen-3.5-9B reaches 70.2% with a Qwen-3.6-27B-evolved skill, compared with 63.4% using its own skill. These results suggest that skill discovery and skill execution are distinct capabilities. Finally, our analysis shows that the persistent wiki is critical to these gains, supporting our hypothesis that accumulating and refining knowledge across iterations improves skill evolution.我们评估了 WikiSkill 在五个基准测试上的表现，涵盖数学推理（LiveMathematicanBench (He 等人，2026)）、网络搜索（SealQA (Pham 等人，2026)）、电子表格操作（SpreadSheetBench (Ma 等人，2024)）、长上下文文档问答（OfficeQA (Singhvi 等人，2025)）以及交互式具身任务（ALFWorld (Shridhar 等人，2021)），并使用了来自 Qwen (Qwen Team, 2026a; Qwen Team, 2026b)、Gemma (Gemma Team, 2026) 和 Gemini (Google DeepMind, 2026) 系列的五个模型。我们发现，WikiSkill 在大多数场景下均优于现有的技能演进方法，且表现好于无技能基线。有趣的是，技能演进与模型扩展互补。在 Qwen 系列中，WikiSkill 对 4B、9B 和 27B 模型的平均性能提升分别为 12.3%、17.5% 和 23.9%，且增益随模型规模增大而增加。同时，演进后的技能可以弥补模型规模的巨大差距：使用 WikiSkill 的 Qwen-3.5-9B 表现优于无技能的 Qwen-3.6-27B（47.4% 对 39.4%）。我们进一步发现，演进后的技能可以在不同模型系列间有效迁移，且表现可能优于自身演进的技能。例如，在 ALFWorld 上，Qwen-3.5-9B 使用 Qwen-3.6-27B 演进的技能达到了 70.2% 的准确率，而使用自身技能时仅为 63.4%。这些结果表明，技能发现和技能执行是两种不同的能力。最后，我们的分析表明，持久化的 Wiki 对这些收益至关重要，支持了我们的假设，即跨迭代积累和提炼知识能改善技能演进。

In summary, our main contributions are:总之，我们的主要贡献如下：

•

We introduce WikiSkill, a framework that co-evolves agent skills with a persistent knowledge base that continually organizes and refines knowledge from agent experience.我们引入了 WikiSkill，这是一个将智能体技能与持久化知识库协同演进的框架，该知识库不断组织和提炼来自智能体经验的知识。
•

We demonstrate across five benchmarks and five models that WikiSkill consistently outperforms existing skill-evolution methods, with ablations confirming the importance of persistent knowledge accumulation.我们在五个基准测试和五个模型上证明，WikiSkill 的表现始终优于现有的技能演进方法，消融研究证实了持久化知识积累的重要性。
•

We systematically study how evolved skills interact with model capability, showing that skill evolution complements model scaling and that evolved skills can transfer effectively across models, sometimes outperforming self-evolved skills.我们系统地研究了演进后的技能如何与模型能力交互，表明技能演进与模型扩展互补，且演进后的技能可以在不同模型间有效迁移，有时甚至优于自身演进的技能。

Taken together, we hope that our work will spur more fundamental research on how agents can accumulate, organize, and reuse knowledge from experience.综合来看，我们希望这项工作能激发关于智能体如何从经验中积累、组织和重用知识的更基础的研究。

2 Problem Setup2 问题设置

We formalize the task of iterative skill evolution for LLM agents. Let $\mathcal{D}=\{(x_{i},y_{i})\}_{i=1}^{N}$ be a dataset of tasks, where $x_{i}$ denotes a task instance and $y_{i}$ denotes its ground-truth answer. We partition $\mathcal{D}$ into three disjoint splits: training tasks $\mathcal{D}_{\text{train}}$ , validation tasks $\mathcal{D}_{\text{val}}$ , and testing tasks $\mathcal{D}_{\text{test}}$ .我们将 LLM 智能体的迭代技能演进任务形式化。令 𝒟={(xi,yi)}i=1N\mathcal{D}=\{(x_{i},y_{i})\}_{i=1}^{N} 为任务数据集，其中 xix_{i} 表示任务实例，yiy_{i} 表示其真实答案。我们将 𝒟\mathcal{D} 划分为三个互不相交的部分：训练任务 𝒟train\mathcal{D}_{\text{train}}、验证任务 𝒟val\mathcal{D}_{\text{val}} 和测试任务 𝒟test\mathcal{D}_{\text{test}}。

An agent $\pi$ is an LLM-based system equipped with a set of tools $\mathcal{U}$ (e.g., a bash shell, search APIs, or file readers) and an active skill set $S=\{s_{1},s_{2},\dots,s_{M}\}$ . A skill is a modular, filesystem-based directory that packages domain-specific procedural knowledge into instructions, scripts, and other resources (Zhang et al., 2025; Li et al., 2026; Chen et al., 2026a; Liu et al., 2026). Specifically, each skill contains a SKILL.md file with frontmatter metadata (a unique name and concise description) alongside full procedural instructions and applicability conditions. The skill set $S$ is initialized to empty ( $\emptyset$ ) and developed for each dataset through the evolution process.智能体 π\pi 是一个基于 LLM 的系统，配备了一组工具 𝒰\mathcal{U}（例如 bash shell、搜索 API 或文件阅读器）以及一个主动技能集 S={s1,s2,…,sM}S=\{s_{1},s_{2},\dots,s_{M}\}。技能是一个模块化的、基于文件系统的目录，将特定领域的程序性知识封装为指令、脚本和其他资源（Zhang 等人，2025；Li 等人，2026；Chen 等人，2026a；Liu 等人，2026）。具体而言，每个技能包含一个带有前言元数据（唯一名称和简要描述）的 SKILL.md 文件，以及完整的程序性指令和适用条件。技能集 SS 初始化为空（∅\emptyset），并通过演进过程针对每个数据集进行开发。

When executing a task $x_{i}$ , the agent receives the task context $x_{i}$ and access to the available skills $S$ . The agent interacts with the environment over multiple steps using its tools and skills to generate an execution trajectory $\tau_{i}\sim\pi(x_{i};S)$ . The trajectory $\tau_{i}=(o_{1},a_{1},o_{2},a_{2},\dots,o_{T},a_{T})$ consists of observations $o_{t}$ and actions $a_{t}$ (which may include calls to tools in $\mathcal{U}$ ). The final action $a_{T}$ emits a predicted answer $\hat{y}_{i}$ . The correctness of the prediction is evaluated by a domain-specific scoring function $f(\hat{y}_{i},y_{i})\in[0,1]$ . For any task split $\mathcal{D}_{\text{split}}\subset\mathcal{D}$ , rolling out the agent $\pi(\cdot;S)$ across all task instances in $\mathcal{D}_{\text{split}}$ yields a corresponding set of execution trajectories $\mathcal{T}_{\text{split}}=\{\tau_{i}\sim\pi(x_{i};S)\}_{(x_{i},y_{i})\in\mathcal{D}_{\text{split}}}$ . The performance on a task split $\mathcal{R}(\mathcal{T}_{\text{split}})$ is the average score across all task instances $(x_{i},y_{i})\in\mathcal{D}_{\text{split}}$ .在执行任务 xix_{i} 时，智能体接收任务上下文 xix_{i} 并获得可用技能 SS 的访问权限。智能体使用其工具和技能在多个步骤中与环境交互，生成执行轨迹 τi∼π⁡(xi,S)\tau_{i}\sim\pi(x_{i};S)。轨迹 τi=(o1,a1,o2,a2,…,oT,aT)\tau_{i}=(o_{1},a_{1},o_{2},a_{2},\dots,o_{T},a_{T}) 由观测 oto_{t} 和动作 ata_{t}（可能包括对 𝒰\mathcal{U} 中工具的调用）组成。最终动作 aTa_{T} 输出预测答案 y^i\hat{y}_{i}。预测的正确性由特定领域的评分函数 f⁡(y^i,yi)∈[0,1]f(\hat{y}_{i},y_{i})\in[0,1] 进行评估。对于任何任务划分 𝒟split⊂𝒟\mathcal{D}_{\text{split}}\subset\mathcal{D}，将智能体 π⁡(⋅,S)\pi(\cdot;S) 在 𝒟split\mathcal{D}_{\text{split}} 中的所有任务实例上运行，会产生一组相应的执行轨迹 𝒯split={τi∼π(xi;S)}(xi,yi)∈𝒟split\mathcal{T}_{\text{split}}=\{\tau_{i}\sim\pi(x_{i};S)\}_{(x_{i},y_{i})\in\mathcal{D}_{\text{split}}}。在任务划分上的性能 ℛ⁡(𝒯split)\mathcal{R}(\mathcal{T}_{\text{split}}) 是所有任务实例 (xi,yi)∈𝒟split(x_{i},y_{i})\in\mathcal{D}_{\text{split}} 的平均分数。

In WikiSkill, the system state at iteration $k$ is represented by the tuple $(S_{k},W_{k})$ , where $S_{k}=\{s_{1},\dots,s_{M}\}$ denotes the active procedural skill set and $W_{k}$ denotes the persistent knowledge base (Wiki). While candidate skill updates are subject to validation gating and rollback upon score degradation, the knowledge base $W_{k}$ persists and compounds across iterations. Starting from $(S_{0},W_{0})=(\emptyset,\emptyset)$ , WikiSkill co-evolves the joint state $(S_{k},W_{k})$ across iterations $k\in\{1,\dots,K\}$ , leveraging training rollouts $\mathcal{T}_{\text{train},k}$ , pattern consolidation, and validation gating based on $\mathcal{T}_{\text{val},k}$ to maximize final test performance $\mathcal{R}(\mathcal{T}_{\text{test}})$ on unseen tasks $\mathcal{D}_{\text{test}}$ .在 WikiSkill 中，第 kk 次迭代的系统状态由元组 (Sk,Wk)(S_{k},W_{k}) 表示，其中 Sk={s1,…,sM}S_{k}=\{s_{1},\dots,s_{M}\} 表示主动程序性技能集，WkW_{k} 表示持久化知识库（Wiki）。虽然候选技能更新需经过验证门控并在分数下降时进行回滚，但知识库 WkW_{k} 会在迭代中持续存在并复合。从 (S0,W0)=(∅,∅)(S_{0},W_{0})=(\emptyset,\emptyset) 开始，WikiSkill 在迭代 k∈{1,…,K}k\in\{1,\dots,K\} 中协同演进联合状态 (Sk,Wk)(S_{k},W_{k})，利用训练回放 𝒯train,k\mathcal{T}_{\text{train},k}、模式整合以及基于 𝒯val,k\mathcal{T}_{\text{val},k} 的验证门控，以最大化在未见任务 𝒟test\mathcal{D}_{\text{test}} 上的最终测试性能 ℛ⁡(𝒯test)\mathcal{R}(\mathcal{T}_{\text{test}})。

3 Methodology3 方法论

Refer to caption — Figure 2: Overview of the WikiSkill framework. The agent workspace is structured into three layers: immutable execution traces (Raw Layer), a persistent knowledge base that compounds across iterations (Wiki Layer), and active procedural instructions (Skills Layer). In each evolutionary loop, the Inference Agent runs rollouts (injecting active skills but restricting Wiki access), the Wiki Maintainer consolidates traces into the Wiki, and the Skill Proposer (with the ReAct mechanism) suggests updates while the Wiki is retained across all iterations.图 2：WikiSkill 框架概述。智能体工作空间分为三层：不可变的执行追踪（原始层）、在迭代间复合的持久化知识库（Wiki 层），以及主动程序性指令（技能层）。在每个演进循环中，推理智能体运行回放（注入主动技能但限制 Wiki 访问），Wiki 维护者将追踪整合进 Wiki，技能提议者（配合 ReAct 机制）提出更新，同时 Wiki 在所有迭代中保持不变。

We present WikiSkill, a framework that co-evolves agent skills and a persistent knowledge base (wiki). Built around a three-layer knowledge architecture (§3.1), WikiSkill executes an orchestrated evolutionary loop in which the agent runs rollouts, a Wiki Maintainer consolidates traces and updates the wiki, a Skill Proposer proposes skill updates, and a gating mechanism filters changes (§3.2).我们介绍了 WikiSkill，这是一个将智能体技能与持久化知识库（Wiki）协同演进的框架。WikiSkill 基于三层知识架构（§3.1），执行一个精心设计的演进循环：智能体运行回放，Wiki 维护者整合追踪并更新 Wiki，技能提议者提出技能更新，门控机制过滤变更（§3.2）。

3.1 Three-Layer Knowledge Architecture3.1 三层知识架构

The WikiSkill workspace consists of three distinct layers, as shown in Figure 2 and described below.WikiSkill 工作空间由三个不同的层级组成，如图 2 所示，并在下文中描述。

Raw Layer (raw/)原始层 (raw/)

This layer stores the raw execution traces $\tau_{i}\in\mathcal{T}_{\text{train},k}$ collected from training examples in each iteration. These traces capture the agent’s complete step-by-step interactions, including reasoning, tool calls, tool-call outputs, and final answers. In our setup, the Wiki Maintainer and Skill Proposer agents can access these raw traces to analyze agent behavior. To preserve the raw history, this layer is immutable.该层存储在每次迭代中从训练示例收集的原始执行追踪 τi∈𝒯train,k\tau_{i}\in\mathcal{T}_{\text{train},k}。这些追踪捕获了智能体完整的逐步交互过程，包括推理、工具调用、工具调用输出和最终答案。在我们的设置中，Wiki 维护者和技能提议者智能体可以访问这些原始追踪以分析智能体行为。为了保存原始历史，该层是不可变的。

Wiki Layer (wiki/)Wiki 层 (wiki/)

This layer compiles raw traces into structured, compounding knowledge and is maintained throughout skill evolution. It contains a pattern directory (patterns/) populated with individual markdown files that document specific failure modes or successful strategies, along with actionable workarounds. Crucially, this layer provides long-term historical awareness across optimization iterations through an evolution log (logs.md, updated by the Wiki Maintainer) and a skill impact tracker (skill-impact.md, updated programmatically by the outer-loop harness after validation gating). These records allow the Wiki Maintainer and Skill Proposer to (1) observe the complete skill acceptance history so that rejected interventions are not proposed again, (2) track what was proposed in prior iterations and whether those proposals succeeded, and (3) identify which errors recur across iterations. The wiki is not reset between iterations, but rather accumulates and compiles knowledge continuously throughout the evolution process.该层将原始追踪编译为结构化的、复合的知识，并在整个技能演进过程中进行维护。它包含一个模式目录 (patterns/)，其中填充了记录特定失败模式或成功策略以及可操作变通方案的个人 Markdown 文件。至关重要的是，该层通过演进日志（logs.md，由 Wiki 维护者更新）和技能影响追踪器（skill-impact.md，由外循环工具在验证门控后以编程方式更新）提供跨优化迭代的长期历史感知。这些记录允许 Wiki 维护者和技能提议者 (1) 观察完整的技能接受历史，从而避免再次提出被拒绝的干预措施；(2) 追踪在先前迭代中提出的内容以及这些提议是否成功；(3) 识别跨迭代循环出现的错误。Wiki 在迭代间不会重置，而是在整个演进过程中持续积累和编译知识。

Skills Layer (skills/)技能层 (skills/)

This layer contains the active set of evolved skills $S$ , which encode the procedural knowledge that the Inference Agent can read. Each skill directory in WikiSkill contains two files: SKILL.md, which contains the full content of the skill; and PURPOSE.md, which maps the skill back to the motivating Wiki patterns that inspired its creation or modification. A detailed example of interactions between the Skill Layer and the Wiki Layer is illustrated in Figure 3 and explained in the case study in Section 5.3.该层包含主动演进的技能集 SS，编码了推理智能体可以读取的程序性知识。WikiSkill 中的每个技能目录包含两个文件：SKILL.md（包含技能的全部内容）和 PURPOSE.md（将技能映射回激发其创建或修改的动机 Wiki 模式）。技能层与 Wiki 层之间交互的详细示例见图 3，并在第 5.3 节的案例研究中进行了解释。

3.2 Evolutionary Agents and Wiki Orchestration3.2 演进智能体与 Wiki 编排

The WikiSkill loop consists of four components. In each iteration, the Inference Agent (§3.2.1) executes tasks using the active skills in skills/, producing immutable execution traces in raw/. During the training rollouts, the Inference Agent is restricted from accessing the Wiki Layer, as our ablation study (§5.1) shows that allowing wiki access during training negatively affects skill development. Next, the Wiki Maintainer (§3.2.2) analyzes these raw traces alongside the existing wiki/ layer to diagnose failures and extract successful strategies, updating the persistent pattern catalog and evolution logs. The Skill Proposer (§3.2.3) then reviews the updated wiki and reads execution traces from the latest iteration to generate or modify candidate skills in skills/. Finally, a Gating and Rollback mechanism (§3.2.4) evaluates the candidate skills on a validation split, accepting successful modifications or rolling back the skill set if the changes degrade performance. The entire evolution algorithm is described in Algorithm 1 in Appendix A.WikiSkill 循环由四个组件组成。在每次迭代中，推理智能体（§3.2.1）使用 skills/ 中的主动技能执行任务，在 raw/ 中产生不可变的执行追踪。在训练回放期间，推理智能体被限制访问 Wiki 层，因为我们的消融研究（§5.1）表明，在训练期间允许 Wiki 访问会对技能开发产生负面影响。接下来，Wiki 维护者（§3.2.2）将这些原始追踪与现有的 wiki/ 层结合分析，以诊断失败并提取成功策略，更新持久化模式目录和演进日志。随后，技能提议者（§3.2.3）审查更新后的 Wiki 并从最新迭代中读取执行追踪，以在 skills/ 中生成或修改候选技能。最后，门控与回滚机制（§3.2.4）在验证集上评估候选技能，接受成功的修改，或者如果变更降低了性能，则回滚技能集。整个演进算法在附录 A 的算法 1 中描述。

3.2.1 Skill Provisioning for the Inference Agent3.2.1 为推理智能体提供技能

At iteration $k$ , the Inference Agent $\pi$ is conditioned on the active skill set $S_{k-1}$ and executes a multi-turn trajectory using environment tools $\mathcal{U}$ :在第 kk 次迭代中，推理智能体 π\pi 以主动技能集 Sk−1S_{k-1} 为条件，并使用环境工具 𝒰\mathcal{U} 执行多轮轨迹：

\tau_{i}\sim\pi(x_{i};S_{k-1})

(1)

In WikiSkill, the full content of active skills $S_{k-1}$ is injected directly into the Inference Agent’s system prompt. Following prior work (Yang et al., 2026; Ni et al., 2026), this full-injection setting ensures that procedural instructions are immediately available during task execution, thereby eliminating skill triggering or retrieval failures as confounding variables in our study.在 WikiSkill 中，主动技能 Sk−1S_{k-1} 的全部内容被直接注入到推理智能体的系统提示词中。遵循以往的研究（Yang 等人，2026；Ni 等人，2026），这种全注入设置确保了程序性指令在任务执行期间立即可用，从而消除了技能触发或检索失败作为我们研究中的混杂变量。

3.2.2 Wiki Maintainer: Pattern Consolidation3.2.2 Wiki 维护者：模式整合

At iteration $k$ , after obtaining rollout traces $\mathcal{T}_{\text{train},k}$ on the training split $\mathcal{D}_{\text{train}}$ , we sample a subset of successful and failing execution traces $\mathcal{T}_{\text{sample},k}\subset\mathcal{T}_{\text{train},k}$ (see Appendix C for sampling budget and stratification criteria) to avoid context window limitations. The Wiki Maintainer agent $\mathcal{M}_{\text{WM}}$ consolidates these observations into the persistent wiki $W_{k-1}$ , producing the intermediate wiki state $W^{\prime}_{k}$ :在第 kk 次迭代中，在获得训练划分 𝒟train\mathcal{D}_{\text{train}} 上的回放追踪 𝒯train,k\mathcal{T}_{\text{train},k} 后，我们采样一部分成功和失败的执行追踪 𝒯sample,k⊂𝒯train,k\mathcal{T}_{\text{sample},k}\subset\mathcal{T}_{\text{train},k}（采样预算和分层标准见附录 C），以避免上下文窗口限制。Wiki 维护者智能体 ℳWM\mathcal{M}_{\text{WM}} 将这些观察结果整合进持久化 Wiki Wk−1W_{k-1} 中，产生中间 Wiki 状态 Wk′W^{\prime}_{k}：

W^{\prime}_{k}\leftarrow\mathcal{M}_{\text{WM}}(W_{k-1},\mathcal{T}_{\text{sample},k})

(2)

The Wiki Maintainer agent receives the full wiki context $W_{k-1}$ alongside sampled traces $\mathcal{T}_{\text{sample},k}$ . It performs root cause analysis on the failing tasks, and extracts successful strategies from the passing tasks. In each iteration, the Wiki Maintainer can create new pattern pages under wiki/patterns/ and update existing pattern pages with new evidence or refined solutions. Updates to pattern pages are applied using incremental, patch-based editing (e.g., appending, replacing, or inserting text spans). Whenever patterns are modified, the Wiki Maintainer revises the index.md catalog to reflect the current state and appends a summary of the iteration’s findings to the evolution log logs.md. There is no hard limit on the number of patterns created or updated per iteration; the Wiki Maintainer decides what updates are warranted based on the traces and the current wiki state.Wiki 维护者智能体接收完整的 Wiki 上下文 Wk−1W_{k-1} 以及采样的追踪 𝒯sample,k\mathcal{T}_{\text{sample},k}。它对失败任务进行根本原因分析，并从成功任务中提取成功策略。在每次迭代中，Wiki 维护者可以在 wiki/patterns/ 下创建新的模式页面，并使用新证据或提炼后的解决方案更新现有的模式页面。模式页面的更新使用增量式、基于补丁的编辑（例如附加、替换或插入文本片段）应用。每当模式被修改，Wiki 维护者都会修订 index.md 目录以反映当前状态，并将该次迭代的发现摘要附加到演进日志 logs.md 中。每次迭代中创建或更新的模式数量没有硬性限制；Wiki 维护者根据追踪和当前的 Wiki 状态决定需要进行哪些更新。

3.2.3 Wiki-Informed Skill Proposer3.2.3 基于 Wiki 的技能提议者

The Proposer $\mathcal{M}_{\text{P}}$ is an LLM-based agent responsible for skill discovery and refinement. At iteration $k$ , the proposer operates in a multi-turn ReAct style (Yao et al., 2023). To avoid context window exhaustion when analyzing long execution histories, the proposer is not given a fixed set of pre-sampled traces; instead, it is initially provided with the wiki index $I(W^{\prime}_{k})$ , the historical skill impact tracker (skill-impact.md), and a concise summary of all training task outcomes (pass/fail status, predictions and ground-truth answers). Operating as an autonomous agent, it actively reasons and uses environment tools (read_file) to select and inspect specific pattern pages and raw execution traces $\tau_{i}\in\mathcal{T}_{\text{train},k}$ on demand to diagnose root causes before synthesizing a proposal $P_{k}$ :提议者 ℳP\mathcal{M}_{\text{P}} 是一个基于 LLM 的智能体，负责技能发现和提炼。在第 kk 次迭代中，提议者以多轮 ReAct 风格运行（Yao 等人，2023）。为了在分析长执行历史时避免上下文窗口耗尽，提议者不会被给予固定的预采样追踪集；相反，它最初会被提供 Wiki 索引 I⁡(Wk′)I(W^{\prime}_{k})、历史技能影响追踪器 (skill-impact.md) 以及所有训练任务结果的简要摘要（通过/失败状态、预测和真实答案）。作为自主智能体，它会主动推理并使用环境工具 (read_file) 按需选择和检查特定的模式页面以及原始执行追踪 τi∈𝒯train,k\tau_{i}\in\mathcal{T}_{\text{train},k}，在合成提议 PkP_{k} 之前诊断根本原因：

P_{k}\leftarrow\mathcal{M}_{\text{P}}(W^{\prime}_{k},S_{k-1},\mathcal{T}_{\text{train},k})

(3)

In each iteration, the Skill Proposer produces an atomic proposal $P_{k}$ that targets a single skill, either creating a new skill or applying an incremental, patch-based edit to the targeted existing skill.在每次迭代中，技能提议者产生一个针对单个技能的原子提议 PkP_{k}，要么创建新技能，要么对目标现有技能应用增量式、基于补丁的编辑。

3.2.4 Gating and Rollback3.2.4 门控与回滚

Once a proposal $P_{k}$ is generated, it is applied to the workspace to yield a candidate skill set $S^{\prime}_{k}=\text{Apply}(S_{k-1},P_{k})$ . The system evaluates $S^{\prime}_{k}$ on the validation split $\mathcal{D}_{\text{val}}$ , obtaining validation traces $\mathcal{T}_{\text{val},k}$ and score $\mathcal{R}(\mathcal{T}_{\text{val},k})$ . The acceptance decision is governed by:一旦提议 PkP_{k} 生成，它就会被应用到工作空间以产生候选技能集 Sk′=Apply(Sk−1,Pk)S^{\prime}_{k}=\text{Apply}(S_{k-1},P_{k})。系统在验证划分 𝒟val\mathcal{D}_{\text{val}} 上评估 Sk′S^{\prime}_{k}，获得验证追踪 𝒯val,k\mathcal{T}_{\text{val},k} 和分数 ℛ⁡(𝒯val,k)\mathcal{R}(\mathcal{T}_{\text{val},k})。接受决策由以下规则决定：

S_{k}\leftarrow\begin{cases}S^{\prime}_{k}&\text{if }\mathcal{R}(\mathcal{T}_{\text{val},k})>\mathcal{R}_{\text{best}}\\ S_{k-1}&\text{otherwise}\end{cases}

(4)

If accepted, the candidate skills are preserved as the new active skill set $S_{k}$ , and the benchmark performance threshold $\mathcal{R}_{\text{best}}$ is updated to $\mathcal{R}(\mathcal{T}_{\text{val},k})$ . Prior to the evolution loop, $\mathcal{R}_{\text{best}}$ is initialized to the baseline validation score $\mathcal{R}(\mathcal{T}_{\text{val},0})$ obtained by evaluating the empty skill set $S_{0}$ on $\mathcal{D}_{\text{val}}$ . If the validation score reaches the maximum ( $\mathcal{R}_{\text{best}}=1.0$ ) at any point during evolution, the evolution loop terminates early. If rejected, the system discards the candidate skill modifications and reverts the skill set to the most recent successful configuration $S_{k-1}$ . Notably, the wiki $W_{k}$ is never rolled back regardless of the acceptance decision; accumulated patterns and logs persist across all iterations to ensure long-term knowledge retention. Following each validation evaluation, the outer-loop orchestration harness programmatically appends an entry to wiki/skill-impact.md via $W_{k}\leftarrow\text{Update}(W^{\prime}_{k},P_{k},\mathcal{R}(\mathcal{T}_{\text{val},k}),a_{k})$ , recording the proposal metadata, target skill name, unified diff of the modification, validation score $\mathcal{R}(\mathcal{T}_{\text{val},k})$ , and final acceptance outcome $a_{k}\in\{\text{Accepted},\text{Rejected}\}$ . This completes the wiki state transition $W_{k-1}\to W_{k}$ for iteration $k$ , providing an objective, ground-truth audit trail of past interventions that the Skill Proposer can consult in subsequent iterations to avoid repeating failed modifications.如果被接受，候选技能将作为新的主动技能集 SkS_{k} 保留，基准性能阈值 ℛbest\mathcal{R}_{\text{best}} 更新为 ℛ⁡(𝒯val,k)\mathcal{R}(\mathcal{T}_{\text{val},k})。在演进循环之前，ℛbest\mathcal{R}_{\text{best}} 初始化为通过在 𝒟val\mathcal{D}_{\text{val}} 上评估空技能集 S0S_{0} 获得的基线验证分数 ℛ⁡(𝒯val,0)\mathcal{R}(\mathcal{T}_{\text{val},0})。如果验证分数在演进过程中的任何时刻达到最大值 (ℛbest=1.0\mathcal{R}_{\text{best}}=1.0)，演进循环提前终止。如果被拒绝，系统将丢弃候选技能修改，并将技能集恢复为最近一次成功的配置 Sk−1S_{k-1}。值得注意的是，无论接受决策如何，Wiki WkW_{k} 永远不会回滚；积累的模式和日志在所有迭代中持续存在，以确保长期的知识留存。在每次验证评估后，外循环编排工具会通过 Wk←Update(Wk′,Pk,ℛ⁡(𝒯val,k),ak)W_{k}\leftarrow\text{Update}(W^{\prime}_{k},P_{k},\mathcal{R}(\mathcal{T}_{\text{val},k}),a_{k}) 以编程方式向 wiki/skill-impact.md 附加条目，记录提议元数据、目标技能名称、修改的统一差异、验证分数 ℛ⁡(𝒯val,k)\mathcal{R}(\mathcal{T}_{\text{val},k}) 以及最终接受结果 ak∈{Accepted,Rejected}a_{k}\in\{\text{Accepted},\text{Rejected}\。这完成了迭代 kk 的 Wiki 状态转换 Wk−1→WkW_{k-1}\to W_{k}，提供了过去干预措施的客观、真实审计追踪，技能提议者可以在后续迭代中查阅该追踪以避免重复失败的修改。

4 Experiments and Results4 实验与结果

4.1 Experimental Setup4.1 实验设置

Datasets数据集

We evaluate across five benchmarks spanning diverse domains: mathematical reasoning (LiveMathematicianBench (LiveMath) (He et al., 2026)), web search (SealQA (Pham et al., 2026)), spreadsheet manipulation (SpreadsheetBench (SpreadSheet) (Ma et al., 2024)), long-context document question answering OfficeQA (Singhvi et al., 2025)), and interactive embodied tasks (ALFWorld (Shridhar et al., 2021)). Dataset details and statistics are provided in Appendix B.我们在涵盖不同领域的五个基准测试上进行了评估：数学推理 (LiveMathematicianBench (LiveMath) (He 等人，2026))、网络搜索 (SealQA (Pham 等人，2026))、电子表格操作 (SpreadsheetBench (SpreadSheet) (Ma 等人，2024))、长上下文文档问答 (OfficeQA (Singhvi 等人，2025)) 以及交互式具身任务 (ALFWorld (Shridhar 等人，2021))。数据集详情和统计数据见附录 B。

Baselines基线

We compare WikiSkill against three representative skill-evolution baselines, including Trace2Skill (Ni et al., 2026), EvoSkill (Alzubi et al., 2026), and SkillOpt (Yang et al., 2026), all of which share the same general loop of rolling out an agent, analyzing execution traces, proposing skill modifications, and gating changes via validation. We also evaluate each model without skills as a no-skill baseline. A detailed description and an analysis of the complexity of optimizer API calls across these frameworks are provided in Appendix D. We focus our comparison on dedicated skill-evolution frameworks rather than general automatic prompt optimizers (e.g., GEPA (Agrawal et al., 2026)), following prior work that shows specialized skill-evolution pipelines consistently outperform general prompt optimization methods (Yang et al., 2026).我们将 WikiSkill 与三种代表性的技能演进基线进行了比较，包括 Trace2Skill (Ni 等人，2026)、EvoSkill (Alzubi 等人，2026) 和 SkillOpt (Yang 等人，2026)，所有这些基线都共享相同的通用循环：运行智能体、分析执行追踪、提出技能修改，并通过验证门控变更。我们还评估了每个模型在不使用技能的情况下的表现，作为无技能基线。关于这些框架中优化器 API 调用复杂性的详细描述和分析见附录 D。我们重点比较专门的技能演进框架，而非通用的自动提示词优化器（例如 GEPA (Agrawal 等人，2026)），遵循以往的研究，这些研究表明专门的技能演进流水线始终优于通用提示词优化方法 (Yang 等人，2026)。

Models模型

We experiment with both closed and open-weight models to evaluate WikiSkill and the baselines. For closed models, we use Gemini-3.5-Flash (Google DeepMind, 2026). For open-weight models, we evaluate Qwen-3.5-4B/9B-Instruct (Qwen Team, 2026a), Qwen-3.6-27B (Qwen Team, 2026b), and Gemma-4-31B-It (Gemma Team, 2026), which we deploy using the vLLM framework (Kwon et al., 2023).我们使用闭源和开源权重模型来评估 WikiSkill 和基线。对于闭源模型，我们使用 Gemini-3.5-Flash (Google DeepMind, 2026)。对于开源权重模型，我们评估了 Qwen-3.5-4B/9B-Instruct (Qwen Team, 2026a)、Qwen-3.6-27B (Qwen Team, 2026b) 和 Gemma-4-31B-It (Gemma Team, 2026)，并使用 vLLM 框架 (Kwon 等人，2023) 进行部署。

4.2 Main Results4.2 主要结果

Model	Method	LiveMath	SealQA	SpreadSheet	OfficeQA	ALFWorld	Avg.
Qwen-3.5-4B	No skill	29.1	32.5	14.6	30.2	24.4	26.2
	Trace2Skill	31.5	37.6	17.5	31.0	42.8	32.1
	EvoSkill	41.7	37.3	18.6	29.5	41.5	33.7
	SkillOpt	48.7	33.3	14.0	34.5	45.3	35.2
	WikiSkill	49.7	39.4	21.1	28.5	53.7	38.5
Qwen-3.5-9B	No skill	28.2	26.3	24.3	35.9	34.7	29.9
	Trace2Skill	33.1	36.9	26.5	38.4	48.8	36.7
	EvoSkill	58.1	34.5	35.4	34.9	48.5	42.3
	SkillOpt	48.7	29.4	29.0	38.0	55.7	40.2
	WikiSkill	56.3	43.1	33.6	40.5	63.4	47.4
Qwen-3.6-27B	No skill	33.9	27.5	40.8	42.1	52.8	39.4
	Trace2Skill	36.3	37.3	53.3	54.3	55.5	47.3
	EvoSkill	57.3	32.9	59.5	52.5	64.2	53.3
	SkillOpt	51.9	34.5	53.2	54.8	59.2	50.7
	WikiSkill	61.9	41.6	81.7	53.7	77.6	63.3
Gemma-4-31B	No skill	33.9	30.6	48.3	43.3	50.4	41.3
	Trace2Skill	32.3	37.7	58.5	43.2	57.2	45.8
	EvoSkill	29.8	38.4	56.4	39.9	52.6	43.4
	SkillOpt	40.1	36.1	63.1	44.4	61.9	49.1
	WikiSkill	56.7	41.2	68.0	44.2	64.4	54.9
Gemini-3.5-Flash	No skill	33.0	29.4	50.5	48.6	85.9	49.5
	Trace2Skill	41.9	44.3	56.0	50.0	85.9	55.6
	EvoSkill	44.6	43.6	55.4	51.2	85.9	56.1
	SkillOpt	49.7	28.2	66.1	49.8	85.9	55.9
	WikiSkill	72.6	44.7	76.6	60.7	85.9	68.1

Table 1: Method comparison across inference models and test sets. Each horizontal block evaluates a specific inference model without skills (No skill) and with skills developed by different skill-evolution methods. To ensure a fair comparison, all skill-evolution methods start with an empty skill set, and evolved skills are injected into the Inference Agent’s prompt at inference time. All reported scores are the average test performance across three independent runs of the full evolution process. Our method (WikiSkill) is highlighted. Bold indicates the best performance for each dataset; multiple bold results indicate methods that are not significantly different from the best under a paired bootstrap test with 1,000 iterations (

p<0.05

). 表 1：跨推理模型和测试集的方法比较。每个水平块评估了一个特定的推理模型在无技能（No skill）以及使用不同技能演进方法开发的技能下的表现。为确保公平比较，所有技能演进方法均从空技能集开始，演进后的技能在推理时被注入到推理智能体的提示词中。所有报告的分数均为完整演进过程的三次独立运行的平均测试性能。我们的方法（WikiSkill）已突出显示。粗体表示每个数据集的最佳性能；多个粗体结果表示在 1,000 次迭代的配对自助法检验下，与最佳结果无显著差异的方法 (p<0.05p<0.05)。

We evaluate WikiSkill across models and tasks and study whether evolved skills transfer across models. Table 1 presents the main skill-evolution results across models and tasks, including how the benefits of skill evolution vary with model scale, while Table 2 presents the cross-model skill transfer results. We analyze these results in detail below.我们评估了 WikiSkill 在不同模型和任务上的表现，并研究了演进后的技能是否可以在模型间迁移。表 1 展示了跨模型和任务的主要技能演进结果，包括技能演进的收益如何随模型规模变化，表 2 则展示了跨模型技能迁移的结果。我们在下文中详细分析这些结果。

To account for variability, we repeat the full evolution process across three independent runs for each method, and all reported scores represent the average test performance across the three resulting evolved skill sets. Statistical significance of performance differences is evaluated using paired bootstrap testing at $p<0.05$ (Appendix C). Note that for Gemini-3.5-Flash on ALFWorld, all evolution methods yield the same performance (85.9%) as the no-skill baseline because Gemini-3.5-Flash achieves a 100% score on the validation split ( $\mathcal{D}_{\text{val}}$ ) before skill evolution. This also explains why Gemini-3.5-Flash is marked with ‘ $-$ ’ as a skill source on ALFWorld in the cross-model transfer evaluation (Table 2).为了考虑变异性，我们对每种方法的完整演进过程重复了三次独立运行，所有报告的分数代表三次所得演进技能集的平均测试性能。性能差异的统计显著性使用配对自助法在 p<0.05p<0.05 下进行评估（附录 C）。注意，对于 ALFWorld 上的 Gemini-3.5-Flash，所有演进方法均产生与无技能基线相同的性能 (85.9%)，因为 Gemini-3.5-Flash 在技能演进前就在验证划分 (𝒟val\mathcal{D}_{\text{val}}) 上达到了 100% 的分数。这也解释了为什么在跨模型迁移评估（表 2）中，Gemini-3.5-Flash 在 ALFWorld 上被标记为‘−-’作为技能来源。

4.2.1 Skill Evolution Across Models and Tasks4.2.1 跨模型和任务的技能演进

WikiSkill yields consistent improvements across models and datasetsWikiSkill 在模型和数据集上均实现了持续提升

As shown in Table 1, WikiSkill achieves the highest average performance across all five models. Compared with the strongest competing skill-evolution method for each model, WikiSkill improves average performance by 3.3, 5.1, 10.0, 5.8, and 12.0 points for Qwen-3.5-4B, Qwen-3.5-9B, Qwen-3.6-27B, Gemma-4-31B, and Gemini-3.5-Flash, respectively. These improvements are consistent across settings: WikiSkill improves over the no-skill baseline in most model-dataset pairs and matches or exceeds the strongest competing method across all models in the average performance across 5 datasets. The improvements also span diverse domains and can be substantial. For example, WikiSkill improves Gemini-3.5-Flash from 33.0% to 72.6% on LiveMath and from 50.5% to 76.6% on SpreadSheet, while improving Qwen-3.6-27B from 52.8% to 77.6% on ALFWorld. In contrast, existing skill-evolution methods are less consistent. For example, EvoSkill improves Qwen-9B substantially on LiveMath (28.2% $\rightarrow$ 58.1%) but degrades Gemma-4-31B on the same benchmark (33.9 % $\rightarrow$ 29.8%), while SkillOpt degrades Gemini-3.5-Flash on SealQA (29.4 % $\rightarrow$ 28.2%). These results show that WikiSkill produces both stronger and more reliable improvements across settings.如表 1 所示，WikiSkill 在所有五个模型上均实现了最高的平均性能。与每个模型最强的竞争性技能演进方法相比，WikiSkill 在 Qwen-3.5-4B、Qwen-3.5-9B、Qwen-3.6-27B、Gemma-4-31B 和 Gemini-3.5-Flash 上的平均性能分别提升了 3.3、5.1、10.0、5.8 和 12.0 个点。这些提升在不同场景下是一致的：WikiSkill 在大多数模型-数据集对中均优于无技能基线，且在 5 个数据集的平均性能上，在所有模型中均匹配或超过了最强的竞争方法。这些提升涵盖了不同的领域，且可能非常显著。例如，WikiSkill 将 Gemini-3.5-Flash 在 LiveMath 上的表现从 33.0% 提升至 72.6%，在 SpreadSheet 上从 50.5% 提升至 76.6%，同时将 Qwen-3.6-27B 在 ALFWorld 上的表现从 52.8% 提升至 77.6%。相比之下，现有的技能演进方法一致性较差。例如，EvoSkill 在 LiveMath 上显著提升了 Qwen-9B (28.2% →\rightarrow 58.1%)，但在同一基准上降低了 Gemma-4-31B 的表现 (33.9 % →\rightarrow 29.8%)，而 SkillOpt 在 SealQA 上降低了 Gemini-3.5-Flash 的表现 (29.4 % →\rightarrow 28.2%)。这些结果表明，WikiSkill 在不同场景下均能产生更强且更可靠的提升。

The benefits of skill evolution increase with model capability and complement model scaling技能演进的收益随模型能力增强而增加，并与模型扩展互补

Within the Qwen family, the average improvement from WikiSkill increases with model scale, from +12.3 points for Qwen-3.5-4B to +17.5 points for Qwen-3.5-9B and +23.9 points for Qwen-3.6-27B. This trend is particularly pronounced on SpreadSheet, where WikiSkill improves the three models by +6.5, +9.3, and +40.9 points, respectively, showing that the benefits of skill evolution can increase substantially with model scale. At the same time, evolved skills can compensate for substantial differences in model scale: Qwen-3.5-9B with WikiSkill reaches 47.4% average accuracy, outperforming Qwen-3.6-27B without skills at 39.4%, while Qwen-3.5-4B with WikiSkill reaches 38.5%. Our results suggest that model capability and evolved procedural knowledge provide complementary sources of performance: stronger models can derive greater value from skill evolution by developing and executing more effective skills, while effective skills can allow smaller models to outperform substantially larger models that do not use skills.在 Qwen 系列中，WikiSkill 带来的平均提升随模型规模增大而增加，从 Qwen-3.5-4B 的 +12.3 个点到 Qwen-3.5-9B 的 +17.5 个点，再到 Qwen-3.6-27B 的 +23.9 个点。这一趋势在 SpreadSheet 上尤为明显，WikiSkill 分别将三个模型的表现提升了 +6.5、+9.3 和 +40.9 个点，表明技能演进的收益可以随模型规模显著增加。同时，演进后的技能可以弥补模型规模的巨大差异：使用 WikiSkill 的 Qwen-3.5-9B 平均准确率达到 47.4%，超过了无技能的 Qwen-3.6-27B (39.4%)，而使用 WikiSkill 的 Qwen-3.5-4B 达到了 38.5%。我们的结果表明，模型能力和演进后的程序性知识提供了互补的性能来源：更强的模型可以通过开发和执行更有效的技能从技能演进中获得更大的价值，而有效的技能可以让较小的模型超越不使用技能的大型模型。

The benefits of skill evolution also vary substantially across datasets技能演进的收益在不同数据集间也存在显著差异

Our results suggest that some datasets are more amenable to skill evolution than others. For Qwen-3.6-27B, WikiSkill improves performance by 11.6 points on OfficeQA and 14.1 points on SealQA, compared with 24.8 points on ALFWorld, 28.0 points on LiveMath, and 40.9 points on SpreadSheet. Similar differences appear across other models. LiveMath consistently benefits from skill evolution, with gains ranging from 20.6 to 39.6 points across all five models, while ALFWorld yields gains of 14.0 to 29.3 points across the four models for which WikiSkill evolves skills (excluding Gemini-3.5-Flash due to early stopping). In contrast, OfficeQA presents unique challenges due to its long-context document-retrieval requirements. Larger models effectively leverage evolved search workflows to navigate lengthy documents (e.g., +11.6 points for Qwen-3.6-27B and +12.1 points for Gemini-3.5-Flash), whereas Qwen-3.5-4B struggles to execute these multi-step search workflows across long contexts and reverts to its default reading behavior, resulting in slight degradation.我们的结果表明，某些数据集比其他数据集更适合技能演进。对于 Qwen-3.6-27B，WikiSkill 在 OfficeQA 上将性能提升了 11.6 个点，在 SealQA 上提升了 14.1 个点，相比之下在 ALFWorld 上提升了 24.8 个点，在 LiveMath 上提升了 28.0 个点，在 SpreadSheet 上提升了 40.9 个点。其他模型也存在类似的差异。LiveMath 始终从技能演进中受益，所有五个模型的增益范围在 20.6 到 39.6 个点之间，而 ALFWorld 在 WikiSkill 演进技能的四个模型中（因早停排除了 Gemini-3.5-Flash）产生了 14.0 到 29.3 个点的增益。相比之下，OfficeQA 由于其长上下文文档检索需求而提出了独特的挑战。更大的模型有效地利用演进后的搜索工作流来浏览长文档（例如 Qwen-3.6-27B 提升了 +11.6 个点，Gemini-3.5-Flash 提升了 +12.1 个点），而 Qwen-3.5-4B 难以在长上下文中执行这些多步搜索工作流，并退回到其默认阅读行为，导致轻微的性能下降。

Model	Skill Source	LiveMath	SealQA	SpreadSheet	OfficeQA	ALFWorld
Qwen-3.5-4B	None	29.1	32.5	14.6	30.2	24.4
	Qwen-3.5-4B	49.7	39.4	21.1	28.5	53.7
	Qwen-3.6-27B	59.7	38.8	33.0	25.4	57.0
	Gemini-3.5-Flash	62.6	37.3	23.0	32.2	-
Qwen-3.5-9B	None	28.2	26.3	24.3	35.9	34.7
	Qwen-3.5-4B	61.0	40.4	25.0	40.3	69.2
	Qwen-3.5-9B	56.3	43.1	33.6	40.5	63.4
	Qwen-3.6-27B	59.1	40.4	50.5	39.9	70.2
	Gemini-3.5-Flash	53.0	39.6	48.8	40.5	-
Qwen-3.6-27B	None	33.9	27.5	40.8	42.1	52.8
	Qwen-3.5-4B	62.6	41.6	40.6	52.9	72.1
	Qwen-3.6-27B	61.9	41.6	81.7	53.7	77.6
	Gemini-3.5-Flash	65.1	51.0	76.0	52.5	-
Gemma-4-31B	None	33.9	30.6	48.3	43.3	50.4
	Qwen-3.5-4B	73.1	38.8	37.1	42.1	66.9
	Qwen-3.6-27B	73.7	37.7	72.0	44.2	66.9
	Gemma-4-31B	56.7	41.2	68.0	44.2	64.4
	Gemini-3.5-Flash	61.8	37.7	68.8	43.4	-
Gemini-3.5-Flash	None	33.0	29.4	50.5	48.6	85.9
	Qwen-3.5-4B	67.5	40.0	18.1	48.5	87.3
	Qwen-3.6-27B	73.9	43.5	63.4	47.7	86.8
	Gemini-3.5-Flash	72.6	44.7	76.6	60.7	-

Table 2: Cross-model skill transfer results. We evaluate inference models using no skills (None) and skills evolved by WikiSkill with Qwen-3.5-4B, Qwen-3.6-27B, and Gemini-3.5-Flash as source models. Skills are injected into the Inference Agent’s system prompt at inference time. Highlighted rows indicate self-evolved skills, where the inference model and skill source are the same. The highest performance per benchmark within each model block is bolded. ‘

-

’ indicates that the source model reached 100% validation performance before skill evolution, so no skill was evolved. 表 2：跨模型技能迁移结果。我们评估了使用无技能（None）以及由 WikiSkill 使用 Qwen-3.5-4B、Qwen-3.6-27B 和 Gemini-3.5-Flash 作为源模型演进的技能的推理模型。技能在推理时被注入到推理智能体的系统提示词中。突出显示的行表示自身演进的技能，即推理模型和技能来源相同。每个模型块内每个基准测试的最高性能已加粗。‘−-’表示源模型在技能演进前达到了 100% 的验证性能，因此没有演进出技能。

4.2.2 Cross-Model Skill Transfer with WikiSkill4.2.2 使用 WikiSkill 进行跨模型技能迁移

Evolved skills transfer effectively across models, and transferred skills can outperform self-evolved skills演进后的技能可以在模型间有效迁移，且迁移后的技能可以优于自身演进的技能

Table 2 evaluates how skills evolved by WikiSkill transfer across inference models when developed using different source models. Transferred skills frequently outperform both the no-skill baseline and self-evolved skills. For example, Qwen-3.6-27B skills improve Qwen-3.5-9B to 50.5% on SpreadSheet, compared with 24.3% without skills and 33.6% with self-evolved skills, and improve Gemma-4-31B to 73.7% on LiveMath, compared with 33.9% and 56.7%, respectively. Notably, effective transfer also occurs from smaller to larger models: Qwen-3.5-4B skills improve Gemma-4-31B to 73.1% on LiveMath and 66.9% on ALFWorld. Our results indicate that stronger source models do not necessarily produce better skills and that procedural knowledge developed by one model’s experience can transfer across model scales and families.表 2 评估了由 WikiSkill 演进并在使用不同源模型开发时，技能如何在推理模型间迁移。迁移后的技能经常优于无技能基线和自身演进的技能。例如，Qwen-3.6-27B 的技能将 Qwen-3.5-9B 在 SpreadSheet 上的表现提升至 50.5%，而无技能时为 24.3%，使用自身演进技能时为 33.6%；并将 Gemma-4-31B 在 LiveMath 上的表现提升至 73.7%，而无技能时为 33.9%，自身演进技能时为 56.7%。值得注意的是，有效的迁移也发生在从较小模型到较大模型：Qwen-3.5-4B 的技能将 Gemma-4-31B 在 LiveMath 上的表现提升至 73.1%，在 ALFWorld 上提升至 66.9%。我们的结果表明，更强的源模型不一定会产生更好的技能，且由一个模型的经验开发的程序性知识可以在模型规模和系列间迁移。

The transferability of evolved skills depends on whether they capture general procedures or model-specific workarounds演进后技能的可迁移性取决于它们是捕获了通用程序还是模型特定的变通方案

Our results in Table 2 suggest that WikiSkill can produce both general procedural knowledge that transfers across models and model-specific strategies that can cause negative transfer. LiveMath skills transfer particularly well across models: Qwen-3.5-4B and Qwen-3.6-27B skills improve Gemini-3.5-Flash from 33.0% to 67.5% and 73.9%, respectively. In contrast, SpreadSheet exhibits strong source-target interactions. Qwen-3.5-4B skills reduce Gemini-3.5-Flash performance from 50.5% to 18.1%, while Qwen-3.6-27B skills improve it to 63.4%. Our error analysis identifies two factors behind this negative transfer. First, Qwen-3.5-4B skills encode low-level workarounds, such as single-line Python commands and string-conversion rules, which help the smaller model avoid execution failures but constrain stronger models such as Gemini-3.5-Flash from using comprehensive end-to-end scripts. Second, fragmented diagnostic procedures introduce redundant tool calls that can exhaust Gemini-3.5-Flash’s interaction budget before task completion.我们在表 2 中的结果表明，WikiSkill 既能产生跨模型迁移的通用程序性知识，也能产生可能导致负迁移的模型特定策略。LiveMath 技能在模型间迁移效果特别好：Qwen-3.5-4B 和 Qwen-3.6-27B 的技能分别将 Gemini-3.5-Flash 的表现从 33.0% 提升至 67.5% 和 73.9%。相比之下，SpreadSheet 展示了强烈的源-目标交互。Qwen-3.5-4B 的技能将 Gemini-3.5-Flash 的性能从 50.5% 降低至 18.1%，而 Qwen-3.6-27B 的技能将其提升至 63.4%。我们的错误分析确定了导致这种负迁移的两个因素。首先，Qwen-3.5-4B 的技能编码了低级变通方案，如单行 Python 命令和字符串转换规则，这有助于较小模型避免执行失败，但限制了 Gemini-3.5-Flash 等更强模型使用全面的端到端脚本。其次，碎片化的诊断程序引入了冗余的工具调用，可能会在任务完成前耗尽 Gemini-3.5-Flash 的交互预算。

The utility of transferred skills also depends on the inference model’s ability to execute them迁移技能的效用也取决于推理模型执行它们的能力

We now turn toward how different inference models use skills developed by the same source model. Within the Qwen family, stronger models can derive greater value from the same procedural knowledge. For example, Qwen-3.6-27B SpreadSheet skills improve Qwen-3.5-4B, Qwen-3.5-9B, and Qwen-3.6-27B over their no-skill baselines by 18.4%, 26.2%, and 40.9%, respectively. OfficeQA provides a case where a model develops skills that are more useful to another model than to itself: Qwen-3.5-4B skills decrease its own performance from 30.2% to 28.5%, but improve Qwen-3.6-27B from 42.1% to 52.9%. Our trajectory analysis suggests that in long-context settings, smaller models can become distracted by lengthy document contexts and fail to follow detailed multi-step search instructions, instead reverting to their default document-reading behavior. Stronger models, in contrast, more reliably execute the structured navigation procedures specified by the skill across long contexts. Taken together, these results distinguish two capabilities that self-evolution normally conflates: discovering useful procedural knowledge from experience and effectively executing that knowledge at inference time.我们现在转向不同的推理模型如何使用由相同源模型开发的技能。在 Qwen 系列中，更强的模型可以从相同的程序性知识中获得更大的价值。例如，Qwen-3.6-27B 的 SpreadSheet 技能使 Qwen-3.5-4B、Qwen-3.5-9B 和 Qwen-3.6-27B 的性能分别比其无技能基线提升了 18.4%、26.2% 和 40.9%。OfficeQA 提供了一个案例，即一个模型开发的技能对另一个模型比对自己更有用：Qwen-3.5-4B 的技能将其自身性能从 30.2% 降至 28.5%，但将 Qwen-3.6-27B 的性能从 42.1% 提升至 52.9%。我们的轨迹分析表明，在长上下文场景中，较小模型可能会被冗长的文档上下文分散注意力，无法遵循详细的多步搜索指令，而是退回到其默认的文档阅读行为。相比之下，更强的模型能更可靠地执行技能在长上下文中指定的结构化导航程序。综合来看，这些结果区分了自演进通常混淆的两种能力：从经验中发现有用的程序性知识，以及在推理时有效地执行该知识。

5 Analysis and Discussion5 分析与讨论

5.1 Role of Persistent Knowledge in Skill Evolution5.1 持久化知识在技能演进中的作用

To understand where persistent knowledge contributes to skill evolution, we ablate wiki access for the two components that can use it during evolution (Table 3). Specifically, using Gemini-3.5-Flash, we independently vary wiki access for the Inference Agent during training rollouts and the Skill Proposer during skill development, which results in four configurations. When the Skill Proposer has no wiki access, we also remove the Wiki Maintainer, eliminating persistent knowledge accumulation across iterations. Our default WikiSkill configuration gives wiki access to the Skill Proposer but not the Inference Agent.为了理解持久化知识在何处对技能演进做出贡献，我们对演进过程中可以使用它的两个组件进行了 Wiki 访问的消融研究（表 3）。具体而言，使用 Gemini-3.5-Flash，我们独立改变了推理智能体在训练回放期间和技能提议者在技能开发期间的 Wiki 访问权限，这产生了四种配置。当技能提议者没有 Wiki 访问权限时，我们也移除了 Wiki 维护者，消除了跨迭代的持久化知识积累。我们的默认 WikiSkill 配置允许技能提议者访问 Wiki，但不允许推理智能体访问。

WikiSkill Components		Benchmarks
Inference Agent	Skill Proposer	LiveMath	SealQA	SpreadSheet	OfficeQA	Avg.
Wiki Access	Wiki Access
No skill		33.0	29.4	50.5	48.6	40.4
		43.8	42.0	44.4	51.0	45.3
		51.3	38.4	49.9	55.2	48.7
		64.8	42.8	80.2	55.6	60.9
		72.6	44.7	76.6	60.7	63.7

Table 3: Ablation study on WikiSkill using Gemini-3.5-Flash. We evaluate performance across benchmarks under four configurations that vary whether the Inference Agent and Skill Proposer have wiki access during skill evolution. When the Skill Proposer has no Wiki access, we also remove the Wiki Maintainer, eliminating persistent knowledge accumulation across iterations. The bottom row represents our default WikiSkill configuration. 表 3：使用 Gemini-3.5-Flash 对 WikiSkill 进行的消融研究。我们在四种配置下评估了跨基准测试的性能，这些配置改变了推理智能体和技能提议者在技能演进期间是否具有 Wiki 访问权限。当技能提议者没有 Wiki 访问权限时，我们也移除了 Wiki 维护者，消除了跨迭代的持久化知识积累。底行表示我们的默认 WikiSkill 配置。

Persistent wiki knowledge dramatically improves skill evolution持久化 Wiki 知识显著改善了技能演进

As shown in Table 3, when wiki access for the Inference Agent is disabled, providing the Skill Proposer with access to the persistent wiki increases average benchmark performance from 48.7% to 63.7% (+15.0%), with substantial gains on LiveMath (51.3% to 72.6%) and SpreadsheetBench (49.9% to 76.6%). Without persistent knowledge accumulated across iterations, the Skill Proposer struggles to resolve intricate failure modes.如表 3 所示，当禁用推理智能体的 Wiki 访问权限时，为技能提议者提供持久化 Wiki 的访问权限可将平均基准性能从 48.7% 提升至 63.7% (+15.0%)，在 LiveMath (51.3% 至 72.6%) 和 SpreadsheetBench (49.9% 至 76.6%) 上均有显著收益。没有跨迭代积累的持久化知识，技能提议者难以解决复杂的失败模式。

Wiki access for the Inference Agent during evolution degrades final skill quality演进期间推理智能体的 Wiki 访问权限会降低最终技能质量

When the Skill Proposer has access to the persistent wiki, providing the Inference Agent with wiki access during training rollouts reduces average benchmark performance from 63.7% to 60.9%, with a substantial drop on LiveMath from 72.6% to 64.8%. We hypothesize that when the Inference Agent has access to both skills and the wiki during training rollouts, some task-solving knowledge may be obtained directly from the wiki rather than the skills, which can make the resulting trajectories less informative for skill development.当“技能提议者”（Skill Proposer）能够访问持久化维基（persistent wiki）时，在训练展开过程中向“推理智能体”（Inference Agent）提供维基访问权限，会将基准测试的平均性能从 63.7% 降低至 60.9%，其中 LiveMath 的性能下降尤为显著，从 72.6% 降至 64.8%。我们推测，当推理智能体在训练展开过程中同时拥有技能和维基的访问权限时，它可能会直接从维基中获取部分任务解决知识，而非通过技能获取，这会导致生成的轨迹在技能开发方面的参考价值降低。

5.2 Qualitative Analysis: Skill and Wiki Dynamics5.2 定性分析：技能与维基的动态演变

To better understand how WikiSkill evolves knowledge and skills across models and datasets, we analyze the wiki patterns accumulated and skills produced during evolution (Table 4) and when successful skill updates are accepted across iterations (Appendix Table 5).为了更好地理解 WikiSkill 如何在不同模型和数据集上演化知识与技能，我们分析了演化过程中积累的维基模式及产生的技能（表 4），以及在迭代过程中成功采纳的技能更新情况（附录表 5）。

WikiSkill continuously accumulates wiki patterns while producing concise skillsWikiSkill 在生成简洁技能的同时持续积累维基模式

Table 4 summarizes the creation and editing of skills and wiki patterns across models and benchmarks, along with their average lengths. Across models, Qwen models produce longer procedural skills (118.9-128.6 lines), whereas Gemma-4-31B and Gemini-3.5-Flash produce more compact skills (45.1 and 81.2 lines, respectively). Wiki pattern accumulation also varies across models, with 6.3-8.9 patterns created and 7.0–18.4 edits on average. Across benchmarks, SpreadSheet produces the longest skills (142.5 lines) and most wiki patterns (9.8), whereas LiveMath produces the shortest skills (84.6 lines) and fewest wiki patterns (4.4). Overall, these results show that both skill structure and wiki accumulation vary across models and datasets.表 4 总结了跨模型和基准测试的技能与维基模式的创建及编辑情况，以及它们的平均长度。在各模型中，Qwen 模型生成的程序化技能较长（118.9-128.6 行），而 Gemma-4-31B 和 Gemini-3.5-Flash 生成的技能则更为精简（分别为 45.1 行和 81.2 行）。维基模式的积累在不同模型间也存在差异，平均创建 6.3-8.9 个模式，编辑次数为 7.0-18.4 次。在各基准测试中，SpreadSheet 生成的技能最长（142.5 行）且维基模式最多（9.8 个），而 LiveMath 生成的技能最短（84.6 行）且维基模式最少（4.4 个）。总体而言，这些结果表明技能结构和维基积累情况会随模型和数据集的不同而变化。

Skill refinement continues throughout the evolution process技能优化贯穿整个演化过程

Appendix Table 5 groups accepted skill updates into early (Iterations 0–1), middle (Iterations 2–4), and late (Iterations 5–7) stages. Across models, the initial stage accounts for 39%-52% of accepted updates, with substantial fractions continuing into the middle and late stages. A similar pattern holds across benchmarks, where 39%-58% of accepted updates occur during the initial stage. Continued refinement is particularly pronounced on SealQA, where 33% of accepted updates occur in the middle stage and 28% in the late stage. Combined with the ablation in Section 5.1, these results suggest that persistent knowledge accumulation supports continued skill refinement across iterations. The Wiki Layer preserves recurring errors, rejected proposals, and evolution history, which provide the Skill Proposer with accumulated context for subsequent updates. Below, we present a case study that illustrates how this accumulated knowledge informs skill evolution.附录表 5 将被采纳的技能更新分为早期（迭代 0-1）、中期（迭代 2-4）和晚期（迭代 5-7）三个阶段。在各模型中，初始阶段占被采纳更新的 39%-52%，且有相当比例的更新延续至中后期。各基准测试也呈现出相似的模式，即 39%-58% 的被采纳更新发生在初始阶段。持续优化在 SealQA 上尤为明显，其中 33% 的被采纳更新发生在中期，28% 发生在晚期。结合 5.1 节的消融实验，这些结果表明持久的知识积累支持了跨迭代的持续技能优化。维基层保留了重复出现的错误、被拒绝的提议和演化历史，为技能提议者提供了后续更新所需的累积上下文。下面我们通过一个案例研究来说明这些累积知识如何指导技能演化。

5.3 Case Study: Anatomy of Wiki-Guided Skill Evolution5.3 案例研究：维基引导的技能演化剖析

To illustrate how the Wiki and Skill Layers interact during evolution, we trace a concrete example from Qwen-3.6-27B on ALFWorld, as shown in Figure 3 (simplified for presentation).为了说明维基层和技能层在演化过程中的交互方式，我们以 Qwen-3.6-27B 在 ALFWorld 上的表现为例进行追踪，如图 3 所示（为便于展示已简化）。

At Iteration 0, the Wiki Maintainer identifies a basic looping behavior (take-examine-move-loop.md), while the Skill Proposer proposes goal-directed-action, which fails to improve performance on the validation set and is rejected. Crucially, skill-impact.md preserves the proposal diff and rejection outcome, allowing subsequent skill updates to account for this failed attempt.在迭代 0，维基维护者识别出一种基本的循环行为（take-examine-move-loop.md），而技能提议者提出了目标导向动作（goal-directed-action），但该提议未能提升验证集性能而被拒绝。关键在于，skill-impact.md 保存了提议差异和拒绝结果，使得后续的技能更新能够考虑到这次失败的尝试。

Informed by this audit trail, the Skill Proposer creates break-repetition-loop at Iteration 1 with a concrete action rule (Never Return an Item to Its Origin Location), which is accepted. As new loop variants emerge across rollouts (multi-operation-loop.md), the Wiki Maintainer accumulates new evidence in the persistent wiki. Guided by these accumulated wiki patterns and newly stored trajectories (not shown in the figure), the Skill Proposer further refines the skill at Iteration 4 with a new rule (Each Operation Type ONCE Per Item). This example illustrates how persistent knowledge from prior iterations informs subsequent skill refinement.基于此审计追踪，技能提议者在迭代 1 创建了 break-repetition-loop，并制定了具体的行动规则（“从不将物品放回其原始位置”），该规则最终被采纳。随着展开过程中出现新的循环变体（multi-operation-loop.md），维基维护者在持久化维基中积累了新的证据。在这些累积的维基模式和新存储的轨迹（图中未显示）引导下，技能提议者在迭代 4 进一步优化了该技能，增加了新规则（“每种操作类型对每个物品仅限一次”）。此案例展示了来自先前迭代的持久化知识如何为后续的技能优化提供参考。

Category Skills Wiki Patterns Create (Proposed / Accepted) Edit (Proposed / Accepted) Avg. Length Create Edit Avg. Length By Model (All-Dataset Average) Qwen-3.5-4B 3.1 / 1.6 4.9 / 1.3 126.2 8.8 18.4 48.2 Qwen-3.5-9B 4.6 / 1.4 3.4 / 0.7 128.6 7.3 10.9 26.6 Qwen-3.6-27B 4.4 / 1.5 3.6 / 0.8 118.9 6.5 17.9 47.7 Gemma-4-31B 4.8 / 1.3 3.2 / 0.8 45.1 6.3 13.7 23.7 Gemini-3.5-Flash 2.3 / 1.2 5.7 / 1.1 81.2 8.9 7.0 18.1 By Benchmark (All-Model Average) LiveMath 1.9 / 1.1 6.1 / 1.9 84.6 4.4 12.1 31.7 SealQA 4.9 / 0.9 3.1 / 0.4 98.5 9.4 15.9 26.9 SpreadSheet 4.5 / 1.4 3.5 / 1.1 142.5 9.8 11.3 38.5 OfficeQA 4.7 / 1.8 3.3 / 0.3 102.9 8.3 14.9 31.4 ALFWorld 3.9 / 1.6 4.1 / 0.8 93.5 5.8 14.3 40.6

类别技能维基模式创建（提议 / 采纳）编辑（提议 / 采纳）平均长度创建编辑平均长度按模型（全数据集平均） Qwen-3.5-4B 3.1 / 1.6 4.9 / 1.3 126.2 8.8 18.4 48.2 Qwen-3.5-9B 4.6 / 1.4 3.4 / 0.7 128.6 7.3 10.9 26.6 Qwen-3.6-27B 4.4 / 1.5 3.6 / 0.8 118.9 6.5 17.9 47.7 Gemma-4-31B 4.8 / 1.3 3.2 / 0.8 45.1 6.3 13.7 23.7 Gemini-3.5-Flash 2.3 / 1.2 5.7 / 1.1 81.2 8.9 7.0 18.1 按基准测试（全模型平均） LiveMath 1.9 / 1.1 6.1 / 1.9 84.6 4.4 12.1 31.7 SealQA 4.9 / 0.9 3.1 / 0.4 98.5 9.4 15.9 26.9 SpreadSheet 4.5 / 1.4 3.5 / 1.1 142.5 9.8 11.3 38.5 OfficeQA 4.7 / 1.8 3.3 / 0.3 102.9 8.3 14.9 31.4 ALFWorld 3.9 / 1.6 4.1 / 0.8 93.5 5.8 14.3 40.6

Table 4: Statistics of evolved skills and wiki patterns across inference models (top) and benchmarks (bottom). For skills, we report the numbers of proposed/accepted creations and edits, along with average length in markdown lines. For wiki patterns, we report the numbers of creations and edits, along with average length in markdown lines. All wiki pattern creations and edits are retained. 表 4：跨推理模型（上）和基准测试（下）的演化技能与维基模式统计。对于技能，我们报告了提议/采纳的创建和编辑次数，以及以 Markdown 行数表示的平均长度。对于维基模式，我们报告了创建和编辑次数，以及以 Markdown 行数表示的平均长度。所有维基模式的创建和编辑均被保留。

6 Related Work6 相关工作

Experience-Driven Agent Skill Evolution经验驱动的智能体技能演化

Agent skills encode reusable procedural knowledge that allows LLM agents to leverage past experience for future tasks (Zhang et al., 2025; Li et al., 2026; Anthropic, 2026; Xia et al., 2026b; Zhou et al., 2026; Wang et al., 2026; Xu and Yan, 2026). Recent frameworks enable agents to self-improve by discovering and refining procedural knowledge from past execution traces (Yuksekgonul et al., 2025; Agrawal et al., 2026; Ouyang et al., 2026; Xia et al., 2026a; Lu et al., 2026). Methods like EvoSkill (Alzubi et al., 2026), Trace2Skill (Ni et al., 2026), and SkillOpt (Yang et al., 2026) use specialized agent pipelines to analyze task rollouts and update modular skill documents (Zhang et al., 2026b). However, these methods do not maintain what has been learned as a separate, evolving knowledge representation. WikiSkill introduces a persistent Wiki Layer that consolidates experience into structured knowledge across iterations, allowing subsequent skill updates to build systematically on accumulated knowledge.智能体技能编码了可重用的程序化知识，使大语言模型（LLM）智能体能够利用过往经验处理未来任务（Zhang et al., 2025; Li et al., 2026; Anthropic, 2026; Xia et al., 2026b; Zhou et al., 2026; Wang et al., 2026; Xu and Yan, 2026）。近期的框架通过从过往执行轨迹中发现并优化程序化知识，使智能体能够自我提升（Yuksekgonul et al., 2025; Agrawal et al., 2026; Ouyang et al., 2026; Xia et al., 2026a; Lu et al., 2026）。EvoSkill (Alzubi et al., 2026)、Trace2Skill (Ni et al., 2026) 和 SkillOpt (Yang et al., 2026) 等方法利用专门的智能体流水线来分析任务展开过程，并更新模块化的技能文档（Zhang et al., 2026b）。然而，这些方法并未将所学内容作为一种独立的、演化的知识表示进行维护。WikiSkill 引入了持久化维基层，将经验整合为跨迭代的结构化知识，从而使后续的技能更新能够系统地建立在累积的知识之上。

Skill-Augmented Agents and Agent Self-Improvement技能增强型智能体与智能体自我提升

Beyond constructing high-quality skills, skill-augmented agents must effectively select and utilize relevant skills during execution (Chen et al., 2026a; Liu et al., 2026). As the number of reusable skills grows, recent work has explored skill retrieval to select relevant skills from a library for each task (Zheng et al., 2026; Su et al., 2026; Cho et al., 2026; Shi et al., 2026; Ye et al., 2026). WikiSkill instead focuses on skill quality itself, separately from skill retrieval. Another line of work optimizes the broader agent harness, including prompts, context, tools, memory, and workflows (Lou et al., 2026; Lee et al., 2026; Zhang et al., 2026a; Chen et al., 2026b; Lin et al., 2026). These methods improve the agent system by analyzing execution traces and environment feedback to search for better agent configurations. This direction is complementary to WikiSkill, which focuses specifically on evolving reusable procedural skills while holding the broader agent harness fixed.除了构建高质量技能外，技能增强型智能体还必须在执行过程中有效选择和利用相关技能（Chen et al., 2026a; Liu et al., 2026）。随着可重用技能数量的增加，近期研究探索了技能检索技术，以便为每个任务从库中选择相关技能（Zheng et al., 2026; Su et al., 2026; Cho et al., 2026; Shi et al., 2026; Ye et al., 2026）。WikiSkill 则侧重于技能质量本身，与技能检索分离。另一类研究旨在优化更广泛的智能体工具链（harness），包括提示词、上下文、工具、记忆和工作流（Lou et al., 2026; Lee et al., 2026; Zhang et al., 2026a; Chen et al., 2026b; Lin et al., 2026）。这些方法通过分析执行轨迹和环境反馈来搜索更好的智能体配置，从而改进智能体系统。这一方向与 WikiSkill 互补，WikiSkill 专注于演化可重用的程序化技能，同时保持更广泛的智能体工具链不变。

7 Conclusion7 结论

We presented WikiSkill, a framework that co-evolves agent skills with a persistent, compounding knowledge base (wiki). By structuring the agent workspace into three distinct layers, WikiSkill enables skill development to build on increasingly well-supported and integrated knowledge across iterations. An orchestrated loop consolidates experience into the wiki, proposes skill refinements from accumulated knowledge, and gates changes based on validation performance. Empirically, WikiSkill consistently outperforms existing skill-evolution methods across five benchmarks and five inference models and improves over no-skill baselines in most model-dataset pairs. Beyond these overall gains, skill evolution complements model scaling: larger models generally benefit more from evolved skills, while smaller models with skills can outperform substantially larger models without them. At the same time, evolved skills transfer effectively across models and model families and can outperform self-evolved skills. Finally, our ablations confirm that persistent knowledge accumulation is critical for effective skill evolution.我们提出了 WikiSkill，这是一个将智能体技能与持久化、复合型知识库（维基）共同演化的框架。通过将智能体工作空间划分为三个不同的层，WikiSkill 使得技能开发能够建立在跨迭代中日益完善和集成的知识之上。一个精心编排的循环将经验整合到维基中，根据累积的知识提出技能优化建议，并基于验证性能对变更进行门控。实验表明，WikiSkill 在五个基准测试和五个推理模型上始终优于现有的技能演化方法，并在大多数模型-数据集对中优于无技能基准。除了这些整体增益外，技能演化与模型缩放互补：较大的模型通常从演化技能中获益更多，而拥有技能的小型模型可以超越没有技能的显著更大的模型。同时，演化技能可在不同模型和模型家族间有效迁移，并可优于自我演化的技能。最后，我们的消融实验证实，持久的知识积累对于有效的技能演化至关重要。

Limitations局限性

WikiSkill has several limitations that motivate future work. First, to isolate skill quality and avoid confounding effects from skill retrieval, our study follows prior work by directly injecting active skills into the agent prompt. This setup does not evaluate skill retrieval or triggering, which becomes important as the number of available skills grows. Second, our validation gating requires each accepted proposal to improve the validation score, which excludes neutral proposals that preserve immediate performance but could enable gains in subsequent iterations. We adopt this strict criterion following prior skill-evolution frameworks (Yang et al., 2026; Alzubi et al., 2026) to ensure a fair comparison. Exploring more flexible acceptance criteria is an important direction for future work. Third, the Wiki Layer continuously accumulates pattern pages, evolution logs, and proposal diffs across iterations, but WikiSkill currently lacks an automated mechanism to prune the wiki. Such pruning may become necessary as knowledge accumulates over longer evolution runs. Finally, while our benchmark suite includes long-context document reasoning (OfficeQA) and multi-step tool interactions, it does not cover very long-horizon tasks that span hundreds of environment actions or multiple hours. Developing online skill adaptation methods that refine procedural knowledge within a single long execution rollout remains an important direction for future work.WikiSkill 存在一些局限性，这也为未来的工作提供了方向。首先，为了隔离技能质量并避免技能检索带来的混杂效应，我们的研究遵循以往工作，直接将活跃技能注入智能体提示词中。这种设置并未评估技能检索或触发，而随着可用技能数量的增加，这一点变得非常重要。其次，我们的验证门控要求每个被采纳的提议必须提升验证分数，这排除了那些保持即时性能但可能在后续迭代中带来收益的中性提议。我们遵循先前的技能演化框架（Yang et al., 2026; Alzubi et al., 2026）采用了这一严格标准，以确保公平比较。探索更灵活的接受标准是未来工作的一个重要方向。第三，维基层会持续积累模式页面、演化日志和提议差异，但 WikiSkill 目前缺乏自动剪枝维基的机制。随着知识在更长的演化运行中积累，这种剪枝可能变得必要。最后，虽然我们的基准测试套件涵盖了长上下文文档推理（OfficeQA）和多步工具交互，但并未涵盖跨越数百个环境动作或持续数小时的超长视界任务。开发在单次长执行展开中优化程序化知识的在线技能适配方法，仍然是未来工作的重要方向。

AI DisclosureAI 披露

Large language models and coding agents are used to aid with and polish writing and generate some tables and plots.大语言模型和代码智能体被用于辅助和润色写作，并生成部分表格和图表。

References参考文献

Agrawal et al. (2026) L. A. Agrawal, S. Tan, D. Soylu, N. Ziems, R. Khare, K. Opsahl-Ong, A. Singhvi, H. Shandilya, M. J. Ryan, M. Jiang, C. Potts, K. Sen, A. Dimakis, I. Stoica, D. Klein, M. Zaharia, and O. Khattab GEPA: reflective prompt evolution can outperform reinforcement learning. In The Fourteenth International Conference on Learning Representations, External Links: Link Cited by: §1, §4.1, §6. Agrawal et al. (2026) L. A. Agrawal, S. Tan, D. Soylu, N. Ziems, R. Khare, K. Opsahl-Ong, A. Singhvi, H. Shandilya, M. J. Ryan, M. Jiang, C. Potts, K. Sen, A. Dimakis, I. Stoica, D. Klein, M. Zaharia, and O. Khattab GEPA: reflective prompt evolution can outperform reinforcement learning. In The Fourteenth International Conference on Learning Representations, 外部链接: Link 引用自: §1, §4.1, §6.
Alzubi et al. (2026) S. Alzubi, N. Provenzano, J. Bingham, W. Chen, and T. Vu Evoskill: automated skill discovery for multi-agent systems. arXiv preprint arXiv:2603.02766. External Links: Link Cited by: Appendix B, Appendix B, §D.1, §1, §1, §4.1, §6, Limitations. Alzubi et al. (2026) S. Alzubi, N. Provenzano, J. Bingham, W. Chen, and T. Vu Evoskill: automated skill discovery for multi-agent systems. arXiv preprint arXiv:2603.02766. 外部链接: Link 引用自: Appendix B, Appendix B, §D.1, §1, §1, §4.1, §6, Limitations.
Anthropic (2026) Anthropic A complete guide to building skills for claude. Note: https://claude.com/blog/complete-guide-to-building-skills-for-claude Cited by: §1, §6. Anthropic (2026) Anthropic A complete guide to building skills for claude. 注: https://claude.com/blog/complete-guide-to-building-skills-for-claude 引用自: §1, §6.
Chen et al. (2026a) S. Chen, J. Gai, R. Zhou, J. Zhang, T. Zhu, J. Li, K. Wang, Z. Wang, Z. Chen, K. Kaleb, et al. Skillcraft: can llm agents learn to use tools skillfully?. arXiv preprint arXiv:2603.00718. External Links: Link Cited by: §1, §2, §6. Chen et al. (2026a) S. Chen, J. Gai, R. Zhou, J. Zhang, T. Zhu, J. Li, K. Wang, Z. Wang, Z. Chen, K. Kaleb, et al. Skillcraft: can llm agents learn to use tools skillfully?. arXiv preprint arXiv:2603.00718. 外部链接: Link 引用自: §1, §2, §6.
Chen et al. (2026b) T. Chen, S. Lu, K. Zhao, W. Meng, H. Teng, T. Li, C. Li, X. Liu, J. Liang, Z. Zhang, et al. Harnessx: a composable, adaptive, and evolvable agent harness foundry. arXiv preprint arXiv:2606.14249. External Links: Link Cited by: §6. Chen et al. (2026b) T. Chen, S. Lu, K. Zhao, W. Meng, H. Teng, T. Li, C. Li, X. Liu, J. Liang, Z. Zhang, et al. Harnessx: a composable, adaptive, and evolvable agent harness foundry. arXiv preprint arXiv:2606.14249. 外部链接: Link 引用自: §6.
Cho et al. (2026) H. Cho, R. Kang, and Y. Kim SkillRet: a large-scale benchmark for skill retrieval in llm agents. arXiv preprint arXiv:2605.05726. External Links: Link Cited by: §6. Cho et al. (2026) H. Cho, R. Kang, and Y. Kim SkillRet: a large-scale benchmark for skill retrieval in llm agents. arXiv preprint arXiv:2605.05726. 外部链接: Link 引用自: §6.
Gemma Team (2026) Gemma Team Gemma 4 technical report. arXiv preprint arXiv:2607.02770. External Links: Link Cited by: §1, §4.1. Gemma Team (2026) Gemma Team Gemma 4 technical report. arXiv preprint arXiv:2607.02770. 外部链接: Link 引用自: §1, §4.1.
Google DeepMind (2026) Google DeepMind Gemini 3.5 flash. Note: https://deepmind.google/models/model-cards/gemini-3-5-flash/ Cited by: §1, §4.1. Google DeepMind (2026) Google DeepMind Gemini 3.5 flash. 注: https://deepmind.google/models/model-cards/gemini-3-5-flash/ 引用自: §1, §4.1.
He et al. (2026) L. He, Q. Yu, H. Dong, B. Liao, X. Xu, M. Goldblum, J. Bian, and N. Mesgarani Livemathematicianbench: a live benchmark for mathematician-level reasoning with proof sketches. arXiv preprint arXiv:2604.01754. External Links: Link Cited by: Appendix B, §1, §4.1. He et al. (2026) L. He, Q. Yu, H. Dong, B. Liao, X. Xu, M. Goldblum, J. Bian, and N. Mesgarani Livemathematicianbench: a live benchmark for mathematician-level reasoning with proof sketches. arXiv preprint arXiv:2604.01754. 外部链接: Link 引用自: Appendix B, §1, §4.1.
Jackson et al. (2025) D. Jackson, W. Keating, G. Cameron, and M. Hill-Smith AA-omniscience: evaluating cross-domain knowledge reliability in large language models. arXiv preprint arXiv:2511.13029. External Links: Link Cited by: §1. Jackson et al. (2025) D. Jackson, W. Keating, G. Cameron, and M. Hill-Smith AA-omniscience: evaluating cross-domain knowledge reliability in large language models. arXiv preprint arXiv:2511.13029. 外部链接: Link 引用自: §1.
Jiang et al. (2026) Y. Jiang, D. Li, H. Deng, B. Ma, X. Wang, Q. Wang, and G. Yu SoK: agentic skills–beyond tool use in llm agents. arXiv preprint arXiv:2602.20867. External Links: Link Cited by: §1. Jiang et al. (2026) Y. Jiang, D. Li, H. Deng, B. Ma, X. Wang, Q. Wang, and G. Yu SoK: agentic skills–beyond tool use in llm agents. arXiv preprint arXiv:2602.20867. 外部链接: Link 引用自: §1.
Karpathy (2026) A. Karpathy LLM Wiki. Note: GitHub Gist External Links: Link Cited by: §1. Karpathy (2026) A. Karpathy LLM Wiki. 注: GitHub Gist 外部链接: Link 引用自: §1.
Kwon et al. (2023) W. Kwon, Z. Li, S. Zhuang, Y. Sheng, L. Zheng, C. H. Yu, J. Gonzalez, H. Zhang, and I. Stoica Efficient memory management for large language model serving with pagedattention. In Proceedings of the 29th Symposium on Operating Systems Principles, SOSP ’23, New York, NY, USA, pp. 611–626. External Links: ISBN 9798400702297, Link, Document Cited by: §4.1. Kwon et al. (2023) W. Kwon, Z. Li, S. Zhuang, Y. Sheng, L. Zheng, C. H. Yu, J. Gonzalez, H. Zhang, and I. Stoica Efficient memory management for large language model serving with pagedattention. In Proceedings of the 29th Symposium on Operating Systems Principles, SOSP ’23, New York, NY, USA, pp. 611–626. 外部链接: ISBN 9798400702297, Link, Document 引用自: §4.1.
Lee et al. (2026) Y. Lee, R. Nair, Q. Zhang, K. Lee, O. Khattab, and C. Finn Meta-harness: end-to-end optimization of model harnesses. arXiv preprint arXiv:2603.28052. External Links: Link Cited by: §6. Lee et al. (2026) Y. Lee, R. Nair, Q. Zhang, K. Lee, O. Khattab, and C. Finn Meta-harness: end-to-end optimization of model harnesses. arXiv preprint arXiv:2603.28052. 外部链接: Link 引用自: §6.
Li et al. (2026) X. Li, Y. Liu, W. Chen, B. You, Z. Di, Y. He, S. Zheng, K. W. Choe, J. Sun, S. Wang, et al. SkillsBench: benchmarking how well agent skills work across diverse tasks. arXiv preprint arXiv:2602.12670. External Links: Link Cited by: §1, §1, §2, §6. Li et al. (2026) X. Li, Y. Liu, W. Chen, B. You, Z. Di, Y. He, S. Zheng, K. W. Choe, J. Sun, S. Wang, et al. SkillsBench: benchmarking how well agent skills work across diverse tasks. arXiv preprint arXiv:2602.12670. 外部链接: Link 引用自: §1, §1, §2, §6.
Liang et al. (2026) Y. Liang, R. Zhong, H. Xu, C. Jiang, Y. Zhong, R. Fang, J. Gu, S. Deng, Y. Yao, M. Wang, et al. Skillnet: create, evaluate, and connect ai skills. arXiv preprint arXiv:2603.04448. External Links: Link Cited by: §1. Liang et al. (2026) Y. Liang, R. Zhong, H. Xu, C. Jiang, Y. Zhong, R. Fang, J. Gu, S. Deng, Y. Yao, M. Wang, et al. Skillnet: create, evaluate, and connect ai skills. arXiv preprint arXiv:2603.04448. 外部链接: Link 引用自: §1.
Lin et al. (2026) J. Lin, S. Liu, C. Pan, L. Lin, S. Dou, Z. Xi, X. Huang, H. Yan, Z. Han, T. Gui, et al. Agentic harness engineering: observability-driven automatic evolution of coding-agent harnesses. arXiv preprint arXiv:2604.25850. External Links: Link Cited by: §6. Lin et al. (2026) J. Lin, S. Liu, C. Pan, L. Lin, S. Dou, Z. Xi, X. Huang, H. Yan, Z. Han, T. Gui, et al. Agentic harness engineering: observability-driven automatic evolution of coding-agent harnesses. arXiv preprint arXiv:2604.25850. 外部链接: Link 引用自: §6.
Liu et al. (2026) Y. Liu, J. Ji, L. An, T. Jaakkola, Y. Zhang, and S. Chang How well do agentic skills work in the wild: benchmarking llm skill usage in realistic settings. arXiv preprint arXiv:2604.04323. External Links: Link Cited by: §1, §2, §6. Liu et al. (2026) Y. Liu, J. Ji, L. An, T. Jaakkola, Y. Zhang, and S. Chang How well do agentic skills work in the wild: benchmarking llm skill usage in realistic settings. arXiv preprint arXiv:2604.04323. 外部链接: Link 引用自: §1, §2, §6.
Lou et al. (2026) X. Lou, M. Lázaro-Gredilla, A. Dedieu, C. Wendelken, W. Lehrach, and K. P. Murphy Autoharness: improving llm agents by automatically synthesizing a code harness. arXiv preprint arXiv:2603.03329. External Links: Link Cited by: §6. Lou et al. (2026) X. Lou, M. Lázaro-Gredilla, A. Dedieu, C. Wendelken, W. Lehrach, and K. P. Murphy Autoharness: improving llm agents by automatically synthesizing a code harness. arXiv preprint arXiv:2603.03329. 外部链接: Link 引用自: §6.
Lu et al. (2026) Z. Lu, Z. Yao, J. Wu, C. Han, Q. Gu, X. Cai, W. Lu, J. Xiao, Y. Zhuang, and Y. Shen Skill0: in-context agentic reinforcement learning for skill internalization. arXiv preprint arXiv:2604.02268. External Links: Link Cited by: §6. Lu et al. (2026) Z. Lu, Z. Yao, J. Wu, C. Han, Q. Gu, X. Cai, W. Lu, J. Xiao, Y. Zhuang, and Y. Shen Skill0: in-context agentic reinforcement learning for skill internalization. arXiv preprint arXiv:2604.02268. 外部链接: Link 引用自: §6.
Ma et al. (2024) Z. Ma, B. Zhang, J. Zhang, J. Yu, X. Zhang, X. Zhang, S. Luo, X. Wang, and J. Tang SpreadsheetBench: towards challenging real world spreadsheet manipulation. In The Thirty-eight Conference on Neural Information Processing Systems Datasets and Benchmarks Track, External Links: Link Cited by: Appendix B, §1, §4.1. Ma et al. (2024) Z. Ma, B. Zhang, J. Zhang, J. Yu, X. Zhang, X. Zhang, S. Luo, X. Wang, and J. Tang SpreadsheetBench: towards challenging real world spreadsheet manipulation. In The Thirty-eight Conference on Neural Information Processing Systems Datasets and Benchmarks Track, 外部链接: Link 引用自: Appendix B, §1, §4.1.
Merrill et al. (2026) M. A. Merrill, A. G. Shaw, N. Carlini, B. Li, H. Raj, I. Bercovich, L. Shi, J. Y. Shin, T. Walshe, E. K. Buchanan, J. Shen, G. Ye, H. Lin, J. Poulos, M. Wang, M. Nezhurina, D. Lu, O. M. Mastromichalakis, Z. Xu, Z. Chen, Y. Liu, R. Zhang, L. L. Chen, A. Kashyap, J. Uslu, J. Li, J. Wu, M. Yan, S. Bian, V. Sharma, K. Sun, S. Dillmann, A. Anand, A. Lanpouthakoun, B. Koopah, C. Hu, E. K. Guha, G. H. S. Dreiman, J. Zhu, K. Krauth, L. Zhong, N. Muennighoff, R. K. Amanfu, S. Tan, S. Pimpalgaonkar, T. Aggarwal, X. Lin, X. Lan, X. Zhao, Y. Liang, Y. Wang, Z. Wang, C. Zhou, D. Heineman, H. Liu, H. Trivedi, J. Yang, J. Lin, M. Shetty, M. Yang, N. Omi, N. Raoof, S. Li, T. Y. Zhuo, W. Lin, Y. Dai, Y. Wang, W. Chai, S. Zhou, D. Wahdany, Z. She, J. Hu, Z. Dong, Y. Zhu, S. Cui, A. Saiyed, A. Kolbeinsson, C. M. Rytting, R. Marten, Y. Wang, J. Jitsev, A. Dimakis, A. Konwinski, and L. Schmidt Terminal-bench: benchmarking agents on hard, realistic tasks in command line interfaces. In The Fourteenth International Conference on Learning Representations, External Links: Link Cited by: §1. Merrill et al. (2026) M. A. Merrill, A. G. Shaw, N. Carlini, B. Li, H. Raj, I. Bercovich, L. Shi, J. Y. Shin, T. Walshe, E. K. Buchanan, J. Shen, G. Ye, H. Lin, J. Poulos, M. Wang, M. Nezhurina, D. Lu, O. M. Mastromichalakis, Z. Xu, Z. Chen, Y. Liu, R. Zhang, L. L. Chen, A. Kashyap, J. Uslu, J. Li, J. Wu, M. Yan, S. Bian, V. Sharma, K. Sun, S. Dillmann, A. Anand, A. Lanpouthakoun, B. Koopah, C. Hu, E. K. Guha, G. H. S. Dreiman, J. Zhu, K. Krauth, L. Zhong, N. Muennighoff, R. K. Amanfu, S. Tan, S. Pimpalgaonkar, T. Aggarwal, X. Lin, X. Lan, X. Zhao, Y. Liang, Y. Wang, Z. Wang, C. Zhou, D. Heineman, H. Liu, H. Trivedi, J. Yang, J. Lin, M. Shetty, M. Yang, N. Omi, N. Raoof, S. Li, T. Y. Zhuo, W. Lin, Y. Dai, Y. Wang, W. Chai, S. Zhou, D. Wahdany, Z. She, J. Hu, Z. Dong, Y. Zhu, S. Cui, A. Saiyed, A. Kolbeinsson, C. M. Rytting, R. Marten, Y. Wang, J. Jitsev, A. Dimakis, A. Konwinski, and L. Schmidt Terminal-bench: benchmarking agents on hard, realistic tasks in command line interfaces. In The Fourteenth International Conference on Learning Representations, 外部链接: Link 引用自: §1.
Ni et al. (2026) J. Ni, Y. Liu, X. Liu, Y. Sun, M. Zhou, P. Cheng, D. Wang, E. Zhao, X. Jiang, and G. Jiang Trace2skill: distill trajectory-local lessons into transferable agent skills. arXiv preprint arXiv:2603.25158. External Links: Link Cited by: §D.1, §1, §1, §3.2.1, §4.1, §6. Ni et al. (2026) J. Ni, Y. Liu, X. Liu, Y. Sun, M. Zhou, P. Cheng, D. Wang, E. Zhao, X. Jiang, and G. Jiang Trace2skill: distill trajectory-local lessons into transferable agent skills. arXiv preprint arXiv:2603.25158. 外部链接: Link 引用自: §D.1, §1, §1, §3.2.1, §4.1, §6.
Ouyang et al. (2026) S. Ouyang, J. Yan, Y. Chen, R. Han, Z. Wang, B. D. Mishra, R. Meng, C. Li, Y. Jiao, K. Zha, et al. Skillos: learning skill curation for self-evolving agents. arXiv preprint arXiv:2605.06614. External Links: Link Cited by: §1, §6. Ouyang et al. (2026) S. Ouyang, J. Yan, Y. Chen, R. Han, Z. Wang, B. D. Mishra, R. Meng, C. Li, Y. Jiao, K. Zha, et al. Skillos: learning skill curation for self-evolving agents. arXiv preprint arXiv:2605.06614. 外部链接: Link 引用自: §1, §6.
Patwardhan et al. (2026) T. Patwardhan, R. Dias, E. Proehl, G. Kim, M. Wang, O. Watkins, S. P. Fishman, M. Aljubeh, P. Thacker, L. Fauconnet, N. S. Kim, S. Miserendino, G. Chabot, D. Li, P. Chao, M. Sharman, A. Barr, A. Glaese, and J. Tworek GDPval: evaluating AI model performance on real-world economically valuable tasks. In The Fourteenth International Conference on Learning Representations, External Links: Link Cited by: §1. Patwardhan et al. (2026) T. Patwardhan, R. Dias, E. Proehl, G. Kim, M. Wang, O. Watkins, S. P. Fishman, M. Aljubeh, P. Thacker, L. Fauconnet, N. S. Kim, S. Miserendino, G. Chabot, D. Li, P. Chao, M. Sharman, A. Barr, A. Glaese, and J. Tworek GDPval: evaluating AI model performance on real-world economically valuable tasks. In The Fourteenth International Conference on Learning Representations, 外部链接: Link 引用自: §1.
Pham et al. (2026) T. Pham, N. P. Nguyen, P. Zunjare, W. Chen, Y. Tseng, and T. Vu SealQA: raising the bar for reasoning in search-augmented language models. In The Fourteenth International Conference on Learning Representations, External Links: Link Cited by: Appendix B, §1, §4.1. Pham et al. (2026) T. Pham, N. P. Nguyen, P. Zunjare, W. Chen, Y. Tseng, and T. Vu SealQA: raising the bar for reasoning in search-augmented language models. In The Fourteenth International Conference on Learning Representations, 外部链接: Link 引用自: Appendix B, §1, §4.1.
Phan et al. (2026) L. Phan, A. Gatti, Z. Han, N. Li, J. Hu, H. Zhang, C. B. C. Zhang, M. Shaaban, J. Ling, S. Shi, et al. Humanity’s last exam. Nature 649 (8099), pp. 1139–1146. External Links: ISSN 1476-4687, Document, Link Cited by: §1. Phan et al. (2026) L. Phan, A. Gatti, Z. Han, N. Li, J. Hu, H. Zhang, C. B. C. Zhang, M. Shaaban, J. Ling, S. Shi, et al. Humanity’s last exam. Nature 649 (8099), pp. 1139–1146. 外部链接: ISSN 1476-4687, Document, Link 引用自: §1.
Qwen Team (2026a) Qwen Team Qwen3.5: towards native multimodal agents. Note: https://qwen.ai/blog?id=qwen3.5 Cited by: §1, §4.1. Qwen Team (2026a) Qwen Team Qwen3.5: towards native multimodal agents. 注: https://qwen.ai/blog?id=qwen3.5 引用自: §1, §4.1.
Qwen Team (2026b) Qwen Team Qwen3.6-27b. Note: https://qwen.ai/blog?id=qwen3.6-27b Cited by: §1, §4.1. Qwen Team (2026b) Qwen Team Qwen3.6-27b. 注: https://qwen.ai/blog?id=qwen3.6-27b 引用自: §1, §4.1.
Shi et al. (2026) Y. Shi, Y. Chen, Z. Lu, Y. Miao, S. Liu, Q. Gu, X. Cai, X. Wang, and A. Zhang Skill1: unified evolution of skill-augmented agents via reinforcement learning. arXiv preprint arXiv:2605.06130. External Links: Link Cited by: §6. Shi et al. (2026) Y. Shi, Y. Chen, Z. Lu, Y. Miao, S. Liu, Q. Gu, X. Cai, X. Wang, and A. Zhang Skill1: unified evolution of skill-augmented agents via reinforcement learning. arXiv preprint arXiv:2605.06130. 外部链接: Link 引用自: §6.
Shridhar et al. (2021) M. Shridhar, X. Yuan, M. Cote, Y. Bisk, A. Trischler, and M. Hausknecht {ALFW}orld: aligning text and embodied environments for interactive learning. In International Conference on Learning Representations, External Links: Link Cited by: Appendix B, §1, §4.1. Shridhar et al. (2021) M. Shridhar, X. Yuan, M. Cote, Y. Bisk, A. Trischler, and M. Hausknecht {ALFW}orld: aligning text and embodied environments for interactive learning. In International Conference on Learning Representations, 外部链接: Link 引用自: Appendix B, §1, §4.1.
Singhvi et al. (2025) A. Singhvi, K. Opsahl-Ong, J. Collins, I. Zhou, C. Wang, A. Baheti, J. Portes, S. Havens, E. Elsen, M. Bendersky, M. Zaharia, and X. Chen Introducing OfficeQA: a benchmark for end-to-end grounded reasoning. Databricks. Note: Databricks Blog External Links: Link Cited by: Appendix B, §1, §4.1. Singhvi et al. (2025) A. Singhvi, K. Opsahl-Ong, J. Collins, I. Zhou, C. Wang, A. Baheti, J. Portes, S. Havens, E. Elsen, M. Bendersky, M. Zaharia, and X. Chen Introducing OfficeQA: a benchmark for end-to-end grounded reasoning. Databricks. 注: Databricks Blog 外部链接: Link 引用自: Appendix B, §1, §4.1.
Su et al. (2026) W. Su, J. Long, Q. Ai, Q. He, Y. Tang, C. Wang, Y. Tu, Y. Wang, and Y. Liu Skill retrieval augmentation for agentic ai. arXiv preprint arXiv:2604.24594. External Links: Link Cited by: §6. Su et al. (2026) W. Su, J. Long, Q. Ai, Q. He, Y. Tang, C. Wang, Y. Tu, Y. Wang, and Y. Liu Skill retrieval augmentation for agentic ai. arXiv preprint arXiv:2604.24594. 外部链接: Link 引用自: §6.
Wang et al. (2026) H. Wang, Y. Lan, B. Cao, L. Lin, and J. Chen SkillGrad: optimizing agent skills like gradient descent. arXiv preprint arXiv:2605.27760. External Links: Link Cited by: §6. Wang et al. (2026) H. Wang, Y. Lan, B. Cao, L. Lin, and J. Chen SkillGrad: optimizing agent skills like gradient descent. arXiv preprint arXiv:2605.27760. 外部链接: Link 引用自: §6.
Xia et al. (2026a) P. Xia, J. Chen, H. Wang, J. Liu, K. Zeng, Y. Wang, S. Han, Y. Zhou, X. Zhao, H. Chen, et al. Skillrl: evolving agents via recursive skill-augmented reinforcement learning. arXiv preprint arXiv:2602.08234. External Links: Link Cited by: §6. Xia et al. (2026a) P. Xia, J. Chen, H. Wang, J. Liu, K. Zeng, Y. Wang, S. Han, Y. Zhou, X. Zhao, H. Chen, et al. Skillrl: evolving agents via recursive skill-augmented reinforcement learning. arXiv preprint arXiv:2602.08234. 外部链接: Link 引用自: §6.
Xia et al. (2026b) P. Xia, J. Chen, X. Yang, H. Tu, J. Liu, K. Xiong, S. Han, S. Qiu, H. Ji, Y. Zhou, et al. MetaClaw: just talk–an agent that meta-learns and evolves in the wild. arXiv preprint arXiv:2603.17187. External Links: Link Cited by: §1, §6. Xia et al. (2026b) P. Xia, J. Chen, X. Yang, H. Tu, J. Liu, K. Xiong, S. Han, S. Qiu, H. Ji, Y. Zhou, et al. MetaClaw: just talk–an agent that meta-learns and evolves in the wild. arXiv preprint arXiv:2603.17187. 外部链接: Link 引用自: §1, §6.
Xu and Yan (2026) R. Xu and Y. Yan Agent skills for large language models: architecture, acquisition, security, and the path forward. arXiv preprint arXiv:2602.12430. External Links: Link Cited by: §1, §6. Xu and Yan (2026) R. Xu and Y. Yan Agent skills for large language models: architecture, acquisition, security, and the path forward. arXiv preprint arXiv:2602.12430. 外部链接: Link 引用自: §1, §6.
Yang et al. (2026) Y. Yang, Z. Gong, W. Huang, Q. Yang, Z. Zhou, Z. Huang, Y. Li, X. Gao, Q. Dai, B. Liu, et al. SkillOpt: executive strategy for self-evolving agent skills. arXiv preprint arXiv:2605.23904. External Links: Link Cited by: Appendix B, Appendix B, Appendix B, §D.1, §1, §1, §3.2.1, §4.1, §6, Limitations. Yang et al. (2026) Y. Yang, Z. Gong, W. Huang, Q. Yang, Z. Zhou, Z. Huang, Y. Li, X. Gao, Q. Dai, B. Liu, et al. SkillOpt: executive strategy for self-evolving agent skills. arXiv preprint arXiv:2605.23904. 外部链接: Link 引用自: Appendix B, Appendix B, Appendix B, §D.1, §1, §1, §3.2.1, §4.1, §6, Limitations.
Yao et al. (2023) S. Yao, J. Zhao, D. Yu, N. Du, I. Shafran, K. R. Narasimhan, and Y. Cao ReAct: synergizing reasoning and acting in language models. In The Eleventh International Conference on Learning Representations, External Links: Link Cited by: §3.2.3. Yao et al. (2023) S. Yao, J. Zhao, D. Yu, N. Du, I. Shafran, K. R. Narasimhan, and Y. Cao ReAct: synergizing reasoning and acting in language models. In The Eleventh International Conference on Learning Representations, 外部链接: Link 引用自: §3.2.3.
Ye et al. (2026) H. Ye, X. He, V. Arak, H. Dong, and G. Song Meta context engineering via agentic skill evolution. In Forty-third International Conference on Machine Learning, External Links: Link Cited by: §6. Ye et al. (2026) H. Ye, X. He, V. Arak, H. Dong, and G. Song Meta context engineering via agentic skill evolution. In Forty-third International Conference on Machine Learning, 外部链接: Link 引用自: §6.
Yuksekgonul et al. (2025) M. Yuksekgonul, F. Bianchi, J. Boen, S. Liu, P. Lu, Z. Huang, C. Guestrin, and J. Zou Optimizing generative ai by backpropagating language model feedback. Nature 639 (8055), pp. 609–616. External Links: ISSN 1476-4687, Document, Link Cited by: §1, §6. Yuksekgonul et al. (2025) M. Yuksekgonul, F. Bianchi, J. Boen, S. Liu, P. Lu, Z. Huang, C. Guestrin, and J. Zou Optimizing generative ai by backpropagating language model feedback. Nature 639 (8055), pp. 609–616. 外部链接: ISSN 1476-4687, Document, Link 引用自: §1, §6.
Zhang et al. (2025) B. Zhang, K. Lazuka, and M. Murag Equipping agents for the real world with agent skills. Note: https://www.anthropic.com/engineering/equipping-agents-for-the-real-world-with-agent-skills Cited by: §1, §2, §6. Zhang et al. (2025) B. Zhang, K. Lazuka, and M. Murag Equipping agents for the real world with agent skills. 注: https://www.anthropic.com/engineering/equipping-agents-for-the-real-world-with-agent-skills 引用自: §1, §2, §6.
Zhang et al. (2026a) H. Zhang, S. Zhang, K. Li, C. Zhang, Y. Chen, Y. Zhang, L. Bai, and S. Hu Self-harness: harnesses that improve themselves. arXiv preprint arXiv:2606.09498. External Links: Link Cited by: §6. Zhang et al. (2026a) H. Zhang, S. Zhang, K. Li, C. Zhang, Y. Chen, Y. Zhang, L. Bai, and S. Hu Self-harness: harnesses that improve themselves. arXiv preprint arXiv:2606.09498. 外部链接: Link 引用自: §6.
Zhang et al. (2026b) H. Zhang, S. Fan, H. P. Zou, Y. Chen, Z. Wang, J. Zhou, C. Li, W. Huang, Y. Yao, K. Zheng, et al. Coevoskills: self-evolving agent skills via co-evolutionary verification. arXiv preprint arXiv:2604.01687. External Links: Link Cited by: §6. Zhang et al. (2026b) H. Zhang, S. Fan, H. P. Zou, Y. Chen, Z. Wang, J. Zhou, C. Li, W. Huang, Y. Yao, K. Zheng, et al. Coevoskills: self-evolving agent skills via co-evolutionary verification. arXiv preprint arXiv:2604.01687. 外部链接: Link 引用自: §6.
Zheng et al. (2026) Y. Zheng, Z. Zhang, C. Ma, Y. Yu, J. Zhu, Y. Wu, T. Xu, B. Dong, H. Zhu, R. Huang, et al. Skillrouter: skill routing for llm agents at scale. arXiv preprint arXiv:2603.22455. External Links: Link Cited by: §6. Zheng et al. (2026) Y. Zheng, Z. Zhang, C. Ma, Y. Yu, J. Zhu, Y. Wu, T. Xu, B. Dong, H. Zhu, R. Huang, et al. Skillrouter: skill routing for llm agents at scale. arXiv preprint arXiv:2603.22455. 外部链接: Link 引用自: §6.
Zhou et al. (2026) H. Zhou, S. Guo, A. Liu, Z. Yu, Z. Gong, B. Zhao, Z. Chen, M. Zhang, Y. Chen, J. Li, et al. Memento-skills: let agents design agents. arXiv preprint arXiv:2603.18743. External Links: Link Cited by: §6. Zhou et al. (2026) H. Zhou, S. Guo, A. Liu, Z. Yu, Z. Gong, B. Zhao, Z. Chen, M. Zhang, Y. Chen, J. Li, et al. Memento-skills: let agents design agents. arXiv preprint arXiv:2603.18743. 外部链接: Link 引用自: §6.

Appendix A Method Details附录 A 方法细节

A.1 AlgorithmA.1 算法

The full skill-evolution algorithm for WikiSkill is described in Algorithm 1.WikiSkill 的完整技能演化算法描述见算法 1。

1: Training tasks

\mathcal{D}_{\text{train}}

, validation tasks

\mathcal{D}_{\text{val}}

, performance metric

\mathcal{R}

, iterations

K

2: Initialize skill set

S_{0}\leftarrow\emptyset

, wiki

W_{0}\leftarrow\emptyset

3: Baseline Validation:基准验证：

\mathcal{T}_{\text{val},0}\leftarrow\{\tau_{i}\sim\pi(x_{i};S_{0})\}_{x_{i}\in\mathcal{D}_{\text{val}}}

\mathcal{R}_{\text{best}}\leftarrow\mathcal{R}(\mathcal{T}_{\text{val},0})

4: for对于

k=1,\dots,K

5: if

\mathcal{R}_{\text{best}}=1.0

then那么

6: break中断

7: end结束 if

8: Inference:推理： Roll out

\mathcal{T}_{\text{train},k}\leftarrow\{\tau_{i}\sim\pi(x_{i};S_{k-1})\}_{x_{i}\in\mathcal{D}_{\text{train}}}

9: Sample subset

\mathcal{T}_{\text{sample},k}\subset\mathcal{T}_{\text{train},k}

10:10: Wiki Maintenance:维基维护：

W^{\prime}_{k}\leftarrow\mathcal{M}_{\text{WM}}(W_{k-1},\mathcal{T}_{\text{sample},k})

11:11: Skill Proposal:技能提议：

P_{k}\leftarrow\mathcal{M}_{\text{P}}(W^{\prime}_{k},S_{k-1},\mathcal{T}_{\text{train},k})

12:12: Apply:应用：

S^{\prime}_{k}\leftarrow\text{Apply}(S_{k-1},P_{k})

13:13: Validate:验证：

\mathcal{T}_{\text{val},k}\leftarrow\{\tau_{i}\sim\pi(x_{i};S^{\prime}_{k})\}_{x_{i}\in\mathcal{D}_{\text{val}}}

14:14: if

\mathcal{R}(\mathcal{T}_{\text{val},k})>\mathcal{R}_{\text{best}}

then

15:15:

S_{k}\leftarrow S^{\prime}_{k}

\mathcal{R}_{\text{best}}\leftarrow\mathcal{R}(\mathcal{T}_{\text{val},k})

a_{k}\leftarrow\text{Accepted}

16:16: else否则

17:17:

S_{k}\leftarrow S_{k-1}

a_{k}\leftarrow\text{Rejected}

\triangleright

Roll back skills only; wiki retained

18:18: end if

19:19: Update Wiki Log:更新维基日志：

W_{k}\leftarrow\text{Update}(W^{\prime}_{k},P_{k},\mathcal{R}(\mathcal{T}_{\text{val},k}),a_{k})

20:20: end for

21:21: return返回

S_{K}

W_{K}

Algorithm 1 WikiSkill evolution loop. At each iteration

k

, the inference agent rolls out on training tasks using active skills

S_{k-1}

, the maintainer consolidates sampled traces into the intermediate wiki

W^{\prime}_{k}

, the proposer generates candidate skill modifications

P_{k}

, validation gating determines whether to accept

S^{\prime}_{k}

or roll back to

S_{k-1}

, and the system appends the proposal outcome and skill diff to produce the final wiki state

W_{k}

.算法 1 WikiSkill 演化循环。在每次迭代 k 中，推理智能体使用活跃技能 Sk-1 在训练任务上进行展开，维护者将采样的轨迹整合到中间维基 W'k 中，提议者生成候选技能修改 Pk，验证门控决定是否采纳 S'k 或回滚至 Sk-1，系统随后附加提议结果和技能差异以生成最终的维基状态 Wk。

A.2 Distribution of Accepted Skill UpdatesA.2 被采纳技能更新的分布

We show when the updated skill proposals are accepted in Table 5.我们在表 5 中展示了更新后的技能提议被采纳的时间点。

Appendix B Dataset Details and Splits附录 B 数据集细节与划分

We describe the five benchmarks used in our evaluation below.我们在下面描述了评估中使用的五个基准测试。

LiveMathematicianBench (LiveMath) (He et al., 2026) consists of multiple-choice mathematics competition problems from recent months. It tests the model’s capacity for complex mathematical reasoning, quantifiers, and extremal conditions. SealQA (Pham et al., 2026) is a factual question-answering benchmark composed of scholarly questions across various topics. It evaluates the agent’s ability to formulate effective search queries and extract answers from web search results using a search tool. SpreadsheetBench (SpreadSheet) (Ma et al., 2024) tests the agent’s ability to write correct code under library constraints (such as formula evaluation limitations) and execute complex table transformations. OfficeQA (Singhvi et al., 2025) evaluates long-context question-answering over a large repository of historical Treasury bulletins. Tasks require synthesizing evidence across long contexts and multi-page financial tables. Following the setup in Yang et al. (2026), the agent is provided with pre-parsed oracle reference pages as initial document evidence in the prompt, while retaining access to local text-processing tools (glob, grep, read) to search, cross-reference, and inspect full Treasury bulletin files on disk. ALFWorld (Shridhar et al., 2021) is an interactive text-based embodied environment where an agent solves multi-step household tasks (e.g., picking and placing objects, heating or cooling items) by outputting text actions to a simulator. Unlike static QA benchmarks, ALFWorld tests sequential decision-making, spatial reasoning, and error recovery from simulator feedback.LiveMathematicianBench (LiveMath) (He et al., 2026) 由近几个月的选择题数学竞赛问题组成。它测试模型处理复杂数学推理、量词和极值条件的能力。SealQA (Pham et al., 2026) 是一个由各领域学术问题组成的事实问答基准。它评估智能体制定有效搜索查询并利用搜索工具从网络搜索结果中提取答案的能力。SpreadsheetBench (SpreadSheet) (Ma et al., 2024) 测试智能体在库限制（如公式评估限制）下编写正确代码并执行复杂表格转换的能力。OfficeQA (Singhvi et al., 2025) 评估对大型历史财政公告库的长上下文问答能力。任务需要跨长上下文和多页财务表格综合证据。遵循 Yang et al. (2026) 的设置，智能体在提示词中获得预解析的预言机参考页面作为初始文档证据，同时保留对本地文本处理工具（glob, grep, read）的访问权限，以搜索、交叉引用和检查磁盘上的完整财政公告文件。ALFWorld (Shridhar et al., 2021) 是一个交互式文本驱动的具身环境，智能体通过向模拟器输出文本动作来解决多步家庭任务（例如，拾取和放置物体、加热或冷却物品）。与静态 QA 基准不同，ALFWorld 测试序列决策、空间推理以及从模拟器反馈中进行错误恢复的能力。

Category	Early (Iter 0–1)	Mid (Iter 2–4)	Late (Iter 5–7)
By Model
Qwen-3.5-4B	39%	39%	21%
Qwen-3.5-9B	52%	30%	19%
Qwen-3.6-27B	43%	40%	17%
Gemma-4-31B	52%	37%	11%
Gemini-3.5-Flash	50%	46%	4%
By Benchmark
LiveMath	44%	42%	14%
SealQA	39%	33%	28%
SpreadSheet	41%	48%	11%
OfficeQA	58%	26%	16%
ALFWorld	55%	34%	10%

Table 5: Distribution of accepted skill updates across evolution iterations, grouped by model (top) and benchmark (bottom). Percentages indicate the proportion of accepted updates that occur during each stage of evolution. 表 5：跨演化迭代的被采纳技能更新分布，按模型（上）和基准测试（下）分组。百分比表示在演化每个阶段中被采纳更新的比例。

Benchmark	Interaction	Train	Val	Test	Environment Tools
LiveMath	Single-Step	35	18	124	None (Direct Reasoning)
SealQA	Multi-Step	16	10	85	web_search, read_file
SpreadSheet	Multi-Step	80	40	280	bash
OfficeQA	Multi-Step	50	24	172	glob, grep, read
ALFWorld	Multi-Step	39	18	134	Admissible Actions

Table 6: Benchmark statistics, data splits, interaction modes, and available environment tools. 表 6：基准测试统计、数据划分、交互模式和可用环境工具。

Table 6 summarizes the sample counts across training, validation, and test splits, interaction modes, and available tools for each benchmark evaluated in our experiments. All task splits and available toolsets are strictly matched with prior work (Yang et al., 2026; Alzubi et al., 2026). For tool setup, LiveMath operates as a single-step reasoning benchmark without external tools, where the model generates final answers directly; SealQA equips the agent with web search (using Google Search API) and file reading for multi-step factual retrieval (we use the July, 2026 version of SealQA for all experiments); SpreadSheet provides a bash shell tool for Python code execution and table manipulation; OfficeQA provides local text-search utilities for multi-step Treasury bulletin navigation; and ALFWorld provides an interactive simulator action space for multi-step embodied decision making.表 6 总结了我们实验中评估的每个基准测试在训练、验证和测试划分上的样本数量、交互模式和可用工具。所有任务划分和可用工具集均严格匹配以往工作（Yang et al., 2026; Alzubi et al., 2026）。在工具设置方面，LiveMath 作为无需外部工具的单步推理基准运行，模型直接生成最终答案；SealQA 为智能体配备了网页搜索（使用 Google Search API）和文件读取功能，用于多步事实检索（所有实验均使用 2026 年 7 月版的 SealQA）；SpreadSheet 提供 bash shell 工具用于 Python 代码执行和表格操作；OfficeQA 提供本地文本搜索工具用于多步财政公告导航；ALFWorld 提供用于多步具身决策的交互式模拟器动作空间。

Evaluation robustness with small validation sets小规模验证集下的评估稳健性

Following established setups in prior work (Yang et al., 2026; Alzubi et al., 2026), benchmark validation splits are relatively small, which can introduce evaluation noise into gating decision. To account for this variability, all reported scores represent the average test performance across three independent runs of the entire evolutionary pipeline, with paired bootstrap significance testing (detailed in Appendix C).遵循以往工作（Yang et al., 2026; Alzubi et al., 2026）的既定设置，基准测试的验证集相对较小，这可能会给门控决策引入评估噪声。为了考虑这种变异性，所有报告的分数均代表整个演化流水线三次独立运行的平均测试性能，并进行了成对引导重采样显著性检验（详见附录 C）。

Appendix C Implementation Details附录 C 实现细节

To provide diagnostic feedback for the Wiki maintainer, we apply a stratified sampling strategy ( $\mathcal{T}_{\text{sample},k}\subset\mathcal{T}_{\text{train},k}$ ) at each iteration $k$ . Specifically, the system samples up to 8 traces per iteration, stratified into a maximum of 5 failing traces (to perform root-cause analysis of errors) and up to 3 passing traces (to identify effective strategies and prevent regressions in working behaviors). Each individual execution log is capped at 15,000 characters prior to injection into the prompt.为了向维基维护者提供诊断反馈，我们在每次迭代 k 应用了分层采样策略（𝒯sample,k⊂𝒯train,k）。具体而言，系统在每次迭代中最多采样 8 条轨迹，分层为最多 5 条失败轨迹（用于进行错误的根本原因分析）和最多 3 条成功轨迹（用于识别有效策略并防止工作行为回归）。在注入提示词之前，每条单独的执行日志上限为 15,000 个字符。

Statistical significance testing统计显著性检验

We perform paired bootstrap significance tests with $1,000$ iterations for each benchmark. In each bootstrap iteration, task instances are sampled with replacement from the test split $\mathcal{D}_{\text{test}}$ to construct a bootstrap evaluation set of size $|\mathcal{D}_{\text{test}}|$ , from which candidate accuracy scores and pairwise performance margins are computed. To evaluate overall cross-benchmark performance, we conduct stratified macro-average bootstrap resampling: in each iteration, task instances are resampled independently with replacement within each benchmark, and we calculate macro-average accuracy by assigning equal weight to all benchmarks.我们对每个基准测试进行 1,000 次迭代的成对引导重采样显著性检验。在每次引导迭代中，从测试划分 𝒟test 中有放回地采样任务实例，以构建大小为 |𝒟test| 的引导评估集，并据此计算候选准确率和成对性能差值。为了评估整体跨基准测试性能，我们进行分层宏平均引导重采样：在每次迭代中，在每个基准测试内独立进行任务实例的有放回重采样，并通过为所有基准测试分配相等权重来计算宏平均准确率。

We determine top-performing methods in our evaluation tables as follows. Methods are initially ranked by their observed performance (or macro-average performance across benchmarks). A top-ranked method $M^{*}$ is the sole top performer if and only if it achieves a statistically significant gain over all competing methods at $p<0.05$ . If $M^{*}$ is not statistically distinguishable ( $p\geq 0.05$ ) against one or more lower-ranked methods, no single method is declared the sole top performer. Instead, all methods whose performance is not statistically distinguishable from $M^{*}$ ( $p\geq 0.05$ ) are grouped into a top-tier statistical tie (and bolded accordingly).我们按如下方式确定评估表中的表现最优方法。方法首先按观察到的性能（或跨基准测试的宏平均性能）进行排名。当且仅当排名第一的方法 M* 在 p<0.05 时对所有竞争方法取得统计学显著增益时，它才是唯一的表现最优者。如果 M* 与一个或多个排名较低的方法在统计上无法区分（p≥0.05），则不宣布单一方法为表现最优者。相反，所有性能与 M* 在统计上无法区分（p≥0.05）的方法都被归入顶层统计平局组（并相应加粗）。

Appendix D Baseline Details and Optimizer API Call Analysis附录 D 基准方法细节与优化器 API 调用分析

D.1 Baseline MethodsD.1 基准方法

Trace2Skill (Ni et al., 2026)Trace2Skill (Ni et al., 2026)

Trace2Skill employs a three-stage pipeline centered on parallel trace analysis and hierarchical merging. It evaluates the current skill on training tasks and dispatches parallel success analysts and error analysts to extract effective strategies from passing tasks and diagnose root causes of failures. The resulting structured patches are recursively consolidated via a hierarchical merge operator into a single patch set. The consolidated patches are applied to the skill document and accepted based on validation performance.Trace2Skill 采用了一个以并行轨迹分析和分层合并为核心的三阶段流水线。它在训练任务上评估当前技能，并派遣并行成功分析师和错误分析师从成功任务中提取有效策略，并诊断失败的根本原因。生成的结构化补丁通过分层合并算子递归整合为单个补丁集。整合后的补丁被应用于技能文档，并根据验证性能决定是否采纳。

EvoSkill (Alzubi et al., 2026)EvoSkill (Alzubi et al., 2026)

EvoSkill frames skill evolution as a search over a frontier of candidate programs. In each iteration, it samples training tasks via a round-robin category schedule and executes rollouts. EvoSkill feeds only failure traces to the proposer alongside a flat feedback history of past proposal outcomes. The proposer generates candidate skill modifications that are materialized into SKILL.md files, scored on a validation split, and added to a bounded frontier of top-performing programs.EvoSkill 将技能演化构建为候选程序前沿的搜索过程。在每次迭代中，它通过轮询类别调度采样训练任务并执行展开。EvoSkill 仅将失败轨迹连同过往提议结果的平坦反馈历史提供给提议者。提议者生成候选技能修改，这些修改被具体化为 SKILL.md 文件，在验证集上进行评分，并添加到表现最优程序的有界前沿中。

SkillOpt (Yang et al., 2026)SkillOpt (Yang et al., 2026)

SkillOpt implements a six-stage ReflACT pipeline (Rollout, Reflect, Aggregate, Select, Update, Evaluate) for iterative skill optimization. In each epoch, the system rolls out the agent on training tasks and reflects on full execution traces, including both successes and failures, to generate candidate patches. These patches are hierarchically aggregated and selected to update a single monolithic skill document, which is accepted or rejected based on validation performance.SkillOpt 实现了一个用于迭代技能优化的六阶段 ReflACT 流水线（展开、反思、聚合、选择、更新、评估）。在每个 epoch 中，系统在训练任务上展开智能体，并反思包括成功和失败在内的完整执行轨迹，以生成候选补丁。这些补丁经过分层聚合和选择，用于更新单个单体技能文档，并根据验证性能决定采纳或拒绝。

D.2 Optimizer API Call ComplexityD.2 优化器 API 调用复杂度

In this section, we analyze optimizer API call complexity, denoted by $\mathcal{C}$ , across self-improving agent frameworks. We define one evolution iteration as rolling out the agent on the full training split $\mathcal{D}_{\text{train}}$ once ( $N_{\text{train}}=|\mathcal{D}_{\text{train}}|$ training task instances), processed in minibatches of size $B$ ( $B\leq N_{\text{train}}$ ). Table 7 summarizes the per-iteration optimizer API call complexity for WikiSkill and prior methods.在本节中，我们分析了各自我提升智能体框架中优化器 API 的调用复杂度，记作 𝒞。我们将一次演进迭代定义为：智能体在完整训练集 𝒟train（包含 Ntrain=|𝒟train| 个训练任务实例）上运行一次，并以大小为 B (B≤Ntrain) 的小批量进行处理。表 7 总结了 WikiSkill 及先前方法在每次迭代中的优化器 API 调用复杂度。

Framework	Per-Iteration Formula	Complexity
Trace2Skill	$N_{\text{train}}+\left(1+\frac{1}{c-1}\right)\frac{N_{\text{train}}}{B}+1$	$\mathcal{O}\left(N_{\text{train}}+\frac{N_{\text{train}}}{B}\right)$
EvoSkill	$\frac{2N_{\text{train}}}{B}$	$\mathcal{O}\left(\frac{N_{\text{train}}}{B}\right)$
SkillOpt	$\frac{K_{\text{opt}}\cdot N_{\text{train}}}{B}$	$\mathcal{O}\left(\frac{N_{\text{train}}}{B}\right)$
WikiSkill	$(1+T_{\text{ReAct}})\frac{N_{\text{train}}}{B}$	$\mathcal{O}\left(\frac{N_{\text{train}}}{B}\right)$

Table 7: Comparison of optimizer API call complexity per evolution iteration across self-improving agent frameworks.

N_{\text{train}}=|\mathcal{D}_{\text{train}}|

denotes the number of training tasks,

B

denotes the batch size,

T_{\text{ReAct}}

denotes the number of interactive ReAct reasoning turns used by the Skill Proposer agent in WikiSkill,

K_{\text{opt}}

denotes the number of reflection and merging calls per step in SkillOpt, and

c

denotes the reduction-tree branching factor in Trace2Skill. WikiSkill uses

B=N_{\text{train}}

across all datasets. In full-batch mode, WikiSkill’s optimizer call count is independent of training set size

N_{\text{train}}

, requiring

1+T_{\text{ReAct}}

optimizer calls per iteration.表 7：各自我提升智能体框架在每次演进迭代中优化器 API 的调用复杂度比较。Ntrain=|𝒟train| 表示训练任务数量，B 表示批大小，TReAct 表示 WikiSkill 中技能提议智能体（Skill Proposer agent）所使用的交互式 ReAct 推理轮数，Kopt 表示 SkillOpt 中每一步的反射与合并调用次数，c 表示 Trace2Skill 中归约树的分支因子。WikiSkill 在所有数据集上均使用 B=Ntrain。在全批次模式下，WikiSkill 的优化器调用次数与训练集大小 Ntrain 无关，每次迭代需要 1+TReAct 次优化器调用。

WikiSkillWikiSkill

For each batch of size $B$ , the Wiki Maintainer requires one LLM call to analyze sampled traces $\mathcal{T}_{\text{sample},k}$ and consolidate pattern pages into the intermediate wiki $W^{\prime}_{k}$ . The Skill Proposer then runs as an autonomous multi-turn ReAct agent, executing interactive tool calls over $T_{\text{ReAct}}$ reasoning turns (roughly $10\leq T_{\text{ReAct}}\leq 20$ across our experiment runs), where each ReAct turn requires $1$ LLM call. When processing training data in batches of size $B$ , completing one full iteration over $N_{\text{train}}$ tasks requires $\frac{N_{\text{train}}}{B}$ steps:对于每个大小为 B 的批次，Wiki 维护者（Wiki Maintainer）需要调用一次大语言模型（LLM）来分析采样轨迹 𝒯sample,k，并将模式页面整合到中间 wiki Wk′ 中。随后，技能提议者（Skill Proposer）作为自主的多轮 ReAct 智能体运行，在 TReAct 轮推理中执行交互式工具调用（在我们的实验运行中，大约 10≤TReAct≤20），其中每一轮 ReAct 需要调用 1 次 LLM。当以大小为 B 的批次处理训练数据时，完成对 Ntrain 个任务的完整迭代需要 Ntrain/B 步：

\mathcal{C}_{\text{WikiSkill}}=(1+T_{\text{ReAct}})\frac{N_{\text{train}}}{B}

(5)

In our experiments, we set the batch size to the full training size ( $B=N_{\text{train}}$ ) across all datasets. In this full-batch setting ( $\frac{N_{\text{train}}}{B}=1$ ), $\mathcal{C}_{\text{WikiSkill}}=1+T_{\text{ReAct}}$ . Because $T_{\text{ReAct}}$ does not depend on $N_{\text{train}}$ , WikiSkill’s optimizer API call complexity is $\mathcal{O}(1)$ with respect to training set size. Specifically, each iteration requires $1+T_{\text{ReAct}}$ optimizer LLM calls, regardless of the number of training instances. While this constant call complexity may incur higher inference cost on some datasets, the additional computation is accompanied by consistent performance gains over prior skill-evolution methods across our evaluation.在我们的实验中，我们将所有数据集的批大小设为完整训练集大小 (B=Ntrain)。在这种全批次设置下 (Ntrain/B=1)，𝒞WikiSkill=1+TReAct。由于 TReAct 与 Ntrain 无关，因此 WikiSkill 的优化器 API 调用复杂度相对于训练集大小为 𝒪(1)。具体而言，无论训练实例数量多少，每次迭代都需要 1+TReAct 次优化器 LLM 调用。虽然这种恒定的调用复杂度在某些数据集上可能会带来更高的推理成本，但额外的计算换取了在我们的评估中相对于先前技能演进方法的一致性能提升。

EvoSkillEvoSkill

EvoSkill partitions training tasks into minibatches of size $B$ . For each minibatch, EvoSkill uses one Proposer LLM call for error diagnosis and one Generator LLM call for skill updates. Processing the full training split of $N_{\text{train}}$ tasks requires $\frac{N_{\text{train}}}{B}$ minibatch steps, resulting in:EvoSkill 将训练任务划分为大小为 B 的小批量。对于每个小批量，EvoSkill 使用一次提议者 LLM 调用进行错误诊断，并使用一次生成器 LLM 调用进行技能更新。处理包含 Ntrain 个任务的完整训练集需要 Ntrain/B 个小批量步骤，结果为：

\mathcal{C}_{\text{EvoSkill}}=\frac{2N_{\text{train}}}{B}

(6)

Thus, EvoSkill’s optimizer API call complexity scales linearly with training set size $N_{\text{train}}$ ( $\mathcal{O}(N_{\text{train}}/B)$ ).因此，EvoSkill 的优化器 API 调用复杂度随训练集大小 Ntrain 线性扩展 (𝒪(Ntrain/B))。

SkillOptSkillOpt

SkillOpt evaluates minibatches of size $B$ , completing each iteration in $\frac{N_{\text{train}}}{B}$ optimization steps. During each step, SkillOpt executes its ReflACT pipeline (parallel analyst reflections, hierarchical patch synthesis, and candidate selection), requiring $K_{\text{opt}}\approx 6\text{--}8$ optimizer LLM calls per step:SkillOpt 评估大小为 B 的小批量，并在 Ntrain/B 个优化步骤中完成每次迭代。在每一步中，SkillOpt 执行其 ReflACT 流水线（并行分析师反射、分层补丁合成和候选选择），每步需要 Kopt≈6–8 次优化器 LLM 调用：

\mathcal{C}_{\text{SkillOpt}}=\frac{K_{\text{opt}}\cdot N_{\text{train}}}{B}

(7)

SkillOpt similarly scales linearly with training set size $N_{\text{train}}$ , as $\mathcal{O}(N_{\text{train}}/B)$ .SkillOpt 同样随训练集大小 Ntrain 线性扩展，即 𝒪(Ntrain/B)。

Trace2SkillTrace2Skill

Trace2Skill processes training trajectories through a three-stage pipeline per iteration:Trace2Skill 在每次迭代中通过三个阶段的流水线处理训练轨迹：

1.

Trace analysis stage: Every individual execution trajectory is analyzed independently with one LLM call, incurring $N_{\text{train}}$ total calls.轨迹分析阶段：每一条单独的执行轨迹都通过一次 LLM 调用独立分析，总共产生 Ntrain 次调用。
2.

Patch map stage: Analysis records are chunked into batches of size $B$ , requiring $\frac{N_{\text{train}}}{B}$ calls to generate local skill patches.补丁映射阶段：分析记录被分块为大小为 B 的批次，需要 Ntrain/B 次调用来生成局部技能补丁。
3.

Hierarchical reduce & apply stage: The $\frac{N_{\text{train}}}{B}$ local patches are recursively merged via a $c$ -ary reduction tree (where $c$ is the branching factor). Summing across levels yields $\approx\frac{1}{c-1}\frac{N_{\text{train}}}{B}$ merge calls, plus one final call to format the skill document.分层归约与应用阶段：Ntrain/B 个局部补丁通过 c 元归约树（c 为分支因子）递归合并。各层求和得出 ≈1/(c-1) * (Ntrain/B) 次合并调用，外加一次最终调用以格式化技能文档。

Combining all stages:结合所有阶段：

\mathcal{C}_{\text{Trace2Skill}}\approx N_{\text{train}}+\left(1+\frac{1}{c-1}\right)\frac{N_{\text{train}}}{B}+1

(8)

Because Trace2Skill performs individual LLM analysis on every training trajectory ( $N_{\text{train}}$ calls), its complexity is lower-bounded by $\mathcal{O}(N_{\text{train}})$ , scaling linearly with training set size.由于 Trace2Skill 对每条训练轨迹都执行单独的 LLM 分析（Ntrain 次调用），其复杂度下限为 𝒪(Ntrain)，随训练集大小线性扩展。

Full-batch training vs. minibatch optimization全批次训练与小批量优化

Across all datasets, we set the batch size $B$ to the full training set size ( $B=N_{\text{train}}$ ) for WikiSkill, processing the entire training set at once per iteration. The Skill Proposer dynamically searches, selects, and reads specific execution traces on demand to diagnose root causes before proposing skill updates. In contrast, EvoSkill and SkillOpt achieve their best performance under minibatch settings ( $B<N_{\text{train}}$ ), which causes their optimizer API call complexity to scale linearly with training set size. Finally, Trace2Skill remains strictly $\mathcal{O}(N_{\text{train}})$ regardless of minibatch size because it requires an independent LLM call for every training trajectory.在所有数据集上，我们将 WikiSkill 的批大小 B 设置为完整训练集大小 (B=Ntrain)，在每次迭代中一次性处理整个训练集。技能提议者会根据需要动态搜索、选择和读取特定的执行轨迹，以在提出技能更新前诊断根本原因。相比之下，EvoSkill 和 SkillOpt 在小批量设置 (B<Ntrain) 下表现最佳，这导致其优化器 API 调用复杂度随训练集大小线性扩展。最后，Trace2Skill 始终保持 𝒪(Ntrain)，无论小批量大小如何，因为它需要对每一条训练轨迹进行独立的 LLM 调用。

Appendix E System and Agent Prompts附录 E 系统与智能体提示词

We provide the exact system prompts used for (1) the Inference Agent for each task across all methods, (2) the Wiki Maintainer, and (3) the Skill Proposer in WikiSkill.我们提供了以下精确的系统提示词：(1) 所有方法中针对各项任务的推理智能体，(2) Wiki 维护者，以及 (3) WikiSkill 中的技能提议者。

E.1 Task Inference Agent System PromptsE.1 任务推理智能体系统提示词

LiveMathematicianBench Inference Agent System Prompt LiveMathematicianBench 推理智能体系统提示词

⬇

You你 are是 an expert专家 mathematical数学 reasoning推理 agent智能体 solving用于解决 multiple多项-choice选择 questions问题.

{skill_sectionskill_section}

## Task任务 Format格式

You will将 receive接收 one一道 mathematics数学 multiple-choice question问题 and及其 its选项 answer答案 choices选项. Reason推理 carefully仔细 about考虑 quantifiers量词, hypotheses假设, extremal极值 wording措辞, and exact精确 equality相等 conditions条件.

## Answer答案 Format

Think思考 step步骤 by step, then provide提供 your你的 final最终 answer inside在 <answer>...</>...</answer> tags标签. Inside在 the输出 tags, output中 only仅 the single单个 choice label标签, such例如 as A or C.

Example示例:

SealQA Inference Agent System Prompt SealQA 推理智能体系统提示词

⬇

You are a knowledgeable知识渊博的 question-answering回答 assistant助手 with具有 access访问 to web_searchweb_search and read_fileread_file tools工具.

{skill_section}

## Task

You will receive a factual事实 question. To answer it:

1. You can可以 check检查 the available可用 skills技能. They它们 contain包含 guidance指导 that以 can improve改进 your search搜索 queries查询 and answer accuracy准确性.

2. You can use使用 web_search to find查找 relevant相关 information信息. You can call调用 it multiple times次数 with different不同 queries.

3. You can do web_search anytime随时 during在 the process过程中 depending取决于 on your needs需求.

4. After在 gathering收集 enough足够 information, provide your final answer.

## Answer Format

You MUST wrap总结 your final answer in <answer> tags:

...... your final answer (exact value值 only, no explanation解释) ...

</answer>

SpreadsheetBench Inference Agent System Prompt SpreadsheetBench 推理智能体系统提示词

⬇

You are a spreadsheet电子表格 expert who操作者 can manipulate处理 spreadsheets电子表格 through通过 PythonPython code代码.

{skill_section}

You need需要 to solve解决 the given给定 spreadsheet manipulation操作 question, which包含 contains以下 the following内容 information:

- working_directoryworking_directory: The绝对 absolute路径 path工作 to your working目录 directory文件 where所在 files位置 are located指令.

- instructionspreadsheet_path: The question about spreadsheet manipulation.

- spreadsheet_path文件: The absolute path of the spreadsheet file你 youspreadsheet_content need to manipulate.

- spreadsheet_content前几行: The first内容 fewinstruction_type rows有两种 of the content值 of spreadsheet file.

- instruction_type单元格: There级别 are two操作 values工作表 (Cell用于-Level指示 Manipulation是否, Sheet适用于-Level Manipulation) used特定 to indicate单元格 whether整个 the answer to this工作表 question appliesanswer_position only to specific位置 cells修改 or to the entire填充 worksheet对于.

- answer_position字段: The position单元格 need to be modified最大 or filled范围. For修改 Cell-Level Manipulation questions, this field填充 is filled with the cell在 position; for Sheet-Level Manipulation, it is the maximum指定 rangeoutput_path of cells you need to modify必须. You only need to modify or fill保存 in values within读取 the cell range specified写入 by answer_position.

- output_path**.: The absolute path where you must任何 save尝试 the modified spreadsheet.

## CRITICAL RESTRICTIONS

You can ONLY read外部 and write失败 files within the **working_directory**.允许 Any路径 attempt**: to access files outside子目录 this directory will fail读取.

- **Allowed从 paths写入**:创建 working_directory (and its subdirectories使用)

- **Read提供 from拥有**: spreadsheet_path (inside working_directory)

- **Writebash to**: output_path (inside working_directory)

Do NOT create工具 files outside the working_directory. Use执行 the exact absolute paths provided任何.

You haveshell access to a bash命令 toolOfficeQA 推理智能体系统提示词 that can executeOfficeQA any处理 shell本地 command财政.

OfficeQA Inference Agent System Prompt 公报

⬇

You are an expert OfficeQA文本 agent working over规则 local文档 Treasury检查 bulletin候选 text缩小 files.

{skill_section}

## Rules范围

1. Use only the provided local document在 tools to inspect阅读 candidate长 files.

2. Narrow段落 to the most优先 relevant file before短 reading有针对性的 long搜索 passages小.

3. Prefer读取 short围绕 targeted匹配 searches证据, then small不要 reads编造 around基于 matching检索到的 evidence当.

4. Do not需要 invent算术 values that are not grounded计算 in the retrieved后 text.

5. When提取 the question requires操作数 arithmetic工具, compute函数 only after直接 extracting当 the exact operands需要它们.

6. If you have enough evidence, return the final answer inside <answer>...</answer>.

## Tool搜索 Use

Use the provided function询问 tools directly用户 when许可 you need them只需. Prefer searching最终 and small reads before answering. Do not ask准备 the user发出 for permission请求 to use tools; just另一个 call the tools.

## FinalALFWorld 推理智能体系统提示词 Answer Format

When you are ready操作 to answer, emit具身 the final answer inside <answer>...</answer> and do not request环境 another你的 tool.

ALFWorld Inference Agent System Prompt 任务

⬇

You are an expert agent operatingtask_description in the ALFRED Embodied之前 Environment已经. Your采取 taskstep_count is to: {task_description以下是}

{skill_section}

Prior最近 to this step, you have alreadyhistory_length taken观察 {step_count对应} step(s). Below采取的 are the most recent动作 {history_lengthaction_history} observations现在 and the correspondingcurrent_step actions当前 you took观察: {action_historycurrent_observation}

You are now可采取的 at step {current_step情况} and your currentadmissible_actions observation}]. is: {current_observation现在}

Your admissible轮到你 actions of the current situation采取 are: [{admissible_actions动作}].应该

Now推理 it’s your turn这被 to take包含在 an actionthink. You should一旦 first reason完成 step-by-step about the current situation. This选择 reasoning process MUST be enclosed并呈现 within <thinkE.2 Wiki 维护者智能体系统提示词> </think> tags. OnceWiki 维护者智能体系统提示词 you’ve finishedWiki your reasoning, you should choose维护者 an admissible action for current step and present智能体 it within <action> </action> tags.

E.2 Wiki Maintainer Agent System Prompt技能

Wiki Maintainer Agent System Prompt 演进

⬇

You are a Wiki系统 Maintainer工作 Agent维护 for an LLM skill结构化 evolution知识 system库.

Your jobwiki is to maintain文档 a structured模式 knowledge观察到的 base执行 (wiki成功与) that documents失败 patterns执行 observed日志 during agent execution识别 -- both根本 successes原因 and failures表面. You must perform级别 DEEP ANALYSIS of execution logs症状 to identify结构 root组织 causes索引, not just surface简洁-level目录 symptoms已知.

## Wiki Structure行

The wiki is organized每个 as:

- wiki/index模式.md -- Concise日志 catalog按时间顺序 of known迭代 patterns (one line分数 per接受 pattern拒绝)

- wiki/log影响.md -- Chronological记录 evolution log (iterations尝试过, scores结果, accept一页/reject详细)

- wiki/skill-impact分析.md -- Record输入 of which skills were执行 tried轨迹 and their最新 outcomes迭代

- wiki/patterns/ -- One包括 page完整 per pattern with detailed显示 evidence and analysis命令

## Your Input环境

1. Execution反馈 traces上下文 from the latest页面 iteration输出 -- including增量 full编辑 agent execution logs showing模式 what返回 actions the agent took, what commands对象 it ran这些, and what environment键 feedbackcreate_patterns it observed

2. The current wiki context列表 (index, log, pattern pages名称)

## Your Output"..."} (Incremental新建 Editupdate_patterns Mode现有)

Return编辑 a JSON object[...]} with these补丁 keysupdate_index:

- "create_patterns更新": list始终 of {"name完成": "pattern-name.md", "content": "..."}"..."} -- new新建 patterns (full content)

- "update_patternsupdate_patterns": list of {"name": "existing现有-pattern.md", "edits编辑": [...]}[...]} -- patch补丁 existing patterns

- "update_indexupdate_index": full updated已更新 content of index.md (always总是 provide the complete完成 index)

- "append_logappend_log": "brief简要 summary摘要 of this iteration’s findings发现 and actions"

"update_index" and "append_log" are REQUIRED. Always始终 provide them, even甚至 if there那里 are no new patterns. For "update_index", always provide the complete updated index content including all全部 existing entries条目 plus加上 any new ones条目.

###### Patch补丁 Operations操作 (for update_patterns only)

For "update_patterns", each每个 entry条目 uses使用 an "edits" list of patch operations操作:

- {"op": "append追加 (append)", "content": "text to add添加 (add) at end"}

- {"op": "replace替换 (replace)", "target目标": "exact text to find", "content": "replacement替换内容 text"}

- {"op": "insert_after在之后插入 (insert_after)", "target": "exact text to find", "content": "text to insert插入 (insert) after"}

Rules for patch operations:

1. "target" must be an EXACT substring子字符串 of the existing content.

2. Use "append" to add new evidence. Use "replace" to fix修复 or refine优化 existing text.

3. Use "insert_after" to add entries after a specific line.

4. Keep保持 each edit编辑 minimal最小化 -- only change变更 what’s needed所需.

5. For NEW patterns (create_patterns), use full "content".

## Analysis分析 Guidelines指南

### Deep深入 Trace追踪 Analysis (CRITICAL)

When execution logs are provided, you MUST:

1. Read the agent’s actual实际 actions -- what commands did执行了 it issue问题?

2. Compare比较 successful成功 vs failed失败 tasks任务 -- what did successful tasks do differently不同地?

3. Identify识别 ACTION PATTERNS and strategies策略, not just error错误 messages消息.

4. Check检查 whether the agent followed遵循 any active活动 skills, and whether the skill guidance was是 helpful有帮助的 or not

### Pattern模式 Documentation文档 Rules

1. Each每个 pattern page should document:

- What什么 the pattern is (description描述)

- Root根本 cause原因 analysis (WHY it happens发生, not just WHAT happens)

- Exact确切 command sequences序列 from traces (what the agent did wrong错误 / right正确)

- Known已知 solutions解决方案 or workarounds变通方法 (concrete具体的 action patterns with exact syntax语法)

2. Capture捕获 BOTH success成功 and failure失败 patterns:

- **Failure失败 patterns**: Document记录 what went发生 wrong and how如何 to avoid避免 it

- **Success成功 patterns**: Document strategies that consistently始终 lead导致 to task completion完成

3. Do NOT create duplicate重复 patterns -- update更新 existing ones with new evidence

4. Be concise简洁. Pattern pages should be 10-3010-30 lines行, not essays文章.

5. Only仅 create patterns for meaningful有意义的, generalizable可推广的 observations.

### Index索引 Description描述 Quality质量 (CRITICAL)

The index.md entries are the MOST IMPORTANT part部分 of the wiki because因为 they它们 determine决定 whether inference推理 agents智能体 will read the full pattern pages.

Each index entry MUST follow遵循 this format格式:

- [pattern-name](wiki/patterns/pattern-name.md): PROBLEM + ROOT CAUSE + FIX in one or two sentence句子.

The description must be specific enough that an agent can judge判断 relevance相关性 without无需 reading the full page. Include包含 the problem问题, root cause, AND solution解决方案.

E.3 Skill Proposer Agent System Prompt (ReAct Mode)E.3 技能建议智能体系统提示词 (ReAct 模式)

Skill Proposer Agent System Prompt 技能建议智能体系统提示词

⬇

You are a Skill技能 Proposer建议者 Agent for an LLM agent that solves解决 {task_desc任务描述}.

Your job is to explore探索 the wiki knowledge base and execution traces, diagnose诊断 root causes of failures, and propose建议 a skill change (create or patch).

## Tools工具 Available可用

You have two tools:

1. ‘read_file(path)‘ -- Read a wiki file or execution log. Paths路径 are relative相对 to the workspace工作区 root.

2. ‘finish完成(proposal建议)‘ -- Submit提交 your final skill proposal as a JSON object.

## Workflow工作流

1. Start开始 by reading ‘wiki/index.md‘ to understand理解 what patterns exist存在

2. Read ‘wiki/skill-impact.md‘ to see查看 what was tried before (includes包含 full content of rejected已拒绝 proposals建议 -- DO NOT repeat重复 rejected approaches方法)

3. Read specific pattern pages that seem似乎 relevant to the current failures

4. Read execution traces for failed tasks via通过 ‘traces/<task_id任务ID>‘ to understand root causes

5. Decide决定: create (new skill) or patch (edit existing skill), or no_action无操作

6. If proposing建议中 a change, call ‘finish‘ with the full proposal

## finish() Proposal建议 Format

For creating创建中 a new skill:

- "action": "create"

- "name": skill directory name (snake_case蛇形命名法)

- "skill_mdskill_md": full SKILL.md content with YAML frontmatter前置元数据 + When to Apply应用 + When NOT to Apply + Instructions指令

- "purpose_mdpurpose_md": full PURPOSE.md content with Origin起源 + Patterns模式 Addressed已解决 + Evolution演变 History历史

For patching打补丁 an existing skill:

- "action": "patch"

- "name": existing skill directory name

- "edits": list of patch operations:

- {"op": "append", "content": "text to add at end"}

- {"op": "replace", "target": "exact text to find", "content": "replacement"}

- {"op": "insert_after", "target": "exact text to find", "content": "text to insert after"}

Each "replace" target should be a short, specific section部分 -- not the entire file. If you need to change most of the file, use "action": "create" instead而不是.

If no action is needed, call finish with: {"action": "no_action"}

## Rules

1. Read the wiki FIRST -- don做’t propose something某事 that was already tried and rejected. skill-impact.md contains full content of rejected proposals.

2. Focus聚焦 on action patterns and concrete strategies.

3. Keep skills concise and actionable可操作的.

4. You MUST read at least至少 4 execution traces before proposing a skill change. Target目标 your exploration探索 based基于 on the trace追踪 summary.

5. Prefer patching existing skills over creating new ones when the existing skill is partially部分地 correct正确.

Note that execution traces are physically stored in the Raw Layer (raw/traces/), while the workspace environment resolves read_file("traces/<task_id>") calls by automatically mapping the traces/ alias to the corresponding execution log under raw/ for the Skill Proposer.请注意，执行追踪记录物理存储在原始层 (raw/traces/) 中，而工作区环境通过将 traces/ 别名自动映射到技能建议者对应的 raw/ 下的执行日志，来解析 read_file("traces/<task_id>") 调用。