文章背景与核心概要
随着大语言模型（LLM）驱动的编程智能体在软件工程领域的广泛应用，如何让智能体适应不同开发者的个性化习惯成为了一个关键挑战。本文探讨了从开发者与编程智能体的历史交互中提炼出的“个性化技能”是否能有效提升智能体在未来任务中的表现。
研究团队开发了一套通过基于规则的引导（rule-based bootstrapping）和基于证据的精炼（evidence-grounded refinement）来提取技能的框架，并在206个真实开发者-智能体交互会话中对这些技能进行了评估。研究结果表明，跨多个开发者汇总而成的“通用技能（generic skills）”比个性化技能能够带来更一致、更显著的性能提升。只有当特定的开发者偏好具有极高的频率且与未来任务高度相关时，个性化技能才展现出有效性。这表明，在当前的智能体架构下，具备广泛可迁移性的程序性知识比局部的、开发者特定的偏好建模更为稳健。
Do Personalized Skills Help Coding Agents? An Empirical Study of Developer Interaction Histories
Authors: Shuyan Huang, Kai Du, Andrew Lan
Date: August 10, 2026
Subject: Software Engineering (cs.SE); Artificial Intelligence (cs.AI)
Identifier: arXiv:2608.10319
Summary
This study investigates whether "personalized skills"—distilled from a developer's past interactions with coding agents—can effectively improve agent performance on future tasks. Using a framework that extracts skills via rule-based bootstrapping and evidence-grounded refinement, the authors evaluated these skills against generic and baseline models across 206 real-world developer-agent sessions.
The findings indicate that generic skills (pooled across multiple developers) provide more consistent and significant gains than personalized ones. Personalized skills only demonstrate effectiveness when specific developer preferences appear with high frequency and relevance to future tasks, suggesting that broadly transferable procedural knowledge is currently more robust than individualized preference modeling.
本研究探讨了从开发者与编程智能体过往交互中提炼出的“个性化技能”是否能有效提升智能体在后续任务中的表现。作者使用了一个通过基于规则的引导和基于证据的精炼来提取技能的框架，并在206个真实的开发者-智能体交互会话中，将这些个性化技能与通用模型及基准模型进行了评估。
研究结果表明，通用技能（跨多个开发者汇总所得）比个性化技能能够提供更一致、更显著的性能提升。只有当特定的开发者偏好以高频率出现并且与未来任务高度相关时，个性化技能才会显现出效果；这表明，在当前阶段，具备广泛可迁移性的程序性知识比个性化的偏好建模更加稳健。
Research Overview
Large language model (LLM)-powered agents have evolved from simple code-completion tools into sophisticated solvers for complex software engineering tasks. As these agents become integral to development workflows, the ability to adapt to individual developer preferences becomes a key challenge.
Key Objectives
- Skill Extraction: Developing a framework to extract reusable preferences from interaction traces without modifying underlying model parameters.
- Evaluation: Utilizing a reproducible replay framework with an interactive, trajectory-conditioned LLM-based human developer simulator.
- Benchmarking: Comparing personalized skills against:
- No-skill baselines.
- Generic-skill baselines (pooled).
- Other-user-skill baselines.
研究概述
由大语言模型（LLM）驱动的智能体已经从简单的代码补全工具演变为解决复杂软件工程任务的高级求解器。随着这些智能体逐渐融入开发工作流，如何适应个体开发者的偏好成为了一个核心挑战。
核心目标
- 技能提取： 开发一个框架，从交互轨迹中提取可重用的偏好，而无需修改底层模型参数。
- 评估方法： 利用一个可复现的回放框架，并配备基于交互式、轨迹条件化的基于LLM的人类开发者模拟器。
- 基准测试： 将个性化技能与以下基准进行对比：
- 无技能基准（No-skill baselines）。
- 通用技能基准（汇总所得）。
- 其他用户技能基准。
Key Findings
- Personalization Limits: Personalized skills offer only small and inconsistent improvements over no-skill baselines.
- Generic Superiority: Generic skills, derived from a broader pool of developer interactions, consistently yield the most significant performance gains.
- Contextual Success: Personalization is most effective when a developer’s history contains frequent, relevant examples that directly map to future tasks.
- Strategic Insight: The research suggests that while personalization is a compelling goal, current agent architectures benefit more from broadly transferable procedural knowledge than from narrow, developer-specific preference signals.
核心发现
- 个性化的局限性： 与无技能基准相比，个性化技能只能带来微小且不稳定的性能提升。
- 通用技能的优越性： 来源于更广泛开发者交互池的通用技能，能够持续带来最显著的性能收益。
- 情境化的成功： 只有当开发者的历史记录中包含频繁且相关的示例、并能直接映射到未来任务时，个性化才最有效。
- 战略洞察： 研究表明，尽管个性化是一个引人注目的目标，但当前的智能体架构从广泛可迁移的程序性知识中获得的收益，要大于从狭隘的、特定于开发者的偏好信号中获得的收益。
Access & Resources
Metadata: 15 pages, 10 figures. Submitted August 10, 2026.
访问与资源
元数据：15页，10张图表。提交于 2026年8月10日。