38 lines
1.3 KiB
Markdown
38 lines
1.3 KiB
Markdown
---
|
||
title: "自进化变异 — Self-Evolutionary Mutation"
|
||
created: 2026-06-19
|
||
updated: 2026-06-19
|
||
type: concept
|
||
tags: [mutation, tool-generation, data-augmentation, ace-router]
|
||
sources:
|
||
- https://arxiv.org/abs/2601.08276
|
||
---
|
||
|
||
# 自进化变异(Self-Evolutionary Mutation)
|
||
|
||
## 定义
|
||
|
||
Self-Evolutionary Mutation 是 ACE-Router 的候选空间扩展策略:通过**LLM 驱动的受控变异**,从已有工具合成行为多样化的新工具变体,丰富训练数据的判别难度。
|
||
|
||
## 五种变异算子
|
||
|
||
| 算子 | 示例 |
|
||
|------|------|
|
||
| **Function Enhancement** | `get_stock_prices` → `get_stock_prices_with_news` |
|
||
| **Parameter Mutation** | 修改参数类型、必选/可选、范围 |
|
||
| **Workflow Chaining** | 组合多个操作为流水线工具 |
|
||
| **Helper Operation** | 创建辅助/便利包装 |
|
||
| **Usage Extension** | 扩展至新的使用场景 |
|
||
|
||
## 为什么要变异
|
||
|
||
- 真实工具的**功能相似性**使 embedding 难以区分
|
||
- 变异生成的"近义词"工具迫使路由器学会精确辨识
|
||
- 在噪声实验中,注入的变异工具对 GPT-4o 和 Gemini 造成了严重干扰(准确率降至 28-32%),而 ACE-Router 保持 56%
|
||
|
||
## 参考
|
||
|
||
- [[candidate-graph|候选图]]
|
||
- [[ace-router|ACE-Router]]
|
||
- [[yao-ace-router-2026|论文]]
|