Taxonomy note
本页的五类回路是 reader-facing review lenses：Specification-to-Execution、Search、Evaluator、Reflection、Population/Archive。论文中的 method families 是文献分类，两者不是冲突 taxonomy；读项目时先用这五个 lens 提问，再回到论文方法族定位。
Taxonomy / Survey Rank 2
五类进化回路：把自进化从口号变成机制地图。
The definition page tells you whether a system belongs in scope. The five loops tell you how it improves: specification, search, evaluation, reflection, and archive or population pressure.
Specify Search Evaluate Reflect Archive
One sentence
五类回路不是五个标签，而是五种“反馈变成改进”的路径
一句话
自进化系统要么把目标变成可执行流程，要么搜索候选，要么评估候选，要么把失败写成可复用经验，要么在种群和 archive 中保留多条改进路径。
三句话
只会反思的系统可能只是会写复盘。只会搜索的系统可能很快过拟合评估器。把五类回路分清楚，读者才知道一个项目到底改了什么、谁在选择、结果是否能复用。
给读者的核心问题
不要问“它是不是用了 agent”。问：mutable object 是什么，feedback signal 是什么，verifier 是否独立，retention 是否影响下一次行为。
Loop matrix
五类回路速查表
| Loop | 可变对象 | 反馈信号 | 验证器 | 保留机制 |
|---|---|---|---|---|
| 规范到执行环 Specification-to-Execution | task spec, workflow, pipeline, tool plan | execution success, user goal fit, pipeline result | runner, tests, task acceptance | reusable workflow template or execution trace |
| 搜索环 Search | prompt, code, architecture, routing, hyperparameters | score, benchmark, cost, transfer result | evaluator, benchmark, validation suite | best candidate, archive, design pattern |
| 评估环 Evaluator | test harness, judge, benchmark gate, acceptance policy | pass/fail, score, regression, safety signal | independent evaluator or hidden task | promotion gate and evidence log |
| 反思环 Reflection | memory, lesson, prompt context, skill note | failed trajectory, critique, environment feedback | retry result, external check, human review | typed memory, skill, retrieval item |
| 种群环 Population | candidate pool, archive, lineage, variants | fitness, diversity, novelty, robustness | evaluator plus archive policy | parent-child lineage and elite archive |
Five loops
逐个看：每个 loop 到底回答什么
Loop 1 / Specification-to-Execution
📋→⚡ 规范到执行环
Specification-to-Execution Loop
将自然语言目标自动转化为可运行的 ML/智能体管线。
- 什么时候用：Use this lens when the question is whether an agent can turn a fuzzy goal into a repeatable process.
- 代表系统：AutoML-GPT · AutoML-Agent
- 常见失败：The system optimizes a distorted specification, or never creates a runnable artifact.
- 核心洞察：The gap between "what you want" and "what runs" is the first evolution bottleneck.
Loop 2 / Search
🔍 搜索环
Search Loop
探索架构、提示词、代码、智能体和超参数空间，寻找更优设计。
- 什么时候用：Use this lens when the system proposes multiple candidates and selects among them.
- 代表系统：OPRO · EvoPrompting · ADAS · OpenEvolve
- 常见失败：The search space becomes random walk, local optimum, or benchmark overfitting.
- 核心洞察：LLMs as search operators unlock combinatorial spaces that gradient methods cannot reach.
Loop 3 / Evaluator
✅ 评估环
Evaluator Loop
测试、基准验证候选改动是否为真正的改进。
- 什么时候用：Use this lens before trusting any claimed self-improvement result.
- 代表系统：FunSearch · Self-Debug · CodeEvolve
- 常见失败：The agent learns the evaluator rather than the task: Goodhart, leakage, weak tests.
- 核心洞察：Without a reliable evaluator, self-improvement is indistinguishable from random walk.
Loop 4 / Reflection
🪞 反思环
Reflection Loop
将失败转化为记忆，从反馈信号生成修正候选。
- 什么时候用：Use this lens when failures are converted into future context.
- 代表系统：Reflexion · Self-Refine · STaR · ExPeL
- 常见失败：Incorrect reflections become reusable errors, causing memory pollution or self-confirmation.
- 核心洞察：The reflection loop is where experience becomes reusable knowledge.
Loop 5 / Population
👥 种群环
Population Loop
维护多个候选方案，跨代选择、变异和重组。
- 什么时候用：Use this lens when multiple candidate lineages survive across generations.
- 代表系统：AlphaEvolve/MAP-Elites · LLaMEA · AgentEvolver
- 常见失败：The population collapses too early, or diversity is kept without useful selection pressure.
- 核心洞察：Population-level evolution discovers solutions no single trajectory can reach.
Composition
真实系统通常是回路组合，不是单环表演
| 系统类型 | 常见组合 | 判断重点 |
|---|---|---|
| Prompt and memory agent | Reflection + Evaluator | The agent learns from failures only if memories are typed, scoped, and re-tested. |
| Code repair agent | Specification-to-Execution + Evaluator + Reflection | The strongest feedback comes from execution traces, tests, hidden cases, and rollback logs. |
| Architecture search system | Search + Evaluator + Population | Candidate agent designs need a reliable benchmark and an archive, not only a best-so-far score. |
| Algorithm discovery system | Search + Evaluator + Population | The evaluator must be strong enough to reject beautiful but wrong programs. |
| Agent-swarm harness | Specification-to-Execution + Search + Evaluator | Roles, topology, handoff, shared state, and review gates become the mutable design space. |
How to use
读项目、写论文、做产品时怎么用
Does it count as self-evolving?
Start with the definition page, then check whether any loop retains a verified change.
Which project category is it?
Classify by mutable object and feedback signal, not by project name.
Where is the strongest evidence?
Prefer evaluator and population loops when tests, benchmarks, archives, and lineage are visible.
What should engineers build first?
Build the evaluator loop before expanding search or reflection.
What should reviewers distrust?
Distrust loops where the same model proposes, judges, stores, and celebrates the change alone.
Method families
五类回路和七个方法族怎样连接
68 papers / 34.7%
提示词/搜索优化
Prompt/Search Optimization
Largest family: LLMs rewriting their own instructions.
OPRO · EvoPrompting · Self-Refine · ACE
51 papers / 26.0%
奖励/强化学习/自博弈
Reward/RL/Self-Play
Self-generated rewards eliminate human annotation bottlenecks.
STaR · Self-Rewarding LM · RISE · Absolute Zero
28 papers / 14.3%
代码/自我修改
Code/Self-Modification
Code as the mutable substrate — agents rewrite their own source.
ADAS · DGM · Godel Agent · AlphaEvolve
16 papers / 8.2%
多智能体反思/辩论
Multi-Agent Reflection/Debate
Social pressure as selection mechanism.
Multi-Agent Debate · MAE · SPIRAL
16 papers / 8.2%
记忆/知识进化
Memory/Knowledge Evolution
Accumulated experience becomes heritable skill libraries.
Voyager · ExPeL · Memory-R1 · ReasoningBank
13 papers / 6.6%
Web/工具/环境适配
Web/Tool/Environment Adaptation
Real-world environments as the ultimate fitness function.
WebEvolver · RAGEN · SICA
4 papers / 2.0%
评估/安全/治理
Evaluation/Safety/Governance
Smallest but most critical: guarding against misevolution.
IterAlign · Constitutional AI
FAQ
四个常见误读
Are the five loops mutually exclusive?
No. Real systems usually combine them. The loops are lenses for asking what changed, what selected it, and how the change survived.
Why is evaluator loop treated as central?
Because every other loop depends on a trustworthy signal. Without an evaluator, self-improvement is indistinguishable from drift.
Can a system be self-evolving without changing weights?
Yes. Prompt, memory, workflow, tool policy, code, and archive changes can be self-evolution if they are feedback-driven, verified, and retained.
Why make this a topic page when a mechanisms page already exists?
The mechanisms page serves survey navigation; this page is an evergreen reader guide that connects definitions, paper framing, and evidence boundaries.