生成-验证
Generator → Verifier
One agent generates, another verifies.
FunSearch, CodeEvolve
Mechanisms · 机制详解
五类工作回路
这是当前 Survey 用来阅读 196 篇论文展示集的 working taxonomy，不是领域已定标准；每个回路都要继续追问 mutable object、feedback、verifier、retention 和 rollback。
State Machine
自进化状态机
INIT 🏁
→ OBSERVE 👁️
→ EVALUATE 📊
→ DECIDE 🤔
→ UPDATE ✏️
→ VERIFY ✅
→ ARCHIVE 📦
可变对象：prompt · memory · skill · code · architecture · policy
Loop 1 of 5
📋→⚡ 规范到执行环
Specification-to-Execution Loop
机制
User specifies goals in natural language → system decomposes into sub-tasks → selects/configures ML pipelines → executes and evaluates → feeds results back.
代表系统
AutoML-GPTAutoML-Agent
组合模式
Feeds into Search Loop and Evaluator Loop.
⚠️ 失效模式
Ambiguous specifications lead to wrong pipeline selections.
✅ 证据
AutoML-GPT achieves competitive results on tabular benchmarks.
💡 核心洞察
The gap between "what you want" and "what runs" is the first evolution bottleneck.
Loop 2 of 5
🔍 搜索环
Search Loop
机制
Define search space → LLM proposes candidates → evaluator scores → select best → mutate/recombine → repeat.
代表系统
OPROEvoPromptingADASOpenEvolve
组合模式
Core of ADAS, OPRO, AlphaEvolve. Requires Evaluator Loop.
⚠️ 失效模式
Search space too large → random walk. Too small → local optima.
✅ 证据
OPRO discovers prompts outperforming human-designed ones. ADAS architectures transfer across tasks.
💡 核心洞察
LLMs as search operators unlock combinatorial spaces that gradient methods cannot reach.
Loop 3 of 5
✅ 评估环
Evaluator Loop
机制
Define tests/benchmarks → run candidate → score objectively → accept/reject → iterate.
代表系统
FunSearchSelf-DebugCodeEvolve
组合模式
Every other loop depends on this.
⚠️ 失效模式
Benchmark contamination. Goodhart's law.
✅ 证据
FunSearch discovered new algorithms verified by formal proof.
💡 核心洞察
Without a reliable evaluator, self-improvement is indistinguishable from random walk.
Loop 4 of 5
🪞 反思环
Reflection Loop
机制
Execute task → observe failure → generate reflection → store in memory → retrieve → revise strategy.
代表系统
ReflexionSelf-RefineSTaRExPeL
组合模式
Feeds into Search Loop. Combines with Population Loop.
⚠️ 失效模式
Reflection can be wrong. Memory grows without bound.
✅ 证据
Reflexion: HumanEval 80→91%. Voyager: skill library compounds.
💡 核心洞察
The reflection loop is where experience becomes reusable knowledge.
Loop 5 of 5
👥 种群环
Population Loop
机制
Initialize diverse population → evaluate → select top → mutate/crossover → archive elites → repeat.
代表系统
AlphaEvolve/MAP-ElitesLLaMEAAgentEvolver
组合模式
AlphaEvolve combines with Evaluator and Search.
⚠️ 失效模式
Premature convergence. Loss of diversity.
✅ 证据
AlphaEvolve: 4×4 complex-valued matrix multiplication with 48 scalar multiplications after 56 years; scope is source-specific.
💡 核心洞察
Population-level evolution discovers solutions no single trajectory can reach.
Collaboration
四种多智能体协作模式
对抗辩论
Proposer ↔ Opponent → Judge
Two agents argue; judge decides.
Multi-Agent Debate
进化锦标赛
Population → Selection → Mutation
Candidates compete; winners reproduce.
AlphaEvolve, DGM
层级监督
Worker → Reviewer → Approver
Multi-level quality gates.
DGM archive system
Methods
七大方法族速查
| 方法族 | 论文数 | 代表方法 | 一句话洞察 |
|---|---|---|---|
| 提示词/搜索优化 | 68 (34.7%) | OPRO, EvoPrompting, Self-Refine, ACE | Largest family: LLMs rewriting their own instructions. |
| 奖励/强化学习/自博弈 | 51 (26.0%) | STaR, Self-Rewarding LM, RISE, Absolute Zero | Self-generated rewards eliminate human annotation bottlenecks. |
| 代码/自我修改 | 28 (14.3%) | ADAS, DGM, Godel Agent, AlphaEvolve | Code as the mutable substrate — agents rewrite their own source. |
| 多智能体反思/辩论 | 16 (8.2%) | Multi-Agent Debate, MAE, SPIRAL | Social pressure as selection mechanism. |
| 记忆/知识进化 | 16 (8.2%) | Voyager, ExPeL, Memory-R1, ReasoningBank | Accumulated experience becomes heritable skill libraries. |
| Web/工具/环境适配 | 13 (6.6%) | WebEvolver, RAGEN, SICA | Real-world environments as the ultimate fitness function. |
| 评估/安全/治理 | 4 (2.0%) | IterAlign, Constitutional AI | Smallest but most critical: guarding against misevolution. |
