六大自进化系统深度分析

它做了什么？

An agent that rewrites its own Python source code to discover better agent architectures.

它怎么进化的？

Maintains an open-ended archive of agent designs. Each generation proposes code modifications, evaluates them on benchmarks, and keeps improvements in the archive for future generations to build on.

📊 证据

💡 核心启示

Code-level self-modification with archival inheritance outperforms fixed-architecture approaches on real-world tasks.

它做了什么？

A meta-agent that writes complete Python-coded agents, searching the space of agent designs automatically.

它怎么进化的？

Uses a "meta-agent" that proposes new agent architectures as Python code, evaluates them on tasks, and iterates. Discovered designs transfer across tasks and base models.

📊 证据

💡 核心启示

Agent design space is searchable — meta-agents can discover non-obvious architectures that human designers miss.

它做了什么？

Combines LLMs with evolutionary computation to discover novel algorithms for math, scheduling, and hardware design.

它怎么进化的？

LLM proposes code mutations -> automated evaluators score -> MAP-Elites maintains diverse population -> selected solutions survive. The paper reports a new matrix multiplication result in a historically studied task.

📊 证据

💡 核心启示

LLM + evolution + strong evaluator can support algorithm discovery on tasks with executable scoring; claims should stay tied to the reported benchmark context.

它做了什么？

An LLM-powered agent that plays Minecraft by building a reusable skill library through autonomous exploration.

它怎么进化的？

Three components: automatic curriculum (what to learn next), skill library (reusable code functions), and iterative prompting (fix failed skills). Skills compound — early discoveries enable later ones.

📊 证据

💡 核心启示

Skill libraries can create compound returns, but stale or mis-retrieved skills can also pollute future behavior.

它做了什么？

Agents that learn from failures by generating verbal reflections stored in episodic memory.

它怎么进化的？

After each failed attempt, the agent generates a textual "reflection" about what went wrong. Future attempts retrieve relevant reflections as additional context, avoiding repeated mistakes.

📊 证据

💡 核心启示

Language can act as a medium of self-improvement when reflections are scoped, retrievable, and checked against task outcomes.

它做了什么？

Language models that generate their own training rewards, creating a self-improvement loop without external annotation.

它怎么进化的？

The model generates candidate responses → judges them using its own reward model → trains on self-judged best responses via DPO → the improved model becomes a better judge → repeat.

📊 证据

💡 核心启示

The evaluator and the evaluated can be the same system — but only with careful design to prevent reward hacking.