AlphaEvolve 怎么做的？
AlphaEvolve 的核心是一个进化循环：LLM 提出代码改动 → 自动评估器打分 → 好的方案存入数据库 → 从数据库采样组装下一轮 prompt → 重复。先看论文原图，再逐步拆解：
流程图拆解
关键机制：Prompt 是怎么拼出来的？
下面用论文 Figure 3 中的图像分类任务作为具体例子，展示 prompt 的完整拼接过程和 LLM 的输出：
真实示例：给 LLM 看的完整 Prompt（论文 Figure 3b）
以图像分类任务为例，展开看 prompt 的四个拼接部分
Part 1：系统指令
Act as an expert software developer.
Your task is to iteratively improve the
provided codebase. [...]
Part 2：历史高分程序（从数据库采样）
- Prior programs
Previously we found that the following
programs performed well on the task:
top_1_acc: 0.796; neg_eval_log_loss: 0.230 average_score: 0.513
class ConvNet(hk.Module): def __init__(self, num_channels=32, ...): self._conv1 = hk.Conv2D(num_channels, 3) self._conv2 = hk.Conv2D(num_channels*2, 3) self._conv3 = hk.Conv2D(num_channels*4, 3) [...]
top_1_acc: 0.796; neg_eval_log_loss: 0.230 average_score: 0.513
class ConvNet(hk.Module): def __init__(self, num_channels=32, ...): self._conv1 = hk.Conv2D(num_channels, 3) self._conv2 = hk.Conv2D(num_channels*2, 3) self._conv3 = hk.Conv2D(num_channels*4, 3) [...]
告诉 LLM：「之前有人用这个简单 CNN，准确率 79.6%」
Part 3：当前待改进程序（数据库中更高分的版本）
- Current program
Here is the current program we are trying
to improve (you will need to propose a
modification to it below).
top_1_acc: 0.862; neg_eval_log_loss: 0.387 average_score: 0.624
class ConvNet(hk.Module): def __init__(self, ...): self._conv1 = hk.Conv2D(num_channels, 3) self._block1 = ResNetBlock(num_channels) self._block2 = ResNetBlock(num_channels*2, stride=2) self._block3 = ResNetBlock(num_channels*4, stride=2) [...]
top_1_acc: 0.862; neg_eval_log_loss: 0.387 average_score: 0.624
class ConvNet(hk.Module): def __init__(self, ...): self._conv1 = hk.Conv2D(num_channels, 3) self._block1 = ResNetBlock(num_channels) self._block2 = ResNetBlock(num_channels*2, stride=2) self._block3 = ResNetBlock(num_channels*4, stride=2) [...]
告诉 LLM：「有人后来改成了 ResNet，准确率提到了 86.2%，你在这个基础上继续改」
Part 4：任务指令 + diff 格式要求
SEARCH/REPLACE block rules: [...]
Task: Suggest a new idea to improve the code that is inspired by your expert knowledge of optimization and ML. Describe each change with a SEARCH/REPLACE block.
Task: Suggest a new idea to improve the code that is inspired by your expert knowledge of optimization and ML. Describe each change with a SEARCH/REPLACE block.
LLM 的输出：代码 diff（论文 Figure 3c）
LLM 先给出改动理由，然后输出精准的 SEARCH/REPLACE 补丁
LLM 的思考
The current model uses a simple ResNet with
only three blocks. We can improve by:
1) increasing model capacity (more blocks)
2) adding regularization (weight decay)
3) switching to AdamW optimizer
LLM 输出的 diff 补丁
<<<<<<< SEARCH
self._block1 = ResNetBlock(num_channels)
self._block2 = ResNetBlock(num_channels*2, stride=2)
self._block3 = ResNetBlock(num_channels*4, stride=2)
=======
self._block1 = ResNetBlock(num_channels)
self._block2 = ResNetBlock(num_channels, stride=1)
self._block3 = ResNetBlock(num_channels*2, stride=2)
self._block4 = ResNetBlock(num_channels*2, stride=1)
self._block5 = ResNetBlock(num_channels*4, stride=2)
self._block6 = ResNetBlock(num_channels*4, stride=1)
>>>>>>> REPLACE
<<<<<<< SEARCH def optimizer(self, learning_rate): return optax.adam(learning_rate) ======= def optimizer(self, learning_rate): return optax.adamw(learning_rate, weight_decay=1e-4) >>>>>>> REPLACE
<<<<<<< SEARCH def optimizer(self, learning_rate): return optax.adam(learning_rate) ======= def optimizer(self, learning_rate): return optax.adamw(learning_rate, weight_decay=1e-4) >>>>>>> REPLACE
这段 diff 会被自动应用到 Part 3 的当前程序上，生成一个新的候选程序，然后送去评估。如果评分比 0.624 高，就会被存入数据库，成为下一轮 prompt 中的「当前程序」。
