背景
最近做业务需要了解什么是交错思考，以前我只知道thinking机制。
本质上就是在模型正式输出前，让其输出。内部有具体的思考活动。
但是这种方式，在下一次访问llm的时候，一般会把thinking内容去除，只发送实际内容给llm。
什么是交互思考
参考这张图片，其实还是挺清晰的：
这是 GLM-4.7 最具标志性的设计之一。
在多轮对话或长任务执行中，模型可以保留前序推理结构与中间结论，而不是每一轮都“从零开始理解上下文”。这意味着：
从数据上来说，就是每次输出的reasoningDetails保留下来，把thinking的中间内容保留了。
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
|
{
"content": [
{
"text": "，从1输出到100，每输出一个数字换一下行",
"type": "text"
}
],
"type": "user",
"updateTime": "2026-01-12T12:10:18.300+0000"
},
{
"content": [
{
"reasoningContent": "The user is asking me to output numbers from 1 to 100, with each number on a new line. This is a simple programming task that I can help with. Let me create a simple script to do this.\n\nThe user wrote \"ultrathink\" which suggests they want me to think carefully about this, but this is a straightforward task.\n\nI should create a simple script that prints numbers 1 to 100, each on a new line. I'll use Python since it's a common language and very straightforward for this task.",
"reasoningDetails": [
{
"format": "MiniMax-response-v1",
"id": "reasoning-text-1",
"index": 0,
"text": " Python since it's a common language and very straightforward for this task.",
"type": "reasoning.text"
}
],
"text": "我来为你创建一个输出1到100数字的程序，每个数字占一行：\n",
"type": "text"
},
{
"toolCallId": "111",
"toolName": "edit_file",
"toolParams": "{\"target_file\": \"print_numbers.py\", \"instructions\": \"Create a Python script that prints numbers from 1 to 100, each on a new line\", \"code_edit\": \"# 输出1到100的数字，每个数字占一行\\nfor i in range(1, 101):\\n print(i)\"}",
"type": "tool_use"
}
],
},
|
ref
(5 封私信 / 11 条消息) 为什么 2025 的基模公司都押注 Interleaved Thinking？ - 知乎
