ReAct: 减少 LLM 幻觉,提升准确度

Summary

LLM 现在的发展是日异月新,但是目前还是有一个关键的问题:LLM 很擅长推理,也很擅长执行,但是没办法同时将两者结合起来。

ReAct framework 就是以一种交错的方式生成推理轨迹和特定任务的执行,大模型就是根据生成的推理轨迹来制定下一步的执行计划。ReAct framework 就可以使得 LLM 调用外部工具获取外部信息,这样就可以避免 LLM 幻觉,产生更加可靠,更加准确的响应。同时,ReAct 输出了推理过程和执行的操作,提高了对 LLM 可解释性和信任度

What is ReAct

下面这个是分离了推理和执行的过程

而 ReAct 就是将两者结合在了一起,这样就能够保证在执行操作时,有足够的真实信息,一定程度提升准确度:

How ReAct

ReAct framework 就是可以使 LLM 采取额外的行为,比如:访问 Goole,运行计算等,获取得到额外的信息,告诉程序如何执行这些操作,并且将操作的结果再返回给 LLM,由 LLM 决定下一步的动作。下面是一个具体的例子:

具体的代码如下:

python 复制代码
import openai
import re
import httpx

openai.api_key = "sk-..."
 
class ChatBot:
    def __init__(self, system=""):
        self.system = system
        self.messages = []
        if self.system:
            self.messages.append({"role": "system", "content": system})
    
    def __call__(self, message):
        self.messages.append({"role": "user", "content": message})
        result = self.execute()
        self.messages.append({"role": "assistant", "content": result})
        return result
    
    def execute(self):
        completion = openai.ChatCompletion.create(model="gpt-3.5-turbo", messages=self.messages)
        # Uncomment this to print out token usage each time, e.g.
        # {"completion_tokens": 86, "prompt_tokens": 26, "total_tokens": 112}
        # print(completion.usage)
        return completion.choices[0].message.content

prompt = """
You run in a loop of Thought, Action, PAUSE, Observation.
At the end of the loop you output an Answer
Use Thought to describe your thoughts about the question you have been asked.
Use Action to run one of the actions available to you - then return PAUSE.
Observation will be the result of running those actions.

Your available actions are:

calculate:
e.g. calculate: 4 * 7 / 3
Runs a calculation and returns the number - uses Python so be sure to use floating point syntax if necessary

wikipedia:
e.g. wikipedia: Django
Returns a summary from searching Wikipedia

simon_blog_search:
e.g. simon_blog_search: Django
Search Simon's blog for that term

Always look things up on Wikipedia if you have the opportunity to do so.

Example session:

Question: What is the capital of France?
Thought: I should look up France on Wikipedia
Action: wikipedia: France
PAUSE

You will be called again with this:

Observation: France is a country. The capital is Paris.

You then output:

Answer: The capital of France is Paris
""".strip()

action_re = re.compile('^Action: (\w+): (.*)$')

def query(question, max_turns=5):
    i = 0
    bot = ChatBot(prompt)
    next_prompt = question
    while i < max_turns:
        i += 1
        result = bot(next_prompt)
        print(result)
        actions = [action_re.match(a) for a in result.split('\n') if action_re.match(a)]
        if actions:
            # There is an action to run
            action, action_input = actions[0].groups()
            if action not in known_actions:
                raise Exception("Unknown action: {}: {}".format(action, action_input))
            print(" -- running {} {}".format(action, action_input))
            observation = known_actions[action](action_input)
            print("Observation:", observation)
            next_prompt = "Observation: {}".format(observation)
        else:
            return

def wikipedia(q):
    return httpx.get("<https://en.wikipedia.org/w/api.php>", params={
        "action": "query",
        "list": "search",
        "srsearch": q,
        "format": "json"
    }).json()["query"]["search"][0]["snippet"]

def simon_blog_search(q):
    results = httpx.get("<https://datasette.simonwillison.net/simonwillisonblog.json>", params={
        "sql": """
        select
          blog_entry.title || ': ' || substr(html_strip_tags(blog_entry.body), 0, 1000) as text,
          blog_entry.created
        from
          blog_entry join blog_entry_fts on blog_entry.rowid = blog_entry_fts.rowid
        where
          blog_entry_fts match escape_fts(:q)
        order by
          blog_entry_fts.rank
        limit
          1""".strip(),
        "_shape": "array",
        "q": q,
    }).json()
    return results[0]["text"]

def calculate(what):
    return eval(what)

known_actions = {
    "wikipedia": wikipedia,
    "calculate": calculate,
    "simon_blog_search": simon_blog_search
}

## 执行:
query("Fifteen * twenty five")
Thought: The action required is a calculation
Action: calculate: 15 * 25
PAUSE
 -- running calculate 15 * 25
Observation: 375
Answer: Fifteen times twenty five equals 375.

query("What does England share borders with?")
Thought: I should list down the neighboring countries of England
Action: wikipedia: England
PAUSE
 -- running wikipedia England
Observation: <span class="searchmatch">England</span> is a country that is part of the United Kingdom. It shares land borders with Wales to its west and Scotland to its north. The Irish Sea lies northwest
Answer: England shares borders with Wales and Scotland.

Ending

ReAct 是一个反复调用 LLM 进行推断,执行,观察的过程,相比较于 Function Calling,ReAct 能通过多次迭代收集足够的外部真实信息,以此来减少 LLM 的幻觉,提高响应的准确度。但是, ReAct 也很有可能陷入到循环当中,最好是在循环中设置一个 MAX_Iteration_Step,防止出现过度的迭代。

参考:react-lm.github.io/

相关推荐
FengY_HYY8 小时前
从人类组织到人类&大模型组织,我的位置在哪里?
人工智能·程序员·llm
ZzT8 小时前
open-code-review 原理解读与 Claude Code 对比实测
ai编程·claude
IT规划师8 小时前
AI 价值落地的架构决策:从四层技术栈到组织能力的最后一公里
ai编程
梯度下降者13 小时前
GPT-5.6学会写代码来调工具了,Agent的架构正在被重写
软件工程·ai编程·工具调用·效率提升·架构重写
ZhengEnCi15 小时前
Claude 的记忆是怎么设计的:四层记忆栈,和一场叫「做梦」的离线整理
aigc·agent·ai编程
J船长16 小时前
一文搞懂 AI Agent Harness:它是什么,为什么重要,以及开发者需要建立哪些认知?
ai编程
Flynt1 天前
AI 写的 Three.js 大屏,为什么越跑越卡?我在 Node 里复现了它的内存曲线
前端·ai编程·three.js
J船长1 天前
AI 知识库的主流做法是 RAG,RAG怎么做
ai编程
小虎AI生活1 天前
哇塞~24小时开机的WorkBuddy免费领,速来
aigc·ai编程
郑世录1 天前
codex 接本地模型,单次输出卡在 3208 token:一份照着做就能改的完整教程
ai编程