Multi-Agent Debate

Multi-Agent debate 模拟不同智能体之间的多轮讨论场景，通常包括几个 solver 和一个 aggregator。 典型情况下，solver 生成并交换他们的答案，而 aggregator 收集并总结答案。

我们实现了 EMNLP 2024 中的示例，其中两个 solver 智能体将按固定顺序讨论一个话题，根据先前的辩论历史表达他们的论点。 在每一轮中，主持人智能体将决定是否可以在当前轮获得最终的正确答案。

import asyncio
import os

from pydantic import Field, BaseModel

from agentscope.agent import ReActAgent
from agentscope.formatter import (
    DashScopeMultiAgentFormatter,
)
from agentscope.message import Msg
from agentscope.model import DashScopeChatModel
from agentscope.pipeline import MsgHub

# 准备一个话题
topic = "两个圆外切且没有相对滑动。圆A的半径是圆B半径的1/3。圆A绕圆B滚动一圈回到起点。圆A总共会旋转多少次？"


# 创建两个辩论者智能体，Alice 和 Bob，他们将讨论这个话题。
def create_solver_agent(name: str) -> ReActAgent:
    """获取一个解决者智能体。"""
    return ReActAgent(
        name=name,
        sys_prompt=f"你是一个名为 {name} 的辩论者。你好，欢迎来到"
        "辩论比赛。我们的目标是找到正确答案，因此你没有必要完全同意对方"
        f"的观点。辩论话题如下所述：{topic}",
        model=DashScopeChatModel(
            model_name="qwen-max",
            api_key=os.environ["DASHSCOPE_API_KEY"],
            stream=False,
        ),
        formatter=DashScopeMultiAgentFormatter(),
    )


alice, bob = [create_solver_agent(name) for name in ["Alice", "Bob"]]

# 创建主持人智能体
moderator = ReActAgent(
    name="Aggregator",
    sys_prompt=f"""你是一个主持人。将有两个辩论者参与辩论比赛。他们将就以下话题提出观点并进行讨论：
``````
{topic}
``````
在每轮讨论结束时，你将评估辩论是否结束，以及话题正确的答案。""",
    model=DashScopeChatModel(
        model_name="qwen-max",
        api_key=os.environ["DASHSCOPE_API_KEY"],
        stream=False,
    ),
    # 使用多智能体格式化器，因为主持人将接收来自多于用户和助手的消息
    formatter=DashScopeMultiAgentFormatter(),
)


# 主持人的结构化输出模型
class JudgeModel(BaseModel):
    """主持人的结构化输出模型。"""

    finished: bool = Field(description="辩论是否结束。")
    correct_answer: str | None = Field(
        description="辩论话题的正确答案，仅当辩论结束时提供该字段。否则保留为 None。",
        default=None,
    )


async def run_multiagent_debate() -> None:
    """运行多智能体辩论工作流。"""
    while True:
        # MsgHub 中参与者的回复消息将广播给所有参与者。
        async with MsgHub(participants=[alice, bob, moderator]):
            await alice(
                Msg(
                    "user",
                    "你是正方，请表达你的观点。",
                    "user",
                ),
            )
            await bob(
                Msg(
                    "user",
                    "你是反方。你不同意正方的观点。请表达你的观点和理由。",
                    "user",
                ),
            )

        # Alice 和 Bob 不需要知道主持人的消息，所以主持人在 MsgHub 外部调用。
        msg_judge = await moderator(
            Msg(
                "user",
                "现在你已经听到了他们的辩论，现在判断辩论是否结束，以及你能得到正确答案吗？",
                "user",
            ),
            structured_model=JudgeModel,
        )

        if msg_judge.metadata.get("finished"):
            print(
                "\n辩论结束，正确答案是：",
                msg_judge.metadata.get("correct_answer"),
            )
            break


asyncio.run(run_multiagent_debate())

Traceback (most recent call last):
  File "/Users/hjt0309/Documents/AI_project/agentscope_doc/agentscope-main/docs/tutorial/zh_CN/src/workflow_multiagent_debate.py", line 117, in <module>
    asyncio.run(run_multiagent_debate())
  File "/opt/miniconda3/lib/python3.12/asyncio/runners.py", line 194, in run
    return runner.run(main)
           ^^^^^^^^^^^^^^^^
  File "/opt/miniconda3/lib/python3.12/asyncio/runners.py", line 118, in run
    return self._loop.run_until_complete(task)
           ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/opt/miniconda3/lib/python3.12/asyncio/base_events.py", line 686, in run_until_complete
    return future.result()
           ^^^^^^^^^^^^^^^
  File "/Users/hjt0309/Documents/AI_project/agentscope_doc/agentscope-main/docs/tutorial/zh_CN/src/workflow_multiagent_debate.py", line 84, in run_multiagent_debate
    await alice(
  File "/Users/hjt0309/Documents/AI_project/agentscope_doc/agentscope-main/src/agentscope/agent/_agent_base.py", line 451, in __call__
    reply_msg = await self.reply(*args, **kwargs)
                ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/Users/hjt0309/Documents/AI_project/agentscope_doc/agentscope-main/src/agentscope/agent/_agent_meta.py", line 120, in async_wrapper
    current_output = await original_func(
                     ^^^^^^^^^^^^^^^^^^^^
  File "/Users/hjt0309/Documents/AI_project/agentscope_doc/agentscope-main/src/agentscope/tracing/_trace.py", line 392, in wrapper
    return await func(self, *args, **kwargs)
           ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/Users/hjt0309/Documents/AI_project/agentscope_doc/agentscope-main/src/agentscope/agent/_react_agent.py", line 309, in reply
    msg_reasoning = await self._reasoning(tool_choice)
                    ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/Users/hjt0309/Documents/AI_project/agentscope_doc/agentscope-main/src/agentscope/agent/_agent_meta.py", line 120, in async_wrapper
    current_output = await original_func(
                     ^^^^^^^^^^^^^^^^^^^^
  File "/Users/hjt0309/Documents/AI_project/agentscope_doc/agentscope-main/src/agentscope/agent/_react_agent.py", line 430, in _reasoning
    res = await self.model(
          ^^^^^^^^^^^^^^^^^
  File "/Users/hjt0309/Documents/AI_project/agentscope_doc/agentscope-main/src/agentscope/tracing/_trace.py", line 604, in async_wrapper
    return await func(self, *args, **kwargs)
           ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/Users/hjt0309/Documents/AI_project/agentscope_doc/agentscope-main/src/agentscope/model/_dashscope_model.py", line 234, in __call__
    parsed_response = await self._parse_dashscope_generation_response(
                      ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/Users/hjt0309/Documents/AI_project/agentscope_doc/agentscope-main/src/agentscope/model/_dashscope_model.py", line 414, in _parse_dashscope_generation_response
    raise RuntimeError(response)
RuntimeError: {"status_code": 401, "request_id": "163ef91b-1986-4f1d-9602-72804412e8a8", "code": "InvalidApiKey", "message": "Invalid API-key provided.", "output": null, "usage": null}

进一步阅读

• 管道 (Pipeline)

Encouraging Divergent Thinking in Large Language Models through Multi-Agent Debate. EMNLP 2024.