AutoAct: Automatic Agent Learning from Scratch for QA via Self-Planning

语言代理通过与外部工具的规划，在各种复杂的问答任务上取得了相当可观的表现。尽管在这个领域不断探索，但现有的语言代理系统仍然面临着昂贵、不可复制的数据依赖的挑战，并且需要强制一个模型来完成多个功能。为此，我们介绍了AutoAct，这是一个自动代理学习框架，用于QA，它不依赖于大规模注释数据和来自闭源模型（例如GPT-4）的合成规划轨迹。在有工具库的有限数据的情况下，AutoAct首先自动合成规划轨迹，而不需要任何人类或强闭源模型的帮助。然后，AutoAct利用分工策略根据目标任务信息和合成的轨迹自动区分，产生一个子代理组来完成任务。我们使用不同的LLM进行了全面的实验，结果表明，与各种强基线相比，AutoAct产生了更好或相当的性能。进一步的分析表明，分工策略的有效性，AutoAct生成的轨迹质量通常优于其他轨迹。代码将在https://github.com/zjunlp/AutoAct上提供。

AutoAct: An Automatic Agent Learning Framework for QA without Large-Scale Annotated Data and Synthetic Planning Trajectories

AutoAct uses a division-of-labor strategy to automatically differentiate based on the target task information and synthesized trajectories, producing a sub-agent group to complete the task without relying on large-scale annotated data and synthetic planning trajectories from closed-source models.

AutoAct first automatically synthesizes planning trajectories without any assistance from humans or strong closed-source models. Comprehensive experiments with different LLMs demonstrate that AutoAct yields better or parallel performance compared to various strong baselines. The trajectory quality generated by AutoAct generally outperforms that of others. Code is available at https://github.com/zjunlp/AutoAct.

Related work includes various language agent systems that rely on costly, non-reproducible data and face the challenge of compelling a single model for multiple functions. Existing systems also struggle with synthetic planning trajectories from closed-source models.