ART Framework: Automatic Reasoning and Tool-use Approach

When using LLMs to complete tasks, alternating between CoT prompts and tool use has proven to be a powerful and robust method. Such approaches typically require handwritten demonstrations for specific tasks and carefully crafted scripts for alternating between generative models and tools.

Paranjape et al. (2023) proposed a new framework called ART (Automatic Reasoning and Tool-use), which uses a frozen LLM to automatically generate programs containing intermediate reasoning steps.

How ART Works

• When receiving a new task, select multi-step reasoning and tool-use demonstrations from the task library.During testing, pause generation when calling external tools, integrate the tool outputs, and then continue generating.

ART guides the model to summarize demonstrations, split new tasks, and use tools appropriately, operating in a zero-shot manner. Additionally, ART supports manual expansion: simply updating the task and tool libraries can fix errors in reasoning steps or add new tools.