Prompting Technique

Automatic Reasoning and Tool-use (ART)

A framework that combines chain-of-thought prompting with external tool use, enabling LLMs to automatically generate intermediate reasoning steps and call tools as a program.

8 min read Based on Paranjape et al., 2023 — Original Source

What is ART?

Combining Chain-of-Thought (CoT) prompting with tool use has proven reliable across many LLM tasks. The downside is that these approaches usually require hand-crafting task-specific demonstrations and carefully scripting when the model pauses for a tool and picks back up.

Automatic Reasoning and Tool-use (ART) proposes a new framework that uses a frozen LLM to automatically generate intermediate reasoning steps as a program. ART goes beyond traditional prompting by letting the model decide when and how to use external tools during its reasoning process.

Research Context

ART was introduced by Paranjape et al. (2023) and represents an important step toward more autonomous AI reasoning. It builds on earlier work in chain-of-thought prompting and tool-augmented language models.

How ART Works

ART follows a structured process that combines task selection, program generation, and tool integration:

Select Examples from a Task Library

Given a new task, ART selects demonstrations of multi-step reasoning and tool use from a pre-built task library. These examples serve as templates showing the model how to decompose similar problems.

Run Program with a Frozen LLM

At test time, the model generates a reasoning program. It pauses generation whenever external tools are called (like search engines, code interpreters, or calculators), integrates their output, and then resumes generation.

Fix Mistakes (Optional)

ART is extensible. Humans can fix mistakes in the reasoning steps or add new tools by simply updating the task and tool libraries. This makes continuous improvement straightforward.

Key Advantages

Zero-shot generalization: ART encourages the model to generalize from demonstrations to decompose a new task and use tools in appropriate places, without task-specific training.
Human extensibility: When the model makes mistakes, humans can correct the reasoning steps or add new tools to the library. This creates a feedback loop for continuous improvement.
Tool integration: ART folds external tool calls (search, code execution, calculations) into the reasoning chain, so the answers stay accurate and grounded in real data.
Frozen LLM: No fine-tuning required. ART works with the model as-is, which means it can be applied to any capable LLM without retraining.

ART in Practice

Here is a simplified example of how an ART-style prompt might look for a research task:

TaskFind the population of Canada and France, then calculate the difference.

Step 1: Search[SEARCH] “Population of Canada 2025”
→ Result: 40.1 million

Step 2: Search[SEARCH] “Population of France 2025”
→ Result: 68.2 million

Step 3: Calculate[CALC] 68.2 - 40.1
→ Result: 28.1 million

AnswerFrance has approximately 28.1 million more people than Canada.

Why This Matters for You

Even if you never build an ART system yourself, understanding it makes you a better prompter. Break a big task into steps, and figure out when the model should reach for an outside tool instead of guessing. That habit carries over to whatever AI you’re using.

When to Use ART

Tasks that require multiple reasoning steps combined with factual lookups
Workflows where accuracy matters and the model needs to verify information externally
Situations where you need reproducible, structured reasoning that humans can audit and correct
Building AI agent systems that need to decide autonomously when to call tools