Personal Second Brain

Most recent

See 344 more →

ART - Automatic multi-step reasoning and tool-use for large language models

2 min read

TL;DR

ART augments frozen LLMs with a task library of few-shot demonstrations that embed tool-use calls, enabling automatic interleaving of reasoning and tool execution without any gradient updates.

Summary

Frozen LLMs are augmented with a task library of few-shot demonstrations that embed tool-use calls such as search and code execution. At inference time, ART retrieves the most relevant demonstrations for a new task and interleaves reasoning steps with tool calls automatically, without gradient updates. The approach generalizes to unseen tasks via simple task descriptions, outperforming zero-shot GPT-3 and chain-of-thought prompting on BigBench and MMLU benchmarks.

Key contributions

  1. Proposes ART, a framework that automatically interleaves multi-step reasoning and tool-use calls in frozen LLMs via retrieved demonstrations.
  2. Introduces a task library of few-shot examples with embedded tool calls, enabling modular and updatable tool knowledge.
  3. Demonstrates zero-shot generalization to new tasks through task descriptions alone, without per-task prompt engineering.
  4. Shows outperformance over chain-of-thought prompting and zero-shot GPT-3 on BigBench and MMLU.

When to cite

  1. When discussing automatic tool-use interleaving in LLMs without fine-tuning or gradient updates.
  2. When citing retrieval-based approaches to few-shot demonstration selection for multi-step reasoning tasks.
  3. When arguing that a task library decouples tool knowledge from the base model, enabling modular updates.
  4. When comparing ART to ReAct or other reasoning-plus-action frameworks for tool-augmented LLMs.

LLM Prompt Engineering

Graph View

Backlinks