Toolformer: Language Models Can Teach Themselves to Use Tools

Abstraction: Self-supervised LLM fine-tuning to autonomously call and integrate external APIs

Key points:

• Toolformer (Meta, Feb 2023) trains LMs to decide which APIs to call, when to call them, what arguments to pass, and how to integrate results into future token predictions
• Training is self-supervised: requires only a handful of demonstrations per API, not large annotated datasets
• Five tool types integrated: calculator, Q&A system, two search engines, translation system, and calendar
• Achieves substantially improved zero-shot performance on downstream tasks, often competitive with much larger models, without degrading core language modeling ability
• Addresses the paradox that large LMs excel at few-shot reasoning yet fail at basic arithmetic and factual lookup where small specialized models dominate

Connections: Meta AI · Tool Use · Large Language Models · AI Agents

Source: https://arxiv.org/abs/2302.04761