Toolformer (2023-02-10)
Recent research proposes Toolformer, a model that can teach itself to use external tools via simple APIs in a self-supervised way. Augmented Language Models (ALMs) can use external modules to expand their context processing ability and learn to reason and use tools while still performing natural language tasks. However, tool-use agents face challenges in deciding when and how to use external tools, and require significant human supervision. Progress has been made on self-supervised learning of tool use with language models, but interactive tool use has not been explored yet.