Home

💻 Code as Action

Introduction

While innovative, projects like ReAct have limitations. Accessing information from platforms like Wikipedia can be time-consuming and inefficient due to the need to create a 'Wikipedia environment.'

🚀 Efficient Language Model Augmentation

Our approach leverages the vast possibilities offered by coding. We distinguish between two types of tasks:

• Controllable Tasks 🔧: These include actions like accessing the internet or performing computations.
• Non-Controllable Tasks 🧩: These are tasks requiring reasoning or strong language priors.

By focusing on this distinction, we can extend language model capabilities efficiently and effectively, making the most of code as an action in our language model.

🎯 Project Goals

The project aims to empower a language learning model (LLM) to generate, execute, debug code, and answer queries. The key goal is to record and reuse this process for better performance.

Consider a user request:

"Give me the Amazon URL for the last Booker Prize-winning book."

This task involves accessing Wikipedia or a search engine and searching Amazon. If our LLM has reusable abstracted codes for such tasks, we can perform more complex tasks more efficiently.

The project aligns with the web version of Voyager. 🌐

The workflow is as follows:

• Step 1: User makes a request.
• Step 2: LLM generates code to execute the request.
• Step 3: The code is executed and debugged if necessary.
• Step 4: LLM answers the user's query based on the results.
• Step 5: The process is recorded and abstracted for future use.

With time, the LLM becomes more efficient and versatile, leading to enhanced performance.