[NeurIPS 2023] LLM-based Autonomous Agents - LG AI Research Blog

With the emergence of GPT-4 in 2023, LLM (Large Language Model) has achieved success in various fields beyond AI. Accordingly, at NeurIPS 2023, research on Autonomous Agents leveraging the vast prior knowledge and generalization capabilities of LLMs was covered as a key topic.

An autonomous agent refers to a computing system that acts autonomously to achieve intended outcomes while interacting with its environment. This research is broadly divided into general-purpose agents and domain-specific agents.

General-purpose LLM agent research focuses on reasoning and planning to solve complex decision-making problems. Key research cases are as follows.

• Tree of Thoughts (ToT): A prompting framework that goes beyond the existing linear CoT approach, enabling the LLM to explore and evaluate multiple reasoning paths on its own. It boosted the success rate on the 'Game of 24' problem from 4% to 74%.
• Reflexion: A framework that improves the agent's action trajectory through natural language feedback, instead of updating the neural network's weights.
• AdaPlanner: Supports adaptive planning that improves the generated plan on its own based on feedback from the environment.
• SwiftSage: Inspired by the dual-process theory of cognitive science, it consists of a Swift module responsible for intuitive, fast thinking and a Sage module responsible for deep thinking.

In addition, research specialized in specific domains is also actively underway, including Web Agent that performs web browsing, Coding Agent that writes code, and Embodied Agent that interacts with the physical environment.