[NeurIPS 2023] Large Language Model-based Autonomous Agents

Following the emergence of GPT-4 in 2023, research on Autonomous Agents that leverage LLMs' prior knowledge and generalization capabilities has risen as a core topic at NeurIPS 2023.

General-purpose LLM agent research focuses on strengthening reasoning and planning capabilities to solve complex sequential decision-making problems.

• Tree of Thoughts (ToT): Proposes a framework that goes beyond existing linear CoT by having the agent evaluate and explore multiple reasoning paths on its own.
• Reflexion: Improves the agent's action trajectory through natural language feedback instead of updating neural network weights.
• AdaPlanner: Provides adaptive planning that self-refines plans by incorporating environmental feedback.
• SwiftSage: Applies dual-process theory from cognitive science, combining intuitive thinking (Swift) and deliberate thinking (Sage).

Domain-specific LLM agent research addresses specialized tasks performed in particular environments.

• Web Agents: The Mind2Web dataset provides an environment for evaluating and developing agents' ability to manipulate real-world websites.
• Coding Agents: Interactive coding research such as InterCode, which implements natural language instructions as code, is underway.