Paper Information If LLM Is the Wizard, Then Code Is the Wand: A Survey on How Code Empowers Large Language Models to Serve as Intelligent Agents https://arxiv.org/abs/2401.00812

• improving programming and reasoning skills

• using external tools with executable steps

• performing self-improvement from feedback

Functionalities of Code Pretraining

1. Performance gain: programming, reasoning, structure generation

2. Tool Use: generate executable steps during decision-making (Planning, CoT)

3. Interaction (Agent): self-improvement

1. enhance perception and planning skills

2. direct action primitive grounding and modular memory organization

3. providing self-correction and self-improvement

Code improves performance

1. Directly improve programming skills

2. Empower complex reasoning (CoT → PoT)

3. Capture structured knowledge
1. structural reasoning
2. markup language understanding

Code improves tool using

Type of tools

Code improves self-correction

1. Selection based method
1. majority voting
2. re-ranking

2. Prompted-based method

3. Fine-tuning

Challenges

1. Causality between code pretraining and LLM’s performance improvement

2. Enhancing reasoning beyond code

3. Improving multi-turn interactions