Program of Thought: Definition and Examples

Program of Thought (PoT) is an advanced prompting method introduced by Chen et al. in 2022, which separates reasoning from computation execution. Rather than asking the model to solve a problem step by step in natural language (as in Chain of Thought), it is asked to generate a program — typically in Python — that encodes the solution logic. This program is then executed by an external interpreter to obtain the final answer.

The intuition behind this approach is simple: language models are excellent at understanding a problem and formulating a solution strategy, but they frequently make arithmetic or logical calculation errors. By delegating execution to a code interpreter, we combine the best of both worlds — the LLM's linguistic understanding and the computational precision of an execution engine.

PoT has proven particularly effective on math reasoning benchmarks like GSM8K and AQuA, where it significantly outperforms classic Chain of Thought. It also excels in problems involving iterations, complex data structures, or financial calculations — areas where natural language reasoning reaches its limits.

This technique is part of a broader movement to augment LLMs with external tools. It is now natively integrated into several AI agent frameworks and constitutes one of the foundations of the "code as reasoning" approach gaining popularity in the community.