Program of Thoughts
Program of thoughts is a reasoning technique in which the model generates code — typically Python — to solve a numerical or logical problem, then executes the code to obtain the answer. It separates reasoning from computation: the language model handles the plan and the interpreter handles the arithmetic.
This reduces the arithmetic errors that pure chain-of-thought tends to accumulate when stringing together multi-step calculations in natural language. The trade-off is an execution dependency: a code-running tool must be in the loop, and the model must produce runnable code rather than pseudocode. Program of thoughts suits quantitative tasks — finance, physics, word problems with many steps — more than open-ended reasoning.
Example
Asked to compute compound interest on $12,400 at 4.75% APR over 8 years with monthly compounding, a program-of-thoughts prompt instructs the model to emit a Python snippet: `p=12400; r=0.0475/12; n=8*12; print(p*(1+r)**n)`. The snippet runs in a sandbox and returns the numeric result, which the model then wraps in a plain-English answer. Pure chain-of-thought on the same problem frequently drops a digit in the exponent step.
Frequently asked questions
What is Program of Thoughts?
- Program of thoughts is a reasoning technique in which the model generates code — typically Python — to solve a numerical or logical problem, then executes the code to obtain the answer. It separates reasoning from computation: the language model handles the plan and the interpreter handles the arithmetic.
How does Program of Thoughts work?
- This reduces the arithmetic errors that pure chain-of-thought tends to accumulate when stringing together multi-step calculations in natural language. The trade-off is an execution dependency: a code-running tool must be in the loop, and the model must produce runnable code rather than pseudocode.
Can you give an example of Program of Thoughts?
- Asked to compute compound interest on $12,400 at 4.75% APR over 8 years with monthly compounding, a program-of-thoughts prompt instructs the model to emit a Python snippet: `p=12400; r=0.0475/12; n=8*12; print(p*(1+r)**n)`. The snippet runs in a sandbox and returns the numeric result, which the model then wraps in a plain-English answer. Pure chain-of-thought on the same problem frequently drops a digit in the exponent step.