reasoning-machines/pal

The framework interleaves LLM-generated text reasoning with executable Python code, delegating arithmetic and procedural computation to a runtime interpreter rather than relying on the model alone. It provides a `ProgramInterface` class that orchestrates prompt execution, code generation, and answer extraction across OpenAI APIs (Codex, GPT-3.5-turbo), with few-shot prompting examples for math, symbolic reasoning, and question-answering tasks. The implementation includes evaluation scripts for benchmark datasets like GSM8k and the authors' GSM-hard variant.

518 stars. No commits in the last 6 months.