Program-of-Thought Prompting Outperforms Chain-of-Thought by 15% (2022)
11 days ago
- #program synthesis
- #numerical reasoning
- #language models
- Introduces 'Program of Thoughts' (PoT) to disentangle computation from reasoning in numerical tasks.
- PoT uses language models to generate reasoning as programs, with computation handled externally.
- Evaluated on math and financial-QA datasets, showing 12% average performance gain over Chain-of-Thoughts (CoT).
- Combining PoT with self-consistency decoding achieves state-of-the-art (SoTA) performance on math datasets.
- All data and code are released on GitHub.