The paper introduces improved performance by prompting LLMs with "natural language embedded programs (NLEP)". No task-specific prompt is needed.
Paper: https://arxiv.org/abs/2309.10814
An automatic NLEP generation toolkit is opensourced: https://github.com/luohongyin/langcode
Example Colab notebook is included in the Github repo.
This work introduces the following features of NLEP
1. NLEP is a full python program that prints the target response of LLMs.
2. Task-general NLEP prompting outperforms task-specific chain-of-thought prompting on math, symbolic, and natural language.
3. Enable the chain-of-thought reasoning ability of small models (RoBERTa) on text classification
4. Hierarchical instructing via program completion.
