3 ms·
Program generation is all you need? For math, symbolic, natural language, etc.
- hyluo 3y agoThe 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 https://arxiv.org/abs/2309.10814 An automatic NLEP generation toolkit is opensourced: https://github.com/luohongyin/langcode 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.